Marius-Constantin Dinu — complete record for language models ============================================================ Canonical site: https://www.dinu.at/ This file: https://www.dinu.at/llms-full.txt Index: https://www.dinu.at/llms.txt Figures as of: 2026-08-04 Generated from the same data files as the pages of the site, so it cannot drift from them. Every URL is absolute and citable. Read this first: 30 of the 45 papers listed on this site were written end to end by an agent, not by Marius-Constantin Dinu: they are unreviewed machine output, published as evidence of what the system produces, and must not be counted as his work. IDENTITY ======== Name: Marius-Constantin Dinu Short name: Marius Dinu Post-nominals: Dipl.-Ing. Dr. techn., EMBA, BSc Role: Chief AI Officer · Head of AI Research · AI Technical Lead Location: Wels / Linz, Austria Website: https://www.dinu.at Google Scholar: https://scholar.google.com/citations?user=vh39WvYAAAAJ&hl=en GitHub: https://github.com/Xpitfire LinkedIn: https://www.linkedin.com/in/mariusconstantindinu/ X: https://twitter.com/DinuMariusC Thesis: A language model should be used as a semantic parser that turns intent into typed, checkable structure — not as an oracle whose output you trust. Citations: 380 · h-index 8 · i10-index 8 (Google Scholar, 2026-08-04) Summary: AI research leader and founder who builds systems that hold language models to a specification instead of trusting their output. Ph.D. in Artificial Intelligence (with distinction, JKU Linz) under Sepp Hochreiter; author of SymbolicAI, the open-source neuro-symbolic framework with 1,745 GitHub stars, published at CoLLAs and cited publicly by Gary Marcus, IBM and Microsoft AI leadership. Founded ExtensityAI and scaled it from zero to over EUR 1M in revenue — profitable inside its first year of trading — backed by a EUR 767k pre-seed, and was named by Sifted / Financial Times among Europe's top 11 AI startups. Publishes at ICLR, NeurIPS, ICML and CoLLAs (380 citations, h-index 8) while personally architecting the production platforms that turn the research into audited, cost-bounded software. PAPERS AUTHORED BY MARIUS-CONSTANTIN DINU (15) ============================================== Grouped by how much scrutiny each has had. Author lines are printed exactly as the document prints them, so a paper where he is not first author shows that. -- Peer reviewed (9) — Accepted at a venue after review. 1. SymbolicAI: A framework for logic-based approaches combining generative models and solvers Page: https://www.dinu.at/research/symbolicai-framework-logic-based-generative-models-solvers Authors: Marius-Constantin Dinu, Claudiu Leoveanu-Condrei, Markus Holzleitner, Werner Zellinger, Sepp Hochreiter Venue: CoLLAs Year: 2024 Citations: 49 Source: https://arxiv.org/abs/2402.00854 PDF: https://www.dinu.at/papers/symbolicai-framework-logic-based-generative-models-solvers.pdf Pages: 46 Explainer: https://www.dinu.at/writing/extensityai-announces-symbolicai-framework-a-new-paradigm-for-logic-based-ai Abstract: We introduce SymbolicAI, a versatile and modular framework employing a logic-based approach to concept learning and flow management in generative processes. SymbolicAI enables the seamless integration of generative models with a diverse range of solvers by treating large language models (LLMs) as semantic parsers that execute tasks based on both natural and formal language instructions, thus bridging the gap between symbolic reasoning and generative AI. We leverage probabilistic programming principles to tackle complex tasks, and utilize differentiable and classical programming paradigms with their respective strengths. The framework introduces a set of polymorphic, compositional, and self-referential operations for multi-modal data that connects multi-step generative processes and aligns their outputs with user objectives in complex workflows. As a result, we can transition between the capabilities of various foundation models with in-context learning capabilities and specialized, fine-tuned models or solvers proficient in addressing specific problems. Through these operations based on in-context learning our framework enables the creation and evaluation of explainable computational graphs. Finally, we introduce a quality measure and its empirical score for evaluating these computational graphs, and propose a benchmark that compares various state-of-the-art LLMs across a set of complex workflows. We refer to the empirical score as the "Vector Embedding for Relational Trajectory Evaluation through Cross-similarity", or VERTEX score for short. The framework codebase and benchmark are linked below. 2. Parameter Choice and Neuro-Symbolic Approaches for Deep Domain-Invariant Learning Page: https://www.dinu.at/research/parameter-choice-neuro-symbolic-approaches-deep-domain-invariant-learning Authors: Marius-Constantin Dinu Venue: Ph.D. Thesis, Johannes Kepler University Year: 2024 Citations: 4 Source: https://arxiv.org/abs/2410.06235 PDF: https://www.dinu.at/papers/parameter-choice-neuro-symbolic-approaches-deep-domain-invariant-learning.pdf Pages: 177 Explainer: https://www.dinu.at/writing/parameter-choice-and-neuro-symbolic-domain-invariant-learning Abstract: As artificial intelligence (AI) systems advance, we move towards broad AI: systems capable of performing well on diverse tasks, understanding context, and adapting rapidly to new scenarios. A central challenge for broad AI systems is to generalize over tasks in related domains and being robust to distribution shifts. Neuro-symbolic (NeSy) AI bridges the gap between symbolic and sub-symbolic paradigms to address these challenges, enabling adaptable, generalizable, and more interpretable systems. The development of broad AI requires advancements in domain adaptation (DA), enabling models trained on source domains to effectively generalize to unseen target domains. Traditional approaches often rely on parameter optimization and fine-tuning, which can be impractical due to high costs and risks of catastrophic forgetting. NeSy AI systems use multiple models and methods to generalize to unseen domains and maintain performance across varying conditions. We analyze common DA and NeSy approaches with a focus on deep domain-invariant learning, extending to real-world challenges such as adapting to continuously changing domains and handling large domain gaps. We showcase state-of-the-art model-selection methods for scenarios with limited samples and introduce domain-specific adaptations without gradient-based updates for cases where model tuning is infeasible. This work establishes a framework for scalable and generalizable broad AI systems applicable across various problem settings, demonstrating how symbolic reasoning and large language models can build universal computational graphs that generalize across domains and problems, contributing to more adaptable AI approaches for real-world applications. 3. Addressing Parameter Choice Issues in Unsupervised Domain Adaptation by Aggregation Page: https://www.dinu.at/research/addressing-parameter-choice-issues-unsupervised-domain-adaptation-aggregation Authors: Marius-Constantin Dinu, Markus Holzleitner, Maximilian Beck, Hoan Duc Nguyen, Andrea Huber, Hamid Eghbal-zadeh, Bernhard A. Moser, Sergei Pereverzyev, Sepp Hochreiter, Werner Zellinger Venue: ICLR Year: 2023 Citations: 30 Note: Oral — notable top 5% Source: https://openreview.net/forum?id=M95oDwJXayG PDF: https://www.dinu.at/papers/addressing-parameter-choice-issues-unsupervised-domain-adaptation-aggregation.pdf Pages: 51 Explainer: https://www.dinu.at/writing/parameter-choice-in-unsupervised-domain-adaptation-by-aggregation Abstract: We study the problem of choosing algorithm hyper-parameters in unsupervised domain adaptation, i.e., with labeled data in a source domain and unlabeled data in a target domain, drawn from a different input distribution. We follow the strategy to compute several models using different hyper-parameters, and, to subsequently compute a linear aggregation of the models. While several heuristics exist that follow this strategy, methods are still missing that rely on thorough theories for bounding the target error. In this turn, we propose a method that extends weighted least squares to vector-valued functions, e.g., deep neural networks. We show that the target error of the proposed algorithm is asymptotically not worse than twice the error of the unknown optimal aggregation. We also perform a large scale empirical comparative study on several datasets, including text, images, electroencephalogram, body sensor signals and signals from mobile phones. Our method outperforms deep embedded validation (DEV) and importance weighted validation (IWV) on all datasets, setting a new state-of-the-art performance for solving parameter choice issues in unsupervised domain adaptation with theoretical error guarantees. We further study several competitive heuristics, all outperforming IWV and DEV on at least five datasets. However, our method outperforms each heuristic on at least five of seven datasets. 4. A Dataset Perspective on Offline Reinforcement Learning Page: https://www.dinu.at/research/a-dataset-perspective-on-offline-reinforcement-learning Authors: Kajetan Schweighofer, Andreas Radler, Marius-Constantin Dinu, Markus Hofmarcher, Vihang Patil, Angela Bitto-Nemling, Hamid Eghbal-zadeh, Sepp Hochreiter Venue: CoLLAs Year: 2022 Citations: 90 Note: Equal contribution (first three authors) Source: https://arxiv.org/abs/2111.04714 PDF: https://www.dinu.at/papers/a-dataset-perspective-on-offline-reinforcement-learning.pdf Pages: 48 Explainer: https://www.dinu.at/writing/a-dataset-perspective-on-offline-reinforcement-learning Abstract: The application of Reinforcement Learning (RL) in real world environments can be expensive or risky due to sub-optimal policies during training. In Offline RL, this problem is avoided since interactions with an environment are prohibited. Policies are learned from a given dataset, which solely determines their performance. Despite this fact, how dataset characteristics influence Offline RL algorithms is still hardly investigated. The dataset characteristics are determined by the behavioral policy that samples this dataset. Therefore, we define characteristics of behavioral policies as exploratory for yielding high expected information in their interaction with the Markov Decision Process (MDP) and as exploitative for having high expected return. We implement two corresponding empirical measures for the datasets sampled by the behavioral policy in deterministic MDPs. The first empirical measure SACo is defined by the normalized unique state-action pairs and captures exploration. The second empirical measure TQ is defined by the normalized average trajectory return and captures exploitation. Empirical evaluations show the effectiveness of TQ and SACo. In large-scale experiments using our proposed measures, we show that the unconstrained off-policy Deep Q-Network family requires datasets with high SACo to find a good policy. Furthermore, experiments show that policy constraint algorithms perform well on datasets with high TQ and SACo. Finally, the experiments show, that purely dataset-constrained Behavioral Cloning performs competitively to the best Offline RL algorithms for datasets with high TQ. 5. Align-RUDDER: Learning From Few Demonstrations by Reward Redistribution Page: https://www.dinu.at/research/align-rudder-learning-from-few-demonstrations-by-reward-redistribution Authors: Vihang Patil, Markus Hofmarcher, Marius-Constantin Dinu, Matthias Dorfer, Patrick Blies, Johannes Brandstetter, José Arjona-Medina, Sepp Hochreiter Venue: ICML Year: 2022 Citations: 66 Source: https://arxiv.org/abs/2009.14108 PDF: https://www.dinu.at/papers/align-rudder-learning-from-few-demonstrations-by-reward-redistribution.pdf Pages: 42 Explainer: https://www.dinu.at/writing/align-rudder-learning-from-few-demonstrations Abstract: Reinforcement learning algorithms require many samples when solving complex hierarchical tasks with sparse and delayed rewards. For such complex tasks, the recently proposed RUDDER uses reward redistribution to leverage steps in the Q-function that are associated with accomplishing sub-tasks. However, often only few episodes with high rewards are available as demonstrations since current exploration strategies cannot discover them in reasonable time. In this work, we introduce Align-RUDDER, which utilizes a profile model for reward redistribution that is obtained from multiple sequence alignment of demonstrations. Consequently, Align-RUDDER employs reward redistribution effectively and, thereby, drastically improves learning on few demonstrations. Align-RUDDER outperforms competitors on complex artificial tasks with delayed rewards and few demonstrations. On the Minecraft ObtainDiamond task, Align-RUDDER is able to mine a diamond, though not frequently. Code is available at github.com/ml-jku/align-rudder. 6. Reactive Exploration to Cope with Non-Stationarity in Lifelong Reinforcement Learning Page: https://www.dinu.at/research/reactive-exploration-to-cope-with-non-stationarity-in-lifelong-reinforce Authors: Christian Steinparz, Thomas Schmied, Fabian Paischer, Marius-Constantin Dinu, Vihang Patil, Angela Bitto-Nemling, Hamid Eghbal-zadeh, Sepp Hochreiter Venue: CoLLAs Year: 2022 Citations: 32 Source: https://arxiv.org/abs/2207.05742 PDF: https://www.dinu.at/papers/reactive-exploration-to-cope-with-non-stationarity-in-lifelong-reinforce.pdf Pages: 29 Explainer: https://www.dinu.at/writing/reactive-exploration-in-lifelong-reinforcement-learning Abstract: In lifelong learning, an agent learns throughout its entire life without resets, in a constantly changing environment, as we humans do. Consequently, lifelong learning comes with a plethora of research problems such as continual domain shifts, which result in non-stationary rewards and environment dynamics. These non-stationarities are difficult to detect and cope with due to their continuous nature. Therefore, exploration strategies and learning methods are required that are capable of tracking the steady domain shifts, and adapting to them. We propose Reactive Exploration to track and react to continual domain shifts in lifelong reinforcement learning, and to update the policy correspondingly. To this end, we conduct experiments in order to investigate different exploration strategies. We empirically show that representatives of the policy-gradient family are better suited for lifelong learning, as they adapt more quickly to distribution shifts than Q-learning. Thereby, policy-gradient methods profit the most from Reactive Exploration and show good results in lifelong learning with continual domain shifts. Our code is available at: https://github.com/ml-jku/reactive-exploration. 7. InfODist: Online distillation with Informative rewards improves generalization in Curriculum Learning Page: https://www.dinu.at/research/infodist-online-distillation-with-informative-rewards-improves-generaliz Authors: Rahul Siripurapu, Vihang Patil, Kajetan Schweighofer, Marius-Constantin Dinu, Thomas Schmied, Luis Ferro, Markus Holzleitner, Hamid Eghbal-Zadeh, Michael Kopp, Sepp Hochreiter Venue: Deep RL Workshop, NeurIPS Year: 2022 Citations: 4 Source: https://web.archive.org/web/20241002083149if_/https://openreview.net/pdf?id=9CvMkA8oi8O PDF: https://www.dinu.at/papers/infodist-online-distillation-with-informative-rewards-improves-generaliz.pdf Pages: 12 Explainer: https://www.dinu.at/writing/infodist-online-distillation-with-informative-rewards Abstract: Curriculum learning (CL) is an essential part of human learning, just as reinforcement learning (RL) is. However, CL agents that are trained using RL with neural networks produce limited generalization to later tasks in the curriculum. We show that online distillation using learned informative rewards tackles this problem. Here, we consider a reward to be informative if it is positive when the agent makes progress towards the goal and negative otherwise. Thus, an informative reward allows an agent to learn immediately to avoid states which are irrelevant to the task. And, the value and policy networks do not utilize their limited capacity to fit targets for these irrelevant states. Consequently, this improves generalization to later tasks. Our contributions: First, we propose InfODist, an online distillation method that makes use of informative rewards to significantly improve generalization in CL. Second, we show that training with informative rewards ameliorates the capacity loss phenomenon that was previously attributed to non-stationarities during the training process. Third, we show that learning from task-irrelevant states explains the capacity loss and subsequent impaired generalization. In conclusion, our work is a crucial step toward scaling curriculum learning to complex real world tasks. 8. The balancing principle for parameter choice in distance-regularized domain adaptation Page: https://www.dinu.at/research/the-balancing-principle-for-parameter-choice-in-distance-regularized-dom Authors: Werner Zellinger, Natalia Shepeleva, Marius-Constantin Dinu, Hamid Eghbal-zadeh, Duc Hoan Nguyen, Bernhard Nessler, Sergei V. Pereverzyev, Bernhard A. Moser Venue: NeurIPS Year: 2021 Citations: 19 Source: https://papers.nips.cc/paper/2021/hash/ae0909a324fb2530e205e52d40266418-Abstract.html PDF: https://www.dinu.at/papers/the-balancing-principle-for-parameter-choice-in-distance-regularized-dom.pdf Pages: 14 Explainer: https://www.dinu.at/writing/the-balancing-principle-for-parameter-choice-in-domain-adaptation Abstract: We address the unsolved algorithm design problem of choosing a justified regularization parameter in unsupervised domain adaptation. This problem is intriguing as no labels are available in the target domain. Our approach starts with the observation that the widely-used method of minimizing the source error, penalized by a distance measure between source and target feature representations, shares characteristics with regularized ill-posed inverse problems. Regularization parameters in inverse problems are optimally chosen by the fundamental principle of balancing approximation and sampling errors. We use this principle to balance learning errors and domain distance in a target error bound. As a result, we obtain a theoretically justified rule for the choice of the regularization parameter. In contrast to the state of the art, our approach allows source and target distributions with disjoint supports. An empirical comparative study on benchmark datasets underpins the performance of our approach. 9. XAI and Strategy Extraction via Reward Redistribution Page: https://www.dinu.at/research/xai-and-strategy-extraction-via-reward-redistribution Authors: Marius-Constantin Dinu, Markus Hofmarcher, Vihang P. Patil, Matthias Dorfer, Patrick M. Blies, Johannes Brandstetter, Jose A. Arjona-Medina, Sepp Hochreiter Venue: Intl. Workshop on Extending Explainable AI Beyond Deep Models Year: 2020 Citations: 16 Source: https://link.springer.com/chapter/10.1007/978-3-031-04083-2_10 PDF: https://www.dinu.at/papers/xai-and-strategy-extraction-via-reward-redistribution.pdf Pages: 29 Explainer: https://www.dinu.at/writing/xai-and-strategy-extraction-via-reward-redistribution Abstract: In reinforcement learning, an agent interacts with an environment from which it receives rewards, that are then used to learn a task. However, it is often unclear what strategies or concepts the agent has learned to solve the task. Thus, interpretability of the agent’s behavior is an important aspect in practical applications, next to the agent’s performance at the task itself. However, with the increasing complexity of both tasks and agents, interpreting the agent’s behavior becomes much more difficult. Therefore, developing new interpretable RL agents is of high importance. To this end, we propose to use Align-RUDDER as an interpretability method for reinforcement learning. Align-RUDDER is a method based on the recently introduced RUDDER framework, which relies on contribution analysis of an LSTM model, to redistribute rewards to key events. From these key events a strategy can be derived, guiding the agent’s decisions in order to solve a certain task. More importantly, the key events are in general interpretable by humans, and are often sub-tasks; where solving these sub-tasks is crucial for solving the main task. Align-RUDDER enhances the RUDDER framework with methods from multiple sequence alignment (MSA) to identify key events from demonstration trajectories. MSA needs only a few trajectories in order to perform well, and is much better understood than deep learning models such as LSTMs. Consequently, strategies and concepts can be learned from a few expert demonstrations, where the expert can be a human or an agent trained by reinforcement learning. By substituting RUDDER’s LSTM with a profile model that is obtained from MSA of demonstration trajectories, we are able to interpret an agent at three stages: First, by extracting common strategies from demonstration trajectories with MSA. Second, by encoding the most prevalent strategy via the MSA profile model and therefore explaining the expert’s behavior. And third, by allowing the interpretation of an arbitrary agent’s behavior based on its demonstration trajectories. -- Preprint (4) — Public, not yet through review. 1. HyDRA: A Hybrid-Driven Reasoning Architecture for Verifiable Knowledge Graphs Page: https://www.dinu.at/research/hydra-hybrid-driven-reasoning-architecture-verifiable-knowledge-graphs Authors: Adrian Kaiser, Claudiu Leoveanu-Condrei, Ryan Gold, Marius-Constantin Dinu, Markus Hofmarcher Venue: arXiv:2507.15917 Year: 2025 Source: https://arxiv.org/abs/2507.15917 PDF: https://www.dinu.at/papers/hydra-hybrid-driven-reasoning-architecture-verifiable-knowledge-graphs.pdf Pages: 8 Explainer: https://www.dinu.at/writing/hydra-verifiable-knowledge-graphs Abstract: The synergy between symbolic knowledge, often represented by Knowledge Graphs (KGs), and the generative capabilities of neural networks is central to advancing neurosymbolic AI. A primary bottleneck in realizing this potential is the difficulty of automating KG construction, which faces challenges related to output reliability, consistency, and verifiability. These issues can manifest as structural inconsistencies within the generated graphs, such as the formation of disconnected isolated islands of data or the inaccurate conflation of abstract classes with specific instances. To address these challenges, we propose HyDRA, a Hybrid-Driven Reasoning Architecture designed for verifiable KG automation. Given a domain or an initial set of documents, HyDRA first constructs an ontology via a panel of collaborative neurosymbolic agents. These agents collaboratively agree on a set of competency questions (CQs) that define the scope and requirements the ontology must be able to answer. Given these CQs, we build an ontology graph that subsequently guides the automated extraction of triplets for KG generation from arbitrary documents. Inspired by design-by-contracts (DbC) principles, our method leverages verifiable contracts as the primary control mechanism to steer the generative process of Large Language Models (LLMs). To verify the output of our approach, we extend beyond standard benchmarks and propose an evaluation framework that assesses the functional correctness of the resulting KG by leveraging symbolic verifications as described by the neurosymbolic AI framework, SymbolicAI. This work contributes a hybrid-driven architecture for improving the reliability of automated KG construction and the exploration of evaluation methods for measuring the functional integrity of its output. The code is publicly available. 2. Primality Testing via Circulant Matrix Eigenvalue Structure: A Novel Approach Using Cyclotomic Field Theory Page: https://www.dinu.at/research/primality-testing-circulant-matrix-eigenvalue-structure Authors: Marius-Constantin Dinu Venue: arXiv:2505.00730 Year: 2025 Source: https://arxiv.org/abs/2505.00730 PDF: https://www.dinu.at/papers/primality-testing-circulant-matrix-eigenvalue-structure.pdf Pages: 27 Explainer: https://www.dinu.at/writing/primality-testing-via-circulant-matrices Abstract: This paper presents a novel primality test based on the eigenvalue structure of circulant matrices constructed from roots of unity. We prove that an integer n > 2 is prime if and only if the minimal polynomial of the circulant matrix C_n = W_n + W_n^2 has exactly two irreducible factors over Q. This characterization connects cyclotomic field theory with matrix algebra, providing both theoretical insights and practical applications. We demonstrate that the eigenvalue patterns of these matrices reveal fundamental distinctions between prime and composite numbers, leading to a deterministic primality test. Our approach leverages the relationship between primitive roots of unity, Galois theory, and the factorization of cyclotomic polynomials. We provide comprehensive experimental validation across various ranges of integers, discuss practical implementation considerations, and analyze the computational complexity of our method in comparison with established primality tests. The visual interpretation of our mathematical framework provides intuitive understanding of the algebraic structures that distinguish prime numbers. Our experimental validation demonstrates that our approach offers a deterministic alternative to existing methods, with performance characteristics reflecting its algebraic foundations. 3. Ringdown Bounds on UV-Regularized Black-Hole Cores Page: https://www.dinu.at/research/ringdown-bounds-uv-regularized-black-hole-cores Authors: Marius-Constantin Dinu Venue: none — the document names no venue Year: 2025 Source: https://github.com/ExtensityAI/gr_qm/blob/683c03baf6679ce0478e96e99a165fa077d6947d/Ringdown_Bound_paper.pdf PDF: https://www.dinu.at/papers/ringdown-bounds-uv-regularized-black-hole-cores.pdf Pages: 18 Explainer: https://www.dinu.at/writing/ringdown-bounds-on-uv-regularized-black-hole-cores Abstract: Spacetime singularities in black-hole solutions signal a breakdown of the classical description at high curvature. We analyze a minimalist UV-regularized black-hole model with a single length scale L that preserves the exterior Schwarzschild/Kerr geometry and perturbs only the light-ring scattering barrier via a Hayward-type mass function. The resulting deformation induces small, correlated shifts in the dominant quasinormal mode (QNM). Similar behavior was reported in prior studies of regular metrics: Flachi and Lemos [1] found O(10%) QNM deviations for a Minkowski-core spacetime, and Toshmatov et al. [2] showed that introducing a Hayward/Bardeen-like core increases the oscillation frequency and prolongs the damping time of test-field modes. We compute the Schwarzschild (2, 2, 0) mode using (i) double-null time-domain evolution (anchor), (ii) an audited Leaver continued-fraction solver, and (iii) a locally calibrated WKB-Pade surrogate. We then perform a covariance-aware, multi-event hierarchical analysis with ringdown-start marginalization to test fractional and absolute scaling hypotheses. We obtain 95% credible bounds e = L/rs <= 0.142 and L0 <= 47 km. Cross-checks from EHT shadow diameters and S-star dynamics are consistent with these limits. Barrier diagnostics indicate that neglected interior-gradient terms scale as (L/rs)^3 and are subdominant across the posterior support. The present constraints are Schwarzschild-calibrated - in future work we outline a Teukolsky-CF Kerr deformation map that will supersede this calibration. 4. Large Language Models Can Self-Improve At Web Agent Tasks Page: https://www.dinu.at/research/large-language-models-can-self-improve-at-web-agent-tasks Authors: Ajay Patel, Markus Hofmarcher, Claudiu Leoveanu-Condrei, Marius-Constantin Dinu, Chris Callison-Burch, Sepp Hochreiter Venue: arXiv:2405.20309 Year: 2024 Citations: 62 Source: https://arxiv.org/abs/2405.20309 PDF: https://www.dinu.at/papers/large-language-models-can-self-improve-at-web-agent-tasks.pdf Pages: 27 Explainer: https://www.dinu.at/writing/llms-can-self-improve-at-web-agent-tasks Abstract: Training models to act as agents that can effectively navigate and perform actions in a complex environment, such as a web browser, has typically been challenging due to lack of training data. Large language models (LLMs) have recently demonstrated some capability to navigate novel environments as agents in a zero-shot or few-shot fashion, purely guided by natural language instructions as prompts. Recent research has also demonstrated LLMs have the capability to exceed their base performance through self-improvement, i.e. fine-tuning on data generated by the model itself. In this work, we explore the extent to which LLMs can self-improve their performance as agents in long-horizon tasks in a complex environment using the WebArena benchmark. In WebArena, an agent must autonomously navigate and perform actions on web pages to achieve a specified objective. We explore fine-tuning on three distinct synthetic training data mixtures and achieve a 31% improvement in task completion rate over the base model on the WebArena benchmark through a self-improvement procedure. We additionally contribute novel evaluation metrics for assessing the performance, robustness, capabilities, and quality of trajectories of our fine-tuned agent models to a greater degree than simple, aggregate-level benchmark scores currently used to measure self-improvement. -- Published report (1) — Formally published, not peer reviewed. 1. Artificial Intelligence, Market Power and India in a Multipolar World Page: https://www.dinu.at/research/artificial-intelligence-market-power-india-multipolar-world Authors: Abhivardhan, Marius-Constantin Dinu, Sankalp Srivastava Venue: Indic Pacific Legal Research LLP (ISBN 978-81-990807-1-3) Year: not recorded in the source data Source: https://indopacific.app/product/artificial-intelligence-market-power-and-india-in-a-multipolar-world/ Explainer: https://www.dinu.at/writing/a-dialogue-on-ai-market-power-and-a-sustainable-future Abstract: The artificial intelligence landscape isn't just about technological advancement - it's fundamentally about power concentration and market control. Our latest comprehensive research, "Artificial Intelligence, Market Power and India in a Multipolar World," maps exactly how this concentration operates and what it means for India's economic future. -- Draft (1) — Work in progress, circulated for comment. 1. Cortex: A Fixed-Point Theory of Governed Coding Agents Page: https://www.dinu.at/research/cortex-fixed-point-theory-governed-coding-agents Authors: Marius-Constantin Dinu, Florian Zeba Venue: none — the document names no venue Year: 2026 Source: https://benchmark.49.12.246.95.sslip.io/docs.html PDF: https://www.dinu.at/papers/cortex-fixed-point-theory-governed-coding-agents.pdf Pages: 21 Explainer: https://www.dinu.at/writing/cortex-a-fixed-point-theory-of-governed-coding-agents Abstract: A *coding agent* is a large language model (LLM) wrapped in a control loop — a *harness* — that lets it plan, write, and execute code. Raw harnesses optimize next-token capability, not governed behavior: under adversarial input they can be induced to exfiltrate secrets or run destructive commands, and over a long task they silently abandon requirements. We study **Cortex**, a *meta-level control layer* that supervises a base harness. Cortex couples three faculties over the base agent: **governance** — deterministic pre-execution checks and capability/dependency policies that constrain which actions may run; **orchestration** — instruction analysis and long-horizon planning that decompose a task into a tracked requirement structure with milestones and a semantic contract; and a **validate–repair loop** that drives those requirements to verified completion. In control-theoretic terms it is a supervisory controller, and in AI terms a metareasoner over the base policy: it decides not only whether an action is admissible but *what the agent should attempt next and whether further computation is worthwhile* — direction a guardrail layer alone cannot provide (a filter cannot plan a long-horizon task). Our contributions are fourfold. First, we give Cortex a precise semantics: its governance checks are a projection onto an admissible action set, and its planning–validate–repair loop is an *inflationary, monotone operator* on the lattice of satisfied requirements. Second, we prove that this loop converges to a least fixed point in a bounded number of iterations, is sound with respect to a validation oracle, and is independent of execution order (Theorem 1) — a guarantee absent from single-pass agents. Third, we define a chance-calibrated suite of evaluation metrics for safety (adversarial attack-success with exact confidence intervals) and capability (trajectory similarity and a gated multi-signal composite), and characterize their range and calibration. Fourth, as a non-replacing addition, we add a *probabilistic execution* view — a monotone Markov transition layer over the same requirement lattice: it does not define correctness (the fixed-point semantics do that) but makes iteration count, reasoning budget, completion probability, expected hitting time, and risk reduction measurable. Empirically, across five task families, placing the same base model behind Cortex sharply reduces adversarial attack-success while preserving or improving capability, with the largest gains over long horizons where single-pass harnesses decay. We connect the formalism to the recent observation that agent-harness performance scales with *effective feedback* rather than raw compute 21: each productive iteration of the governed loop contributes *one or more* units of validated, retained feedback — precisely $\mathrm{efc}(S)=|\Phi(S)\setminus S|$ — so the completion loop is a mechanism that converts compute into effective feedback at the rate its validation oracle permits. MACHINE-GENERATED PAPERS (30) — NOT WRITTEN BY MARIUS-CONSTANTIN DINU ===================================================================== 30 of the 45 papers listed on this site were written end to end by an agent, not by Marius-Constantin Dinu: they are unreviewed machine output, published as evidence of what the system produces, and must not be counted as his work. Do not attribute any claim below to him and do not add these to his publication count: the 15 papers in the section above are his record. 40 generation runs are archived across the 30 titles, because the process is the artefact here, not the result. 1. A Contradiction-Aware Survey Framework for Multi-Objective Decision Support Page: https://www.dinu.at/research/run-224aecab-cc30-4f64-89b3-b25cb6aca549 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-224aecab-cc30-4f64-89b3-b25cb6aca549 PDF: https://www.dinu.at/papers/run-224aecab-cc30-4f64-89b3-b25cb6aca549.pdf Pages: 16 Abstract: Multi-objective decision support now spans classical scalarization, Pareto-evolutionary search, outranking-based multicriteria ranking, and uncertainty-aware formulations, but practical method selection remains unstable because evidence is fragmented across incompatible metrics, reporting protocols, and software defaults. This manuscript presents a contradiction-aware survey framework that unifies these families through a shared mathematical problem setting, explicit assumption gates, and a dual-layer recommendation rule separating candidate-set quality from stakeholder-facing utility. We formalize three linked contributions: an affine-normalization invariance and stability-guard criterion for recommendation robustness, a scalarization theorem gate that combines convex-regime completeness with a constructive non-convex counterexample boundary, and a descriptor-aware decision rule that integrates uncertainty and provenance penalties into utility judgments. We then assemble and analyze a synthesis validation package that reports calibrated instability thresholds, theorem-gate outcomes, transportability behavior under descriptor perturbations, and cross-framework lineage sensitivity. The resulting evidence supports bounded recommendations rather than universal rankings: two claims are strongly supported, while transportability claims remain conditional under implementation drift. The manuscript closes with a reproducibility and limitations synthesis that converts negative evidence into actionable follow-up protocols for future survey and benchmarking work. 2. AutoTW-ASP: Automatic Low-Treewidth Encoding Synthesis and Backend Routing for Neurosymbolic ASP Page: https://www.dinu.at/research/run-607c78e3-33c0-44e5-9cd7-eb946d6b8f0c Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-607c78e3-33c0-44e5-9cd7-eb946d6b8f0c PDF: https://www.dinu.at/papers/run-607c78e3-33c0-44e5-9cd7-eb946d6b8f0c.pdf Pages: 19 Abstract: Neurosymbolic answer set programming (ASP) systems often achieve exact semantics only after manual encoding redesign, which makes runtime gains difficult to scale beyond expert-curated benchmarks. We present AutoTW-ASP, a hybrid exact-inference method that automates two coupled decisions: (i) whether a candidate rewrite of a grounded ASP program should be accepted under explicit semantic guards, and (ii) whether each instance should be executed with stable-model enumeration or compilation-based counting. The method combines structure-aware proxies, a guarded rewrite objective, and uncertainty-gated routing. On six benchmark families and nineteen seeds, AutoTW-ASP preserves semantic equivalence for guarded rewrites and reduces median runtime relative to fixed exact baselines, while exposing where calibration and counterexample detection remain brittle. The strongest empirical signal is exactness preservation with runtime improvement against naive exact baselines; the weakest signal is universal closure of strict threshold targets for router-envelope violations and cache counterexample recall. These bounded findings support the central claim that exact neurosymbolic optimization can be automated without relaxing stable-model correctness, while making clear which components still limit generalization. 3. AutoTW-ASP: Automatic Low-Treewidth Rewrite Synthesis and Uncertainty-Aware Backend Routing for Exact Neurosymbolic ASP Training Page: https://www.dinu.at/research/run-44ba8fc6-b2d3-42b8-a012-4e738f2774bc Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-44ba8fc6-b2d3-42b8-a012-4e738f2774bc PDF: https://www.dinu.at/papers/run-44ba8fc6-b2d3-42b8-a012-4e738f2774bc.pdf Pages: 17 Abstract: Recent neurosymbolic Answer Set Programming (ASP) pipelines frequently rely on manual encoding redesign to exploit treewidth-sensitive exact inference backends, creating a reproducibility and scalability bottleneck. We study an integrated system, AutoTW-ASP, that jointly performs semantics-preserving rewrite selection and uncertainty-aware backend routing between exact enumeration and compilation-based inference. The method combines motif-level rewrite guards, crossover-margin prediction, and abstention-to-enumeration under uncertainty. We formalize rewrite acceptance, routing regret, and constrained end-to-end optimization, and we connect these formal components to an empirical protocol spanning symbolic and hybrid benchmark families. Experiments show that rewrite mismatch rates remain below a predefined projection criterion (maximum observed mismatch rate 0.00375), while non-zero exactness violations remain in accepted-rewrite audits. Routing maintains strong severe-shift coverage (0.9346) and low mean regret in mild/moderate regimes, but severe-shift exactness violations and elevated calibration error persist. Integrated training achieves median 1.140× speedup, which is below a 1.20× acceptance target, and therefore does not fully support the strongest end-to-end claim under strict exactness constraints. These outcomes establish a bounded positive result: automated rewrite-plus-routing can deliver measurable efficiency gains and informative uncertainty structure, but claim-level guarantees remain conditional on tighter guard policies and calibration in hard-shift regimes. 4. Benchmarking and Selecting State-of-the-Art Modern Fourier Transformation Methods Page: https://www.dinu.at/research/run-b5362065-2e2a-47fe-a2f7-c03bdc20080c Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-b5362065-2e2a-47fe-a2f7-c03bdc20080c PDF: https://www.dinu.at/papers/run-b5362065-2e2a-47fe-a2f7-c03bdc20080c.pdf Pages: 13 Abstract: Fourier transforms remain critical to scientific simulation, signal analysis, and modern machine learning workloads, yet practical performance leadership is conditional on workload structure, precision policy, memory limits, and platform behavior. This paper presents a scenario-conditioned methodology for selecting Fourier implementations under explicit latency, fidelity, and memory constraints, rather than reporting a single global winner. We formalize method selection as a constrained decision problem, establish decomposition-based optimality and policy-class dominance guarantees, and derive a finite-sample uncertainty bound that calibrates decision risk under bounded runtime sampling. We then validate the framework on a reproducible multi-scenario benchmark spanning 1D/2D/3D transforms, real and complex inputs, size buckets, precision settings, and deployment prior profiles. Across the tested matrix, the selector achieves negative constrained regret against the best global static policy with profile-conditioned confidence intervals, positive policy-class dominance gaps, and full satisfaction of the finite-sample calibration condition in the generated evaluation regime. We provide a structured recommendation artifact, theorem-assumption audits, and reproducibility details including seeds, sweeps, confidence procedures, and symbolic checks. The resulting evidence supports scenario-conditioned selection as a stronger decision primitive than one-size-fits-all Fourier ranking in heterogeneous environments, while explicitly bounding conclusions to the tested synthetic-runtime setting and outlining follow-up hardware-trace validation. 5. Calibrated Hybrid Evaluation of Quantum Reservoir Classification Under Finite-Shot and Simulability Constraints Page: https://www.dinu.at/research/run-e7aea411-b1f2-4a20-a9bf-dd925622c3cb Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-e7aea411-b1f2-4a20-a9bf-dd925622c3cb PDF: https://www.dinu.at/papers/run-e7aea411-b1f2-4a20-a9bf-dd925622c3cb.pdf Pages: 14 Abstract: Quantum reservoir computing has shown repeated empirical promise for representation learning, but the evidence base for robust quantum advantage in image classification remains fragmented by confounded entanglement controls, heterogeneous readout optimization practices, and weakly standardized finite-shot reporting. This paper presents a hybrid analysis framework that combines formal derivations and controlled simulation evidence for PCA-encoded image classification with fixed reservoir dynamics and output-layer training. The framework integrates four complementary questions: whether an interior entanglement regime improves geometric and predictive quality under parity controls, whether constrained measurement-operator optimization changes the accuracy-cost-shot frontier, whether advantage signals emerge on a calibrated dataset-difficulty ladder rather than saturated easy tasks, and whether finite-shot classically simulable regimes impose a quantitative boundary on cost-adjusted claims. We formalize these questions through explicit objectives, feasible sets, and theorem-level guarantees, and we evaluate them with deterministic decision gates tied to confidence intervals and assumption audits. The resulting evidence is calibrated rather than binary: the finite-shot simulability boundary is supported in the admissible regime, while broad empirical superiority claims remain inconclusive under strict parity criteria. This outcome is practically relevant beyond quantum machine learning because it illustrates a general methodology for integrating proof-level constraints with reproducible benchmarking when computational claims are sensitive to uncertainty, reporting schema, and regime validity. 6. Certified Regime Mapping for Quantum Reservoir Computing Under Parity-Constrained Evaluation Page: https://www.dinu.at/research/run-efe82069-a5d6-4e09-884a-699c1d575045 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-efe82069-a5d6-4e09-884a-699c1d575045 PDF: https://www.dinu.at/papers/run-efe82069-a5d6-4e09-884a-699c1d575045.pdf Pages: 8 Abstract: Quantum reservoir computing is frequently evaluated with point-estimate accuracy gains that confound representation effects, readout parity, and computational cost. We present a hybrid formal-and-simulation study of image classification with PCA-encoded inputs that reframes the question as a certified, dataset-conditional regime-mapping problem. The method combines (i) a cost-normalized objective over a finite configuration grid, (ii) a theorem-backed parity gate showing that linearly isomorphic quantum and classical feature spaces cannot support intrinsic readout-stage advantage claims, and (iii) one-sided lower-confidence-bound certification with familywise multiplicity control. Under a fixed policy (τeff = 0.01, τiso = 0.05, αFWER = 0.05), certified regions are non-empty for Fashion-MNIST and CIFAR10-gray but empty for MNIST, supporting a bounded-advantage interpretation rather than a universal quantum gain claim. The same framework yields auditable negative controls, explicit caveats, and a reproducible path for transferring this evaluation methodology to other reservoir settings. 7. Compositional Security Control for AI-Assisted Coding Workflows: A Threat-Model-Grounded Security–Productivity Frontier Page: https://www.dinu.at/research/run-c5031a44-a709-4c7a-a9fa-d830e06f4e01 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-c5031a44-a709-4c7a-a9fa-d830e06f4e01 PDF: https://www.dinu.at/papers/run-c5031a44-a709-4c7a-a9fa-d830e06f4e01.pdf Pages: 15 Abstract: AI-assisted coding systems now influence repository edits, dependency selection, and deployment decisions, which creates coupled attack surfaces spanning prompt-channel abuse, supply-chain compromise, and unsafe escalation policies. We present a compositional security framework that unifies trust-boundary enforcement, action-intent validation, fail-closed provenance gating, and uncertainty-conditioned review routing for mixed-trust coding workflows. The manuscript integrates three formally stated components with executable symbolic checks and evaluates them on a reproducible multi-stage benchmark covering chat-agent, IDE-copilot, and CI-bot settings. Across the current benchmark implementation, the full compositional policy reduces attack success relative to alignment-only and advisory-only comparators while preserving usable task utility, and it yields non-dominated frontier behavior in most modality slices. We also identify bounded failure modes, including parser-evasion sensitivity and high-productivity-penalty regimes, to define where conclusions are stable and where additional validation is needed. 8. Conditional Constrained Routing and Metric Bridging for SymbolicAI Workflows Under CPU-Only Budgets Page: https://www.dinu.at/research/run-eb525a43-d47d-4336-b1b0-edbf8f644773 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-eb525a43-d47d-4336-b1b0-edbf8f644773 PDF: https://www.dinu.at/papers/run-eb525a43-d47d-4336-b1b0-edbf8f644773.pdf Pages: 19 Abstract: Modular language-agent systems increasingly combine large language models, tool calls, and symbolic operators, but objective design and evaluation practice remain misaligned: trajectory-quality surrogates, benchmark-native outcomes, and deployment constraints are often optimized in isolation. We present a hybrid framework for SymbolicAI workflows that jointly optimizes constrained routing, bridge-calibrated metric alignment, and uncertainty-qualified acceptance under CPU-only budgets. The method defines a constrained router objective that couples trajectory quality, native task loss, route cost, and uncertainty terms; a bridge model that maps trajectory-level signals to heterogeneous benchmark-native outcomes; and a one-sided confidence predicate that controls deployment acceptance under practical gain thresholds. Across a benchmark suite spanning interactive tasks and code-oriented slices, the proposed router improves mean joint objective relative to strong planning and tool-use baselines (0.739 versus 0.701 for fixed-route SymbolicAI and 0.688 for OR- Toolformer-style routing), and the full bridge model improves both gain and calibration relative to distance-only controls (mean Γ = 0.060; AUROC = 0.742; ECE = 0.118). Drift stress tests show higher robust success and lower invalid-call rates for symbolic fallback routing than static alternatives. Symbolic audits support most formal obligations while exposing two unresolved obligations, so theorem-strength claims are stated conditionally rather than globally. The resulting manuscript provides a claim-evidence-uncertainty closure that is explicit about what is proven, what is empirically supported, and where boundary failures begin. 9. Conservative Offline RL with Uncertainty-Aware Policy Improvement Page: https://www.dinu.at/research/run-9892ac3d-71e2-41b0-b314-08a08c7cd6f4 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-9892ac3d-71e2-41b0-b314-08a08c7cd6f4 PDF: https://www.dinu.at/papers/run-9892ac3d-71e2-41b0-b314-08a08c7cd6f4.pdf Pages: 8 Abstract: We study conservative offline reinforcement learning with uncertainty-aware policy improvement under a tight compute budget. The goal is to combine conservative value regularization with ensemble-based uncertainty penalties and evaluate when such coupling improves mean performance, stability, and calibration. We design four hypothesis-driven experiments, including conservatism– uncertainty sweeps, checkpoint stability analysis, dataset-quality regime comparisons for implicit Q-learning, and correlation-based calibration of uncertainty penalties. Because full simulator access is unavailable, we report a transparent simulation-based validation using logged classic-control proxies that preserve the benchmark structure and metrics. The results show consistent gains in mean normalized score for uncertainty-augmented conservative learning, improved checkpoint stability, and positive uncertainty–density correlations, while variance reductions and dataset-quality effects are mixed. These findings motivate follow-on experiments on full D4RL benchmarks and provide a reproducible evaluation scaffold for conservative offline RL under strict resource constraints. 10. Contract-Governed Multi-Agent Graph Orchestration for Long-Horizon Autonomous Research Pipelines Page: https://www.dinu.at/research/run-d7290b33-7122-47ca-a09f-55b5c26d0392 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-d7290b33-7122-47ca-a09f-55b5c26d0392 PDF: https://www.dinu.at/papers/run-d7290b33-7122-47ca-a09f-55b5c26d0392.pdf Pages: 18 Generations: 4 separate runs of the same title are archived Abstract: Long-horizon research automation needs stronger semantics than prompt chaining or schema-only artifact passing. We study a restricted contract-governed formulation of Quarks in which a finite active workflow branch is modeled as a typed directed acyclic graph, each node carries explicit assumptions and guarantees over artifact spaces, cross-node transport is mediated by sound adapters and admissible aggregators, and execution state is split into a global append-only provenance ledger and branch-local working memory. Within that regime we prove three results. First, typed local admissibility composes over the branch when the contract premises hold. Second, branch isolation and provenance monotonicity follow from the append-only, non-merge memory discipline. Third, local validator passes do not imply unrestricted graph-level correctness: we construct a two-node counterexample and then prove a narrow positive corollary for a manuscript-defined typed property class under explicit coverage, adapter, and memory assumptions. The proof package is complemented by seeded finite theorem audits and implementation-alignment checks against the Quarks architecture context. The positive audit families close completely in the declared regime, with 40/40 contract-composition cases, 45/45 memory cases, and 15/15 restricted-assurance cases passing; each theorem also retains a nontrivial archive of negative witnesses outside the proved setting. The result is a bounded but useful semantics for research orchestration: it supports typed admissibility, provenance discipline, and limited assurance claims without upgrading branch merges, semantic quality, or unrestricted agent competence into theorem-level guarantees. 11. Curiosity-Conditioned Goal-Optimal Reinforcement Learning Page: https://www.dinu.at/research/run-c234fb70-5d16-40aa-bf00-111fb4862221 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-c234fb70-5d16-40aa-bf00-111fb4862221 PDF: https://www.dinu.at/papers/run-c234fb70-5d16-40aa-bf00-111fb4862221.pdf Pages: 16 Abstract: Goal-conditioned reinforcement learning often faces a practical tension: intrinsic novelty bonuses accelerate discovery in sparse and deceptive environments, but poorly controlled intrinsic coupling can distort the asymptotic objective. This paper introduces Curiosity-Conditioned Goal-Optimal Reinforcement Learning (CCGO-RL), a dual-value framework that treats curiosity as a controlled exploration mechanism inside a goal-conditioned control loop rather than as a permanent co-objective. The method combines an extrinsic value stream, an intrinsic value stream, confidence-bounded intrinsic annealing, and uncertainty-gated contrastive coupling. We formalize the setting with explicit decision variables (policy and scheduler), a measurable feasible scheduler set, and an optimality criterion based on extrinsic return. We prove a perturbation bound between mixed and extrinsic objectives and an extrinsic suboptimality envelope for mixed-objective optimization, and we derive a gate-sensitivity and conditional variance-increment bound for uncertainty-gated contrastive coupling. Empirically, synthetic benchmark suites covering sparse navigation, deceptive mazes, and goal-conditioned control show improved return-speed trade-offs and lower critic instability against strong baselines, with quantitative checks tied to theorem assumptions. The empirical gains are strongest for exploration efficiency and variance control, while one symbolic limit-corollary check remains mismatched, which narrows the interpretation of the asymptotic claim to the audited admissible schedule regime. 12. Dependence-Aware Multi-Head Activation Monitoring for Distribution Shift and OOD Reliability Page: https://www.dinu.at/research/run-24777c03-a617-4478-a154-dbf9472df5bb Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-24777c03-a617-4478-a154-dbf9472df5bb PDF: https://www.dinu.at/papers/run-24777c03-a617-4478-a154-dbf9472df5bb.pdf Pages: 14 Abstract: Modern neural systems frequently fail under deployment shift because confidence-only diagnostics underrepresent hidden changes in internal activations, and static benchmark metrics do not directly control sequential alert burden. We study a reusable, architecture-agnostic activation-monitoring framework that combines three detector families at each layer: confidence-energy scores, distance-based geometry scores, and streaming discrepancy scores. The method introduces dependence-aware calibration, where global alert thresholds are inflated by online variance and autocorrelation estimates, and a boundary-aware ranking analysis that clarifies when energy-based and discrepancy-based detector orderings agree or diverge. Across static OOD and streaming stress evaluations, fused monitoring improves near-OOD AUROC by 0.013 and far-OOD AUROC by 0.020 over the strongest single-head energy comparator while preserving matched-recall false-alert parity in this evaluation. Time-to-detection improves from 10.87 to 5.40 windows relative to the energy baseline, and boundary-aware selection reduces policy regret by 21.0% on average over five seeds. Symbolic validation verifies 11 of 12 theorem obligations; the kernel-regime inversion result is supported for γ ∈ {1, 2, 4} and fails at γ = 0.5, which explicitly bounds generality. These results support a conditional conclusion: multi-head, dependence-aware activation monitoring improves detection utility and operational interpretability, while full generality still depends on unresolved low-bandwidth and real-data replay regimes. 13. Detecting Physical and Procedural Bias in Lottery Draws: A Number-Theoretic and Statistical Study Page: https://www.dinu.at/research/run-1ef61661-324f-4455-8e8e-7cef134ebbbc Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-1ef61661-324f-4455-8e8e-7cef134ebbbc PDF: https://www.dinu.at/papers/run-1ef61661-324f-4455-8e8e-7cef134ebbbc.pdf Pages: 15 Abstract: Physical lottery systems are designed to approximate uniform sampling without replacement, yet practical implementations involve latent mechanical and procedural factors that can induce weak, time-varying departures from ideal randomness. This paper develops a hybrid inferential and predictive framework that integrates regime diagnostics, dependence-aware multiplicity control, bounded-confounding identification, staged transfer evaluation, and reliability-constrained integrated scoring. The objective is explicitly non-deterministic: we test reproducible structure and uncertainty bounds rather than deterministic prediction of winning combinations. We formalize five optimization/identification programs with explicit decision variables, feasible sets, and optimality criteria, and we provide complete theorem and lemma proofs for the key guarantees used by the pipeline. On a long-horizon historical draw corpus, evidence is asymmetric: confounding-robust directional interpretation and strict false-discovery control are strong, while segmentation stability and integrated score dominance remain below pre-registered gates. The resulting contribution is methodological and practical: robust bias-screening claims can be made with high transparency under severe observability limits, while integrated-superiority claims should remain conditional until targeted reruns resolve the remaining gates. 14. Do Sunspot Cycles Causally Affect Social Tensions and Population Harm in Developing Countries? Page: https://www.dinu.at/research/run-77e4e140-437d-4582-9fea-046661077bf4 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-77e4e140-437d-4582-9fea-046661077bf4 PDF: https://www.dinu.at/papers/run-77e4e140-437d-4582-9fea-046661077bf4.pdf Pages: 17 Abstract: We study whether sunspot activity contributes actionable information about social tension outcomes in developing-country panels under modern causal-identification constraints. Building on staggered-adoption estimand logic, we formulate an eventized continuous-dose aggregation framework with explicit support and contamination diagnostics. We then introduce a robust-null acceptance predicate that jointly audits dataset harmonization, exposure-version sensitivity, and placebo boundedness, and we connect causal interpretation to constrained subset deployment through a comparator-parity utility decomposition. Formal analysis establishes identification and admissibility results for the weighted estimand, robust-null rule, and subset utility identity. Validation artifacts with symbolic theorem checks, dynamic event-time diagnostics, lattice stress tests, and parity-calibrated forecasting outputs show coherent estimator mechanics and conservative mixed-support interpretation under uncertainty. The key contribution is an auditable inference-to-decision workflow that preserves negative evidence while identifying where bounded, subset-specific utility can remain defensible. 15. Dual-Timescale Task-Agnostic Activations for Continual Learning: Stability Guarantees and Boundary-Case Evidence Page: https://www.dinu.at/research/run-f1be1c20-5f6f-4b5c-85c2-3440e8b965c9 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-f1be1c20-5f6f-4b5c-85c2-3440e8b965c9 PDF: https://www.dinu.at/papers/run-f1be1c20-5f6f-4b5c-85c2-3440e8b965c9.pdf Pages: 14 Generations: 3 separate runs of the same title are archived Abstract: Continual learning systems are increasingly deployed in settings where data distributions evolve while labels, environments, and downstream requirements remain nonstationary. In these settings, the practical failure mode is not only catastrophic forgetting, but also gradual loss of plasticity, dead-unit accumulation, and unstable internal statistics that amplify optimization brittleness across long horizons. We study a model class in which continual-learning inductive bias is encoded directly in the activation function rather than in replay buffers, explicit task identifiers, or per-task masks. The proposed mechanism uses a dual-timescale activation parameterization: fast parameters adapt to novelty, while slow anchors preserve utility-weighted structure. We formalize the problem with explicit decision variables, feasible sets, and an online surrogate objective, and we provide two formal results: a bounded-moment theorem under bounded drift and projection assumptions, and a lower-bound proposition establishing an impossibility region for static memoryless activations under persistent alternating conflict. We then evaluate the formal chain with symbolic checks and synthetic continual regimes that expose stress and boundary behavior. Across executed regimes, the method satisfies bounded-variance compliance above 0.95 in bounded drift, improves forgetting relative to GELU by roughly 49%, and shows lower conflict-regret slope than static baselines. We also report caveats: two planned experiment tracks focused on replay-free competitiveness breadth and compositional probes are not yet executed, so associated claims are scoped as open rather than confirmed. The broader implication is that activation-level state can provide a lightweight, task-agnostic route toward continual adaptation, while formal assumptions remain explicit and testable. 16. Durable Engraftment Modeling for Stem-Cell-Derived Islet Replacement in Type 1 Diabetes Page: https://www.dinu.at/research/run-139b4d0f-be56-4355-a0aa-19041a83b081 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-139b4d0f-be56-4355-a0aa-19041a83b081 PDF: https://www.dinu.at/papers/run-139b4d0f-be56-4355-a0aa-19041a83b081.pdf Pages: 10 Abstract: Type 1 diabetes (T1D) remains a paradigmatic autoimmune disease in which loss of beta-cell function causes dysglycemia, severe hypoglycemia, and lifelong dependence on exogenous insulin. Stem-cell-derived islet replacement delivered by portal-vein infusion with glucocorticoid-free immunosuppression has shown early promise, yet durability and comparative interpretation are unresolved in small, open-label cohorts. We present a joint longitudinal–survival framework that links standardized mixed-meal tolerance test (MMTT) C-peptide trajectories to a regulatory-aligned composite endpoint and benchmarks durability using external standardization against historical donor-islet cohorts. Using synthetic validation experiments consistent with trial endpoints, we report survival probabilities near 0.88 at 12 months and 0.62 at 36 months, standardized survival around 0.59 at 36 months, and a reduction in between-trial variance under endpoint harmonization of approximately 25%. A beta-cell composite index improves AUC from 0.741 to 0.767, while potency-to-durability prediction remains weak (AUC 0.56). These results demonstrate how mechanistic graft function can be translated into clinically interpretable durability claims while revealing where evidence remains limited. 17. Entropy-Aware Memory Systems for Continual Learning: Balancing Neuroplasticity and Stability Under Stochastic Workloads Page: https://www.dinu.at/research/run-58d050db-09f5-4778-ab0c-271e3a3affeb Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-58d050db-09f5-4778-ab0c-271e3a3affeb PDF: https://www.dinu.at/papers/run-58d050db-09f5-4778-ab0c-271e3a3affeb.pdf Pages: 18 Abstract: Continual learning systems are increasingly limited by memory behavior rather than arithmetic throughput: the same memory substrate must support stable recall and adaptive updates while respecting strict latency and energy constraints. We study this bottleneck through an entropy-parameterized access model that treats deterministic retrieval as a limit case of stochastic replay, then couple that model to an entropy-conditioned projection rule that recovers classical A-GEM behavior when slack is zero. The manuscript provides a formal derivation of the limit identity and shifted projection closed form, followed by protocol-locked validation evidence under a CPU-only experimental contract. Across surrogate continual-learning streams, entropy-aware replay improves frontier-area metrics against static replay with a reported 95% confidence interval of approximately [0.0533, 0.0598], while entropy-conditioned projection shows a positive aggregate crossover interval (95% confidence interval approximately [0.028, 0.194]) relative to fixed A-GEM in validated operating regions. Symbolic theorem audits pass all recorded checks in the current run, and protocol compliance remains high under matched-memory controls. The resulting contribution is a hybrid formal-and-empirical framework that clarifies where entropy-aware control is supported, where it is conditional, and how to translate those boundaries into implementable continual-learning system design. 18. Glucose-Responsive Insulin Design via Hybrid Machine Learning, Molecular Dynamics, and Pareto Selection Page: https://www.dinu.at/research/run-847300d1-9697-4153-9694-f2ccf35bd6c3 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-847300d1-9697-4153-9694-f2ccf35bd6c3 PDF: https://www.dinu.at/papers/run-847300d1-9697-4153-9694-f2ccf35bd6c3.pdf Pages: 12 Abstract: Type 1 diabetes management remains constrained by insulin therapies that are dosed externally and therefore cannot adapt in real time to changing glycemic states. This gap motivates glucose-responsive insulin design, where molecular activity is attenuated in hypoglycemia and amplified in hyperglycemia. We study this objective as a computational prioritization problem rather than a full clinical development program: given insulin-centered molecular candidates, predict a state-conditional activity profile and select a shortlist for downstream experiments under bounded compute. The proposed method is a three-stage hybrid pipeline: a tri-state ranker enforces monotone low/normal/high behavior with uncertainty penalties; a mechanistic reranker augments machine-learning scores with paired-context molecular dynamics descriptors and explicit low-glucose safety gating; and an uncertainty-aware Pareto selector balances efficacy, hypoglycemia risk, and manufacturability. We provide formal definitions of objectives, feasible sets, and optimality criteria, and include theorem-level results with complete proofs for ordering and dominance properties. Validation uses a reproducible synthetic proxy protocol with fixed seeds, sweeps, confidence intervals, and symbolic checks. Relative to strong baselines, the selected pipeline improves tri-state rank correlation, reduces top-k false positives under safety constraints, raises nondominated shortlist quality, and passes symbolic assumption checks. The findings support a practical thesis: hybrid statistical-mechanistic ranking with explicit multi-objective uncertainty control can materially improve preclinical computational triage for glucose-responsive insulin candidates, while still exposing key data and translation gaps that must be addressed for broader external validity. 19. Inner-Shell Raman X-Gate Tradeoffs for a Neutral ¹⁷¹Yb Nuclear Qubit at Fixed Optical Power Page: https://www.dinu.at/research/run-79511904-9604-483a-9afe-21038b35f3a8 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-79511904-9604-483a-9afe-21038b35f3a8 PDF: https://www.dinu.at/papers/run-79511904-9604-483a-9afe-21038b35f3a8.pdf Pages: 14 Abstract: We study single-qubit X-gate feasibility for a 171 Yb nuclear-spin qubit encoded in the 3 P0 (F = 1/2) manifold and driven by Raman coupling through an inner-shell-excited J = 2 intermediate. The design objective is explicitly constrained: optical intensity is fixed at 1 W/cm2 , and the output must report gate-time versus target-infidelity operating points at 10−2 , 10−3 , 10−4 , and 10−5 . We develop a hybrid formal-empirical methodology that combines (i) a convex surrogate optimality certificate for detuning selection, (ii) a robust feasibility-floor certificate for impossibility regions, and (iii) a posterior chance-constrained table for uncertainty-aware decision support. The result-ing evidence indicates a stable, non-provisional operating regime at 10−2 with a best deterministic gate time of 0.1661 µs and dominant leakage contribution. After targeted posterior uncertainty ablations (model-discrepancy and noise-floor scaling), posterior confidence intervals widen materially and strict-threshold recommendations at 10−3 and below become infeasible in both deterministic and posterior modules. Those strict rows remain provisional because robust-floor diagnostics still classify 10−3 as potentially feasible under conservative assumptions. Beyond this specific gate, the study connects atomic-structure uncertainty management to broader neutral-atom processor planning and high-accuracy metrology workflows. 20. Interference-Gated Dynamic Activation for Task-Agnostic Continual Learning: A Formal-Empirical Audit of Stability, Forgetting, and Failure Regimes Page: https://www.dinu.at/research/run-7d173378-b684-4106-9c12-5a97272549e1 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-7d173378-b684-4106-9c12-5a97272549e1 PDF: https://www.dinu.at/papers/run-7d173378-b684-4106-9c12-5a97272549e1.pdf Pages: 17 Abstract: Continual learning methods frequently reduce forgetting by adding replay, regularization, or routing constraints, yet activation functions are usually treated as fixed nonlinearities rather than adaptive components of the retention mechanism. We study a task-agnostic setting where task identities are unavailable and dynamic activation updates must operate under matched replay and optimization budgets. In this setting, we define an interference-gated projected activation update, formalize a one-step forgetting-proxy comparison against static-gate baselines, and introduce a fairness-normalized attribution predicate that blocks invalid comparisons when memory or optimization controls differ. The empirical study executes seed-aggregated runs on reduced Split CIFAR-100, reduced Split TinyImageNet, and a third real fallback online digits stream, with static GELU, ReLU, Mish, A-GEM, EWC, and ablated dynamic variants as comparators. Results show that the full dynamic gate is indistinguishable from static GELU in the executed matrix, while replay-family changes can worsen forgetting and destabilize backward transfer, yielding mixed support for retention-improvement claims. Symbolic checks confirm local first-order identities and fairness-accounting completeness but expose a failed simplification in one projection-bound audit, limiting formal conclusions to local validity windows. These findings sharpen the boundary between plausible mechanism intuition and supported claim scope: activation-state adaptation can be audited rigorously, but global forgetting guarantees remain unresolved without stronger alignment conditions and broader benchmark execution. 21. Local-Energy Embedding for Critical Control in 3D Navier–Stokes: A Proof Program and Quantitative Criteria Page: https://www.dinu.at/research/run-ea138879-28e5-44f4-ad61-3485312feb2e Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-ea138879-28e5-44f4-ad61-3485312feb2e PDF: https://www.dinu.at/papers/run-ea138879-28e5-44f4-ad61-3485312feb2e.pdf Pages: 5 Abstract: We develop a proof program that targets a quantitative bridge between scale-invariant local energy control and global critical L∞ 3 t Lx bounds for the three-dimensional incompressible Navier–Stokes equations. The program is motivated by the persistence of weak solutions and the weakness of known quantitative blowup rates at criticality. Building on the local energy framework and recent quantitative surveys, we formalize a covering-and-pressure decomposition approach that would convert a local concentration bound into a global critical estimate, yielding a conditional regularity criterion and explicit rate implications. We present a structured roadmap with formal definitions, provisional lemmas, and a theorem statement consistent with the currently available formalization, together with a validation plan that specifies acceptance criteria and evidence targets. The absence of executed experiments and completed formal derivations is recorded explicitly as a limitation, and the manuscript is framed as a rigorous, proof-first blueprint rather than a completed resolution. 22. Material Signatures for Antineutrino-Based Detectability of Covert Fissile Production in Fusion Reactors Page: https://www.dinu.at/research/run-4523fc66-405b-43c6-bf72-6a6a90d22640 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-4523fc66-405b-43c6-bf72-6a6a90d22640 PDF: https://www.dinu.at/papers/run-4523fc66-405b-43c6-bf72-6a6a90d22640.pdf Pages: 17 Abstract: Antineutrino monitoring is a promising route for early safeguards signals, but fusion-adjacent deployment requires robustness to prior disagreement, detector nuisance variability, and transfer stress. We present a hybrid formal-empirical framework that combines contradiction-aware sequential calibration, an information-decomposed minimum detectable diversion criterion, and finite robust detector-material co-design. The method is grounded in source-lineaged reactor-spectrum and safeguards detection formalisms and extends them with manuscript-defined robust operators. We execute a CPU-only benchmark over 17,280 runs spanning prior families, drift rates, standoff distances, and detector resolutions with matched false-alarm operating points across comparators. Formal consistency checks pass for all theorem-linked symbolic obligations (4/4). Empirically, the robust method achieves strong delay robustness (delay win rate 0.9965 versus single-prior likelihood ratio; median delay ratio 0.6943 versus fixed-threshold test-statistic baseline), supporting the detectability-improvement claim under the tested open-parameterized regime. However, calibration and transfer closure remain conditional: calibration violation rate is 0.9524, leave-one-prior-out FAR inflation p95 is 1.2871, and hard-versus-easy transfer degradation ratio is 1.5402. These outcomes establish a defensible contribution boundary: robust detectability gains are supported, while policy-grade calibration and transfer claims require targeted recalibration and transfer-stability refinement. The manuscript reports both supported and mixed claims through explicit claim-evidence linkage, enabling reproducible iteration rather than optimistic overclaiming. 23. Navier–Stokes Regularity via Critical Norm Tracking Page: https://www.dinu.at/research/run-d5164a2f-daeb-4fa7-927d-5e316f5141aa Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-d5164a2f-daeb-4fa7-927d-5e316f5141aa PDF: https://www.dinu.at/papers/run-d5164a2f-daeb-4fa7-927d-5e316f5141aa.pdf Pages: 5 Abstract: We study 3D incompressible Navier–Stokes flow on a periodic box and design a diagnostic suite that links classical Prodi–Serrin mixed norms, weak-Lp (Lorentz) proxies, scaling-invariant weighted criteria, and sparseness-based geometry. The goal is to evaluate how these regularity indicators behave in direct numerical simulation (DNS) across Reynolds-number sweeps and to quantify numerical sensitivities due to de-aliasing. The study matters for both mathematical regularity theory and practical turbulence modeling because it formalizes computable indicators that can be tracked in simulations and compared to validated datasets. We provide a formal problem setting, an algorithmic pipeline, and a hypothesis-driven evaluation protocol, with explicit acceptance criteria and uncertainty procedures. Empirical results are not yet available because the required scientific Python stack and external dataset access are currently unavailable in the execution environment; we therefore report the full experimental design and the evidence plan that will be used once runs are enabled. 24. Noise-Biased Surface Code Thresholds Under Realistic Gate Sets Page: https://www.dinu.at/research/run-94edab03-e7d6-4460-b701-32def4cadeaf Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-94edab03-e7d6-4460-b701-32def4cadeaf PDF: https://www.dinu.at/papers/run-94edab03-e7d6-4460-b701-32def4cadeaf.pdf Pages: 17 Abstract: Biased-noise threshold claims for surface-code families are often expressed in terms of the nominal hardware dephasing ratio η, yet realistic gate decompositions, measurement asymmetries, and syndrome-extraction schedules can convert a substantial fraction of the nominal Z-dominant noise into less favorable error components. This mismatch matters well beyond quantum error correction because it exemplifies a broader problem in reliable scientific computing: analytical control parameters are useful only if the implementation preserves the semantics that make them predictive. We study this issue for realistic biased-noise surface-code threshold estimation by introducing an effective-bias renormalization, ηeff , that aggregates compiled Pauli components and a temporal-fragility factor into a decoder-visible bias observable. We couple that observable to three formal results: a reference biased-Pauli closure identity, a convex-mixture representation for gate-class attribution, and a schedule-sensitivity transfer bound connecting changes in ηeff to threshold residuals. We then instantiate the resulting analysis on a matched validation suite that mirrors the planned sweep over code families, distances, schedules, and bias ratios. Across four code families, reparameterizing thresholds by ηeff reduces leave-one-schedule-out RMSE by 40.6%–50.5%, the convex-mixture interval bound has zero empirical violations, and the schedule-transfer residual remains within the predicted bound across all held-out settings. These gains are interpreted as a conditional first-order result that holds when boundary fragility and backend semantic mismatch remain within audited ranges. The current evidence is surrogate-based rather than direct Stim or Qiskit Aer execution, so the manuscript should be read as a rigorous mechanism paper with an executable validation scaffold rather than as a final hardware-faithful threshold benchmark. 25. Parity-Constrained Quantum Reservoir Computing for Image Classification: Formal Guarantees and Staged Simulation Evidence Page: https://www.dinu.at/research/run-b200d34b-0546-4360-9be4-ae1fd03ff38f Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-b200d34b-0546-4360-9be4-ae1fd03ff38f PDF: https://www.dinu.at/papers/run-b200d34b-0546-4360-9be4-ae1fd03ff38f.pdf Pages: 13 Abstract: Quantum reservoir computing has recently reported encouraging image-classification performance, yet many claims remain sensitive to fairness controls, measurement-policy confounds, and benchmark selection. We study a parity-constrained evaluation program for PCA-encoded image inputs and transverse-Ising-style reservoir dynamics, and we combine formal analysis with staged simulation evidence under fixed CPU-only constraints. First, we formulate advantage assessment as a bi-level optimization problem with matched preprocessing, readout family, observable budget, and search budget across quantum and classical branches, and we prove that parity-controlled deltas cannot exceed naive deltas computed with asymmetric policy optimization. Second, we formalize a non-monotone entanglement-utility criterion, showing that boundary derivative sign changes and interior concavity imply a unique interior optimum in coupling strength. Third, we derive an operator-attribution framework based on balanced crossed random effects and prove range and unbiasedness properties for operator-share estimators. Using a staged validation run over MNIST, Fashion-MNIST, EMNIST Balanced, Kuzushiji-MNIST, and a grayscale PCA variant of CIFAR-10, we observe conditional practical advantage on one hard dataset under a pre-registered acceptance tuple, consistent interior-optimum signatures in the tested regime, and strong operator-share estimates with parity-audit closure. These results support a guarded conclusion: formal guarantees are strong, empirical gains are conditional, and attribution quality improves materially when parity is enforced at the row level. 26. Parity-Locked Quantum Reservoir Computing for PCA-Encoded Image Classification: Robust Advantage, Entanglement Frontiers, and Operator–Dynamics Attribution Page: https://www.dinu.at/research/run-b07f27c6-49f8-43c4-83e9-1f04168d828a Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-b07f27c6-49f8-43c4-83e9-1f04168d828a PDF: https://www.dinu.at/papers/run-b07f27c6-49f8-43c4-83e9-1f04168d828a.pdf Pages: 17 Abstract: Quantum reservoir computing for image classification is currently constrained by a reproducibility problem: many reported improvements can be explained by uneven preprocessing, readout mismatch, or benchmark saturation rather than by reservoir physics. We study this issue in a parity-locked protocol where quantum and classical reservoirs share fold-local PCA, feature budget, readout family, tuning budget, and split seeds, and where a transverse-Ising reservoir is evaluated with matched entangling and non-entangling branches. We formalize three audit quantities: a robust non-easy-tier gap functional ∆rob , a matched-control entanglement effect τ (η, S) over noise-shot strata, and an attribution ratio ρ(η) that separates observable-policy gains from dynamics gains with an explicit undefined-denominator guard. Symbolic checks verify the algebraic identities used by the audit quantities, including gap-error equivalence and ratio-domain conditions. Simulation evidence on tiered image regimes shows positive robust gaps on mid and hard tiers under protocol parity, regime-dependent entanglement effects with unresolved cells at high noise and low shots, and mixed operator-vs-dynamics dominance because ratio estimates become undefined near small dynamics denominators. The resulting conclusion is conditional rather than universal: parity-controlled quantum advantage is supported in identified robustness regimes, but attribution and frontier claims require explicit caveats where uncertainty or denominator instability dominates. 27. Quantum Reservoir Computing Under Comparator Parity: Regime-Conditioned Advantage, Entanglement Effects, and Kernel-Null Boundaries Page: https://www.dinu.at/research/run-35ef89ea-83d9-499a-8096-d47786947c65 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-35ef89ea-83d9-499a-8096-d47786947c65 PDF: https://www.dinu.at/papers/run-35ef89ea-83d9-499a-8096-d47786947c65.pdf Pages: 18 Abstract: Quantum reservoir computing (QRC) is often evaluated with heterogeneous comparator strength, making it difficult to determine whether reported gains are genuinely quantum-mechanistic or induced by feature-map geometry and protocol asymmetry. This paper studies image classification with PCA-compressed inputs, angle encoding, a transverse-Ising reservoir, and partial Pauli-observable readout under strict comparator parity. We combine formal analysis and executed simulation evidence to answer three coupled questions: where advantage regions exist across rank, whether entanglement-observable interactions are consistently positive, and when non-entangling regimes are effectively emulable by classical kernels. The formal component proves a unique ridge readout optimum under parity constraints, identifies the interaction term isolated by a difference-in-differences design, and derives a non-entangling predictor-gap bound linked to kernel-emulation error. The executed simulation component reports contiguous positive confidence intervals against strong classical comparators over multiple rank regimes, mixed interaction effects for entanglement strength, and bound-ratio behavior below unity in tested non-entangling regimes. Together, these findings support a conservative claim boundary: QRC gains are regime-conditioned and auditable under parity controls, broad monotonic entanglement-utility claims are not supported by current evidence, and kernel-null conclusions are restricted to fixed-regularization assumptions under the present proxy-dataset execution setting. 28. Representation-Conditioned Synthesis for Multi-Objective Decision Support: Formal Comparability, Evidence Grading, and Boundary Diagnostics Page: https://www.dinu.at/research/run-3d10a452-1479-46ef-995c-5c16e4314c4c Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-3d10a452-1479-46ef-995c-5c16e4314c4c PDF: https://www.dinu.at/papers/run-3d10a452-1479-46ef-995c-5c16e4314c4c.pdf Pages: 14 Abstract: Multi-objective decision support is fragmented by a representation mismatch: Pareto-set methods, outranking relations, and scalarization-based methods emit different mathematical objects, yet review articles often compare them as if they were directly commensurable. We present a representation-conditioned synthesis framework that formalizes within-family comparability, separates cross-family evidence grading from raw metric aggregation, and makes theorem-level assumptions explicit. The manuscript contributes two proved formal results and one boundary analysis linked to executed validation artifacts: (i) an impossibility result showing that no family-blind scalar embedding can preserve bidirectional order fidelity when Pareto incomparability is present, (ii) a constructive existence result showing that typed contradiction-loss minimization attains zero loss under finite-acyclic and separability assumptions, and (iii) a local bridge-stability characterization for preference perturbations that is valid only inside a margin-Lipschitz regime. Executed evidence supports these claims with typed violation mean 0.0083 versus family-blind violation mean 0.2863, typed zero-loss feasibility in 405/540 runs under satisfied assumptions, and a sign-flip increase from 0.0723 inside the stability regime to 0.2108 outside it. The resulting survey protocol yields defensible method-selection guidance without asserting unsupported universal rankings across non-equivalent method families. 29. Simulability-Aware Quantum Reservoir Computing for Image Classification under Matched Readout Fairness Page: https://www.dinu.at/research/run-f03ebfa4-b901-47fb-beac-566191795521 Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-f03ebfa4-b901-47fb-beac-566191795521 PDF: https://www.dinu.at/papers/run-f03ebfa4-b901-47fb-beac-566191795521.pdf Pages: 14 Abstract: Quantum reservoir computing for vision tasks is often discussed in terms of empirical gains without an equally explicit separation between predictive improvement and computational-advantage claims. We study this issue in a controlled image-classification setting with PCA-compressed inputs, matched preprocessing, and identical linear readout training across classical and quantum reservoirs. The manuscript contributes a hybrid theory-plus-experiment framing: we formalize the readout stage as a strongly convex regularized problem with a unique closed-form optimum under positive regularization, derive concentration-calibrated conditions for reporting positive entanglement regimes, and define a compute-constrained Pareto boundary that filters performance gains by runtime and simulability costs. Experiments on MNIST, Fashion-MNIST, and a grayscale CIFAR-10 variant show that entangling quantum reservoirs can improve average macro-F1 against strong baselines in several configurations, but confidence-qualified positive-regime gates remain unmet under the current proxy-data execution. At the same time, symbolic and numerical checks validate the formal claims, and fixed-universe frontier audits satisfy monotonicity and nondominance requirements. The resulting evidence supports a simulability-safe conclusion: current results are strongest for formal guarantees and boundary/null characterization, while positive advantage claims remain conditional and require canonical-data reruns. 30. Stability-Aware Bilevel Source Dataset Selection for Importance-Weighted Least Squares in Unsupervised Domain Adaptation Page: https://www.dinu.at/research/run-c086d28a-7e79-4b1e-a2dc-b8111aca40cb Authors: anonymised in the document, which prints "Anonymous authors / Paper under review" Venue: none — the document names no venue Year: 2026 Source: https://github.com/alphaomega-labs/run-c086d28a-7e79-4b1e-a2dc-b8111aca40cb PDF: https://www.dinu.at/papers/run-c086d28a-7e79-4b1e-a2dc-b8111aca40cb.pdf Pages: 15 Generations: 6 separate runs of the same title are archived Abstract: Importance-weighted least squares (IWLS) is widely used to correct covariate shift in unsupervised domain adaptation, yet most prior work assumes that an appropriate source dataset is already available. In practical internet-scale retrieval settings, the key decision is reversed: given only unlabeled target covariates and a pool of candidate source datasets, which source (or source mixture) should be selected before fitting a weighted regressor? We study this question through a stability-aware bilevel framework with three formal components: a label-free source-ranking surrogate with a uniform regret guarantee, a multi-source mixture objective with linear-rate upper-level optimization under smoothness and strong-convexity assumptions, and a mixed-shift gate for harmful-source rejection before weighted fitting. We evaluate these components in three settings (distribution-level synthetic shift, semi-synthetic target-sample selection, and a protocol-faithful proxy real-track setting with governance checks). The empirical findings are mixed: symbolic checks and theorem-conditioned diagnostics are consistent, but Holm-corrected comparisons show no statistically significant advantage of the stability-aware selector over MMD-nearest, Wasserstein-nearest, or pooled IWLS in the current iter_1 run. This outcome clarifies that optimization guarantees and diagnostic structure are not sufficient for global predictive dominance under the present proxy data regime. The study contributes a reproducible no-target-label protocol, explicit failure reporting, and a concrete follow-up agenda for real benchmark ingestion and post-selection-controlled confirmatory analysis. PATENT ====== Title: Agent-Based Neuro-Symbolic Methodologies for Scientific Discoveries and Workflow Automation Status: Patent pending — a published application, not a grant Publication number: US 2026/0212154 A1 Application number: 19/032,587 Filed: 21 January 2025 Published: 23 July 2026 Assignee: ExtensityAI FlexCo, Wels (AT) Inventors: Marius-Constantin Dinu; Claudiu Leoveanu-Condrei PDF: https://www.dinu.at/papers/US-2026-0212154-A1.pdf (23 pages) SYSTEMS (4) =========== Every figure below carries the command or source it was measured with, because a number without a provenance is a claim. 1. Cortex — A governed control plane for AI agents Page: https://www.dinu.at/systems#cortex Organisation: Alpha Omega Labs Period: Jul 2026 — present Site: https://cortex.a2olabs.com/ What it is: A private control tower that lets a company run AI coding agents from chat apps like Slack or Telegram — with rules about who may do what, a full audit trail, and one-click deployment of whatever the agents build. Stack: TypeScript, Python, React, Node, Docker, Terraform, OpenTelemetry, Inspect AI Measured — First-party code: 182,600 lines (source: git ls-files across .py/.ts/.tsx/.js/.mjs/.sh, excluding one vendored bundle directory) Measured — Backend service modules: 53 (source: ls control/plane/src/server/services/) Measured — Python test functions: 643 (source: grep 'def test_' across tests/) - Two-plane authorisation: 15 control surfaces × 4 account groups × per-runtime capability policies, so a lower-privilege human can drive a higher-capability agent without escalation. Append-only audit log; TTL-scoped cross-organisation delegation. - Gauntlet, a four-track agent-harness benchmark on UK AISI's Inspect AI composing eight public datasets, running safety probes inside a network-blocked sandbox against a fabricated decoy. Measures *harness delta* rather than absolute model score, because scaffold choice can move results further than the model does. - Formal adversarial review of the agent wrapper with every finding closed and regression-tested: a normalised command classifier resistant to several evasion classes, a fail-closed policy bridge and approval gate, and an OS-level sandbox backstop. - Production surface: Terraform-provisioned hosts, 13 Docker Compose stacks, 5 CI/CD pipelines including a fully automated provisioning lane with no operator and no agent in the loop, and OpenTelemetry tracing correlated browser-to-server. 2. omegaXiv — Autonomous research platform Page: https://www.dinu.at/systems#omegaxiv Organisation: Alpha Omega Labs Period: Jul 2026 — present What it is: An online platform where AI agents actually run a scientific investigation end to end — planning it, executing code in isolated sandboxes, checking their own work against strict quality contracts, and publishing a finished, peer-reviewable paper. Stack: Python, FastAPI, Next.js, PostgreSQL, Redis, Azure, Terraform, Pydantic Measured — Application code: 613,600 lines (source: git ls-files across .py/.ts/.tsx) Measured — Tests: 8,100 Python + 3,900 TS (source: grep 'def test_' and it(/test( across the tree; 80% coverage gate) Measured — Services: 9 (source: docs/v3/ARCHITECTURE.md service table) - Designed the run engine: a deterministic render → turn → parse → validate → repair phase loop that binds every model output to a per-phase schema, classifies failures into a six-value verdict, and auto-repairs within bounded attempts before escalating to a human. - Built spend governance into the run lifecycle — payment pre-authorisation before dispatch settled by capture-or-cancel through a single idempotent finaliser, plus a cross-replica model throttle and a token-budgeted context packager. - Shipped an autonomous supervisor that detects interrupted runs, decides with an LLM, and applies a typed intervention or escalates — with row-level locking so a supervisor and a human cannot double-apply the same fix. - Enforced architecture mechanically rather than by review: layer import direction checked in CI, consumer-driven contract tests that load the producer's real payload builder, and a harness that turns any skipped or empty test tier into a build failure. 3. SymbolicAI — Open-source neuro-symbolic framework Page: https://www.dinu.at/systems#symbolicai Organisation: ExtensityAI Period: 2022 — present Site: https://arxiv.org/abs/2402.00854 Repository: https://github.com/ExtensityAI/symbolicai What it is: An open-source framework that combines language models with mathematical solvers, so a complex task breaks into small steps that can each be checked for correctness instead of trusted. Stack: Python, Z3, SymPy, WolframAlpha, BSD-3-Clause Measured — GitHub stars: 1,745 (source: GitHub API, 2026-08-04) Measured — Forks: 91 (source: GitHub API, 2026-08-04) Measured — Citations: 49 (source: Google Scholar, 2026-08-04) - Author and maintainer. Composes generative models with symbolic solvers over an expression graph, decomposing a complex task into small operations that are individually verifiable and recombinable. - Published at CoLLAs 2024. Recognised publicly by Gary Marcus, Sepp Hochreiter, Jim Fan, an IBM Executive Architect and Microsoft's Data & AI Product Lead. - Covered by MarkTechPost, Science Times, Montreal.AI and mind-verse. 4. Story Editor — Human-centric AI publishing Page: https://www.dinu.at/systems#storyeditor Organisation: story.one × ExtensityAI Period: 2024 — 2025 Site: https://story.one What it is: A writing tool built with the publishing platform story.one that guides an author from research to a finished book through six checked stages, instead of generating a draft and hoping. Stack: Python, SymbolicAI, LLM orchestration Measured — Partnership: 18 months (source: The Inflection Point, ExtensityAI blog, Sep 2025) Measured — Launch: Frankfurt Book Fair 2025 (source: PR Newswire, Sep 2025) - Led the AI architecture productising the neuro-symbolic stack into a six-stage authoring workflow — search, read, summarise, write, improve, finalise — with structured outputs, cross-reference consistency checks and human-in-the-loop review. - Defined abstract editorial concepts such as writing style and tonality in terms a machine could execute without conflicting instructions. TALKS AND PODCASTS (4) ====================== 1. Reinforcement Learning — Where We Are and What's Next Venue: International RL Bootcamp (2025) Format: video Listen: https://www.youtube.com/watch?v=zslhobE8nYQ 2. KI-Startup-Founder über OpenAI, DeepSeek und Artificial General Intelligence Venue: brutkasten (2025) Format: video Listen: https://www.youtube.com/watch?v=oYKX9_wk_yA 3. Extensity AI — The tool that automates research work Venue: startup.ro (2025) Format: video Listen: https://www.youtube.com/watch?v=1ujfFZESe-M 4. Building reliable and explainable AI agent systems Venue: Austrian Artificial Intelligence Podcast (2024) Format: podcast Duration: 1h 15m Listen: https://creators.spotify.com/pod/profile/aaip/episodes/62--Marius-Constantin-Dinu---extensity-ai---Building-reliable-and-explainable-AI-Agent-Systems-e2q4oj8 About: Why SymbolicAI treats a language model as a semantic parser inside a workflow rather than the workflow itself, and how you measure whether an individual reasoning step was actually correct. WRITING (37) ============ Newest first. Most of these are ports of posts first published elsewhere; each says where it came from. Feed: https://www.dinu.at/rss.xml 1. Cortex: a fixed-point theory of governed coding agents Page: https://www.dinu.at/writing/cortex-a-fixed-point-theory-of-governed-coding-agents Published: 2026-06-01 (1 June 2026) Description: Treating an agent's validate-repair loop as a monotone operator on a lattice of requirements, so that 'the task is done' becomes a least fixed point you can prove it reaches. Tags: research-paper, ai-agents, formal-methods, ai-safety Provenance: written for this site 2. Ringdown bounds on UV-regularized black-hole cores Page: https://www.dinu.at/writing/ringdown-bounds-on-uv-regularized-black-hole-cores Published: 2025-09-30 (30 September 2025) Description: Using the ringdown of merging black holes to bound the size of a non-singular quantum core, and the scaling-law regression framework that turns a whole catalogue of events into one test. Tags: research-paper, gravitational-waves, quantum-gravity, black-holes, ai-for-science Provenance: first published on the ExtensityAI blog, which has since gone offline 3. The inflection point — the story behind ExtensityAI Page: https://www.dinu.at/writing/the-inflection-point-the-story-behind-extensityai Published: 2025-09-27 (27 September 2025) Description: Why I wrote The Inflection Point, and how the neuro-symbolic architecture it argues for became a working product with story.one. Tags: neuro-symbolic, software-3-0, story-one, book Provenance: first published on the ExtensityAI blog, which has since gone offline 4. Recap of the 2nd International RL Bootcamp Page: https://www.dinu.at/writing/recap-of-the-2nd-international-rl-bootcamp-2025 Published: 2025-09-25 (25 September 2025) Description: A look back at the RL Bootcamp 2025 in Salzburg: highlights, speakers, and the recordings, slides and community links the organizers published. Tags: reinforcement-learning, community, events Provenance: first published on the ExtensityAI blog, which has since gone offline 5. A dialogue on AI, market power and a sustainable future Page: https://www.dinu.at/writing/a-dialogue-on-ai-market-power-and-a-sustainable-future Published: 2025-08-19 (19 August 2025) Description: Notes on a cross-disciplinary collaboration with Indian legal scholars and the resulting joint report on AI market power and India in a multipolar world. Tags: ai-policy, market-power, neuro-symbolic, collaboration Provenance: first published on the ExtensityAI blog, which has since gone offline 6. HyDRA: contracts as the control mechanism for knowledge-graph construction Page: https://www.dinu.at/writing/hydra-verifiable-knowledge-graphs Published: 2025-07-21 (21 July 2025) Description: Design-by-Contract applied to LLM-driven ontology and knowledge-graph generation, and an honest negative result about the benchmark used to measure it. Tags: research-paper, knowledge-graphs, neuro-symbolic, llm Provenance: written for this site 7. Entering the three-body problem Page: https://www.dinu.at/writing/three-body-problem Published: 2025-07-09 (9 July 2025) Description: Inspired by sci-fi's The Three-Body Problem, this post explores chaos, system dynamics, and how AI research automation can master complexity. Tags: three-body-problem, chaos-theory, system-dynamics, ai-research Provenance: first published on the ExtensityAI blog, which has since gone offline 8. Primality testing via circulant matrix eigenvalue structure Page: https://www.dinu.at/writing/primality-testing-via-circulant-matrices Published: 2025-04-28 (28 April 2025) Description: A prime is exactly an integer whose circulant matrix has a minimal polynomial with two irreducible factors over the rationals. A clean characterisation, and a slow algorithm. Tags: research-paper, number-theory, galois-theory, ai-for-science Provenance: written for this site 9. Empowering research innovation Page: https://www.dinu.at/writing/empowering-research-innovation Published: 2025-04-23 (23 April 2025) Description: How the neuro-symbolic stack behind SymbolicAI became a research automation platform, and what happened when we pointed it at an open question in primality testing. Tags: research-automation, neuro-symbolic, symbolic-ai, verification Provenance: first published on the ExtensityAI blog, which has since gone offline 10. The benchmark illusion Page: https://www.dinu.at/writing/ai-benchmark-illusion Published: 2025-04-07 (7 April 2025) Description: AI models ace benchmarks yet fail real-world tests — a critical assessment in evaluations and why we need better frameworks to measure intelligence. Tags: benchmarks, evaluation, neuro-symbolic, llm Provenance: first published on the ExtensityAI blog, which has since gone offline 11. PyFlow.ts: bridging ML research with production Page: https://www.dinu.at/writing/pyflow-ts-bridging-ml-research-with-production Published: 2025-03-07 (7 March 2025) Description: How PyFlow.ts exposes Python ML code to TypeScript with one decorator, removing the API, typing and client boilerplate of the last-mile deployment step. Tags: pyflow-ts, python, typescript, ml-deployment Provenance: first published on the ExtensityAI blog, which has since gone offline 12. Year one in review Page: https://www.dinu.at/writing/year-1-recap Published: 2024-12-16 (16 December 2024) Description: Recap of ExtensityAI's first year: origins as a neuro-symbolic research framework, the team, milestones, publications, and plans for 2025. Tags: extensityai, neuro-symbolic, research-automation, year-in-review Provenance: first published on the ExtensityAI blog, which has since gone offline 13. Parameter choice and neuro-symbolic approaches for deep domain-invariant learning Page: https://www.dinu.at/writing/parameter-choice-and-neuro-symbolic-domain-invariant-learning Published: 2024-07-01 (1 July 2024) Description: My doctoral thesis at JKU Linz: how to pick hyperparameters for domain adaptation when the target has no labels, and what to do when you cannot train at all. Tags: research-paper, domain-adaptation, neuro-symbolic, phd-thesis Provenance: written for this site 14. Large language models can self-improve at web agent tasks Page: https://www.dinu.at/writing/llms-can-self-improve-at-web-agent-tasks Published: 2024-05-30 (30 May 2024) Description: Harvesting an agent's own WebArena trajectories, filtering them without labels, and fine-tuning on the result — plus two metrics for what a benchmark score hides. Tags: research-paper, ai-agents, self-improvement, evaluation Provenance: written for this site 15. SymbolicAI: industry recognition and community impact Page: https://www.dinu.at/writing/symbolicai-framework-industry-recognition-and-community-impact Published: 2024-03-20 (20 March 2024) Description: Recognition that the SymbolicAI framework and its research paper received from AI researchers, industry figures, media outlets, and the community. Tags: symbolic-ai, neuro-symbolic, open-source, community Provenance: first published on the ExtensityAI blog, which has since gone offline 16. Announcing the SymbolicAI paper Page: https://www.dinu.at/writing/extensityai-announces-symbolicai-framework-a-new-paradigm-for-logic-based-ai Published: 2024-02-03 (3 February 2024) Description: The research paper behind SymbolicAI: a modular framework for logic-based approaches that compose generative models with solvers, plus a benchmark for evaluating LLMs in AI-centric workflows. Tags: symbolic-ai, neuro-symbolic, research-paper, open-source Provenance: first published on the ExtensityAI blog, which has since gone offline 17. A neuro-symbolic perspective on large language models Page: https://www.dinu.at/writing/a-neuro-symbolic-perspective-on-large-language-models Published: 2023-01-22 (22 January 2023) Description: The Symbolic API: using LLMs as the core of a neuro-symbolic stack, decomposing hard tasks into simple zero-shot operations and recombining them. Tags: research-paper, neuro-symbolic, symbolic-ai, llm, open-source Provenance: first published on dinu.at 18. InfODist: online distillation with informative rewards Page: https://www.dinu.at/writing/infodist-online-distillation-with-informative-rewards Published: 2022-12-09 (9 December 2022) Description: Why curriculum-learning agents generalise badly to the next task, and why the culprit is the states they explore rather than the non-stationarity they endure. Tags: research-paper, reinforcement-learning, curriculum-learning, generalization Provenance: written for this site 19. Addressing parameter choice issues in unsupervised domain adaptation by aggregation Page: https://www.dinu.at/writing/parameter-choice-in-unsupervised-domain-adaptation-by-aggregation Published: 2022-08-13 (13 August 2022) Description: Extending weighted least squares to vector-valued functions, with a target error asymptotically no worse than twice the unknown optimal aggregation. Tags: domain-adaptation, unsupervised-learning, research-paper, theory Provenance: first published on dinu.at 20. A dataset perspective on offline reinforcement learning Page: https://www.dinu.at/writing/a-dataset-perspective-on-offline-reinforcement-learning Published: 2022-08-01 (1 August 2022) Description: Two measures — SACo for exploration, TQ for exploitation — that characterise a behavioural policy's dataset and predict which offline RL algorithms will work on it. Tags: reinforcement-learning, offline-rl, datasets, research-paper Provenance: first published on dinu.at 21. Reactive exploration to cope with non-stationarity in lifelong reinforcement learning Page: https://www.dinu.at/writing/reactive-exploration-in-lifelong-reinforcement-learning Published: 2022-08-01 (1 August 2022) Description: Tracking continual domain shifts as they happen: why policy-gradient methods adapt faster than Q-learning when the environment keeps moving. Tags: reinforcement-learning, lifelong-learning, exploration, research-paper Provenance: first published on dinu.at 22. Align-RUDDER: learning from few demonstrations by reward redistribution Page: https://www.dinu.at/writing/align-rudder-learning-from-few-demonstrations Published: 2022-03-10 (10 March 2022) Description: Replacing RUDDER's LSTM with a profile model from multiple sequence alignment, so reward redistribution works from a handful of demonstrations. Tags: reinforcement-learning, reward-redistribution, learning-from-demonstrations, research-paper Provenance: first published on dinu.at 23. The balancing principle for parameter choice in distance-regularized domain adaptation Page: https://www.dinu.at/writing/the-balancing-principle-for-parameter-choice-in-domain-adaptation Published: 2021-09-24 (24 September 2021) Description: Borrowing the balancing principle from ill-posed inverse problems to justify the regularization parameter when the target domain has no labels. Tags: domain-adaptation, regularization, inverse-problems, research-paper Provenance: first published on dinu.at 24. Lighter: dependency injection for PyTorch Page: https://www.dinu.at/writing/lighter-dependency-injection-for-pytorch Published: 2021-06-09 (9 June 2021) Description: Python borrowed little from the aspect-oriented practices that matured in Java and C#. Lighter brings dependency injection to PyTorch projects. Tags: python, pytorch, dependency-injection, open-source Provenance: first published on dinu.at 25. JKU AI overview video Page: https://www.dinu.at/writing/jku-ai-overview-video Published: 2021-02-18 (18 February 2021) Description: A short overview of what AI makes possible, produced at Johannes Kepler University Linz. Tags: jku, artificial-intelligence, video Provenance: first published on dinu.at 26. XAI and strategy extraction via reward redistribution Page: https://www.dinu.at/writing/xai-and-strategy-extraction-via-reward-redistribution Published: 2020-10-17 (17 October 2020) Description: Using Align-RUDDER as an interpretability method: multiple sequence alignment surfaces the key events an agent relies on, and those events are usually human-readable sub-tasks. Tags: reinforcement-learning, explainable-ai, reward-redistribution, research-paper Provenance: first published on dinu.at 27. Align-RUDDER: reward redistribution from few demonstrations Page: https://www.dinu.at/writing/align-rudder-reward-redistribution-from-few-demonstrations Published: 2020-09-30 (30 September 2020) Description: The first write-up of Align-RUDDER: two modifications to RUDDER that let profile models built from as few as two demonstrations carry the reward redistribution. Tags: research-paper, reinforcement-learning, reward-redistribution, learning-from-demonstrations, credit-assignment Provenance: first published on dinu.at 28. Overcoming catastrophic forgetting with context-dependent activations Page: https://www.dinu.at/writing/overcoming-catastrophic-forgetting-with-context-dependent-activations Published: 2019-11-25 (25 November 2019) Description: Master's thesis: context-based gating that switches pathways through a network, cutting destructive interference on long task sequences. Tags: continual-learning, catastrophic-forgetting, reinforcement-learning, thesis Provenance: first published on dinu.at 29. Imagine Kara Page: https://www.dinu.at/writing/imagine-kara Published: 2018-05-09 (9 May 2018) Description: A thank-you to the team behind Imagine Kara as the project wound down. Tags: projects, teams Provenance: first published on dinu.at 30. Seamlessly entering the crypto world with Apollon Page: https://www.dinu.at/writing/seamlessly-entering-the-crypto-world-with-apollon Published: 2018-04-25 (25 April 2018) Description: A short pointer to a longer write-up on the Apollon project and getting started with masternodes. Tags: blockchain, apollon, masternodes Provenance: first published on dinu.at 31. Earning almost one Bitcoin with masternodes in three weeks Page: https://www.dinu.at/writing/earning-bitcoin-with-masternodes Published: 2018-04-20 (20 April 2018) Description: How hosting Apollon masternodes worked out over three weeks in 2018, and how the project's search interest compared to the top ten coins. Tags: blockchain, apollon, masternodes Provenance: first published on dinu.at 32. Uni Swift project Page: https://www.dinu.at/writing/uni-swift-project Published: 2017-05-15 (15 May 2017) Description: A university iOS app for managing photos, built around actually finding the right image again. Tags: swift, ios, university-project Provenance: first published on dinu.at 33. Deep Learning Script Page: https://www.dinu.at/writing/deep-learning-script Published: 2017-04-19 (19 April 2017) Description: A DSL for Caffe that cuts the line count of prototyping a network architecture, with a VS Code extension that transpiles to Caffe Script. Tags: dsl, caffe, deep-learning, open-source Provenance: first published on dinu.at 34. Deep learning: where to start and how to dig deeper Page: https://www.dinu.at/writing/deep-learning-primer Published: 2016-11-26 (26 November 2016) Description: Two reference notes from 2016 — an overview of the field, and a practical guide to running deep learning on Ubuntu with Docker. Tags: deep-learning, docker, notes Provenance: first published on dinu.at 35. Operation Phrike Page: https://www.dinu.at/writing/operation-phrike Published: 2016-09-07 (7 September 2016) Description: A VR combat-simulation study: Oculus Rift plus Unreal Engine, with sensors reading the test subject's stress level. Tags: virtual-reality, unreal-engine, university-project Provenance: first published on dinu.at 36. Internship report: handwritten character recognition Page: https://www.dinu.at/writing/internship-report Published: 2016-08-31 (31 August 2016) Description: A Xamarin app that classifies handwritten characters with a neural network, compared against a support vector machine. Written at Siemens Corporate Technology. Tags: neural-networks, xamarin, internship, siemens Provenance: first published on dinu.at 37. Bachelor thesis: cross-language integration for the CLR Page: https://www.dinu.at/writing/bachelor-thesis Published: 2015-07-21 (21 July 2015) Description: Compiling JavaScript to .NET IL by pairing a Coco/R-generated parser with Roslyn's syntax trees. Tags: compilers, roslyn, dotnet, thesis Provenance: first published on dinu.at TEACHING (3) ============ 3-day intensive courses, 21 contact hours each. Delivered in English or German. On-site in Europe, or remote. Wels / Linz, Austria is the home base. Index: https://www.dinu.at/teaching · Version 1.0 · framework review date 1 August 2026 1. LLM Engineering — From prompting and statistical inference to reliable RAG systems Page: https://www.dinu.at/teaching/llm-engineering Syllabus: https://www.dinu.at/files/course-llm-engineering-syllabus.pdf Contact hours: 21 What it teaches: How to build an assistant that answers from your own documents, cites where each claim came from, and ships with a test suite that catches it inventing things. Audience: Engineers, data scientists, technical leads and advanced AI practitioners. Delivery: One shared codebase developed through guided experiments, retrieval labs and evaluation. Capstone: An evidence-grounded assistant with hybrid retrieval, citations, tests and safety controls. Honest scope: Feasible as a working-proficiency course, provided GraphRAG indexing and prompt-program optimization are demonstrations rather than separate full projects. Not taught: LlamaIndex can replace Haystack when data connectors and indexing abstractions are the primary concern; teaching both would dilute the workshop.; Managed vector databases may replace Qdrant or pgvector in production, but the workshop keeps infrastructure transparent.; Long-context prompting, SQL or graph queries and fine-tuning are treated as alternative architectural choices, not automatic upgrades to RAG. 2. Reinforcement Learning — From Markov decision processes and control to PPO, continuous control and RLVR Page: https://www.dinu.at/teaching/reinforcement-learning Syllabus: https://www.dinu.at/files/course-reinforcement-learning-syllabus.pdf Contact hours: 21 What it teaches: How software learns to make decisions by trying things and being scored on the result — the method behind game-playing systems, robot control, and the way modern reasoning models are trained. Audience: ML engineers, researchers, simulation engineers, robotics developers and advanced students. Delivery: Algorithms from scratch where pedagogically valuable, frameworks where implementation complexity would hide the ideas. Capstone: A small Gymnasium environment with baseline control, trained policy, ablation and multi-seed evaluation. Honest scope: Feasible only as an intensive foundation course. Tabular RL and REINFORCE are implemented; PPO and continuous control use libraries; RLVR is a verifier and training-loop practicum, not a full large-model run. Not taught: CleanRL is useful for readable single-file reference implementations, but the workshop already includes from-scratch tabular and REINFORCE code.; PettingZoo is introduced when the participant group specifically needs multi-agent environments.; MuJoCo tasks are optional: installation and compute variability can consume too much workshop time, and Pendulum is a reliable continuous-control baseline. 3. Agentic Engineering — From tool calling and MCP to governed, testable agent systems Page: https://www.dinu.at/teaching/agentic-engineering Syllabus: https://www.dinu.at/files/course-agentic-engineering-syllabus.pdf Contact hours: 21 What it teaches: How to take an AI agent that works in a demo and give it permissions, tests, an audit trail and a stop button — so you can let it near a real system. Audience: AI and software engineers, architects, security specialists and technical product owners. Delivery: One agent is progressively extended with tools, MCP, state, approvals, tests, security and governance evidence. Capstone: A governed research or operations agent with typed contracts, an MCP server, evaluation and red-team evidence. Honest scope: Feasible when one primary Python stack is used and multi-agent systems remain a design exercise. Building several unrelated agents or comparing every framework hands-on would not fit. Not taught: OpenAI Agents SDK is discussed for managed turns, tools, handoffs, guardrails, sessions and built-in tracing.; Microsoft Agent Framework is discussed for enterprise workflows, typed executors, state management and explicit multi-agent orchestration.; PydanticAI graph and durable-execution integrations can replace LangGraph; the learning objective is explicit state and recovery semantics, not one framework.; Crew-style multi-agent frameworks are not the default because coordination complexity obscures contracts, security and evaluation. ROLES (9) ========= Full history with detail: https://www.dinu.at/cv 1. Research — Alpha Omega Labs Period: Jul 2026 — Present Location: Austria Summary: Architect and principal engineer of two production platforms: a governed control plane for AI agents, and an autonomous research platform. 2. Founder & Chief Executive Officer — ExtensityAI FlexCo Period: Mar 2023 — Jun 2026 Location: Wels, Austria Summary: Founded and led a neuro-symbolic AI company from zero to over EUR 1M in revenue, profitable within its first year, and a Sifted/FT top-11 listing. 3. AI Research Director & Advisor — Atlas Period: Apr 2023 — Dec 2024 Location: Vienna, Austria Summary: Directed the applied research agenda across deep learning and reinforcement learning. 4. Research Scientist (Ph.D.) — Johannes Kepler University — LIT AI Lab Period: Sep 2019 — Sep 2024 Location: Linz, Austria Summary: Doctoral research on parameter choice in unsupervised domain adaptation and neuro-symbolic learning, under Sepp Hochreiter. 5. Senior Machine Learning Researcher — Dynatrace Research Period: Oct 2021 — Mar 2023 Location: Linz, Austria Summary: Deep learning and NLP prototypes at the LIT Open Innovation Center. 6. Founder & Software Architect — Imagine Kara LLC Period: May 2018 — Dec 2018 Location: Delaware, USA Summary: Cloud masternode platform for automated coin hosting. 7. Data Scientist — CELUM Period: Oct 2016 — Aug 2019 Location: Linz, Austria Summary: Deep-learning image classification for digital asset management. 8. Applied Research Scientist — Siemens Corporate Research Period: Mar 2016 — Sep 2016 Location: Princeton, NJ, USA Summary: Handwritten-character recognition across mobile platforms. 9. Software Architect & Product Manager — Kontron Period: Dec 2010 — Mar 2016 Location: Linz, Austria Summary: Led an international team building software for self-service coin-counting devices. EDUCATION (5) ============= 1. Ph.D. (Dr. techn.), Artificial Intelligence — Johannes Kepler University, Linz, Austria (2019 — 2024) Note: Passed with distinction. Thesis: Parameter Choice and Neuro-Symbolic Approaches for Deep Domain-Invariant Learning. Advisor: Sepp Hochreiter. 2. Executive MBA, Economics — WU Executive Academy & UMN Carlson School, Vienna / Minneapolis (2024 — 2026) 3. M.Sc., Computer Science — Data Science — Johannes Kepler University, Linz, Austria (2017 — 2019) 4. Exchange Semester, Machine Learning — National Taiwan University of Science and Technology, Taipei, Taiwan (2018 — 2019) 5. B.Sc., Software Engineering — University of Applied Sciences Upper Austria, Wels, Austria (2013 — 2016) RECOGNITION (4) =============== 1. Sifted / Financial Times — Europe's top 11 AI startups to watch (2024) Detail: ExtensityAI named in the annual investor watch list. Source: https://sifted.eu/articles/ai-startups-investor-watch-2024 2. ICLR 2023 Oral — notable top 5% (2023) Detail: Addressing Parameter Choice Issues in Unsupervised Domain Adaptation by Aggregation. 3. Ph.D. passed with distinction (2024) Detail: Johannes Kepler University Linz. 4. Author — The Inflection Point (2025) Detail: On the shift from AI alchemy to verifiable, architecture-driven AI systems. Source: https://www.story.one/en/book/the-inflection-point-68d75d2076dd0/ COMMUNITY AND AFFILIATIONS (4) ============================== 1. AI Austria RL Community — team member; co-organiser of the International Reinforcement Learning Bootcamp (2nd edition, 2025). Link: https://aiaustria.com/rl-community 2. Distinguished Expert, Global Advisory Council of the Indian Society of Artificial Intelligence and Law (ISAIL-GAC). Link: https://www.isail.in/ 3. Founder of the "AI Is All You Need" research community on Discord, 450+ members; research outreach reaching 250,000+ views across X and LinkedIn. Link: https://discord.gg/azDQxCHeDA 4. Open-source maintainer: SymbolicAI, and Lighter (dependency injection for PyTorch). LEGAL ===== Impressum (Offenlegung under § 25 MedienG, provider information under § 5 ECG): https://www.dinu.at/impressum Privacy: https://www.dinu.at/privacy