arXiv:2608. 01743v1 Announce Type: new Abstract: Reinforcement learning (RL) has become a central paradigm for large language model (LLM) post-training, but optimization toward new objectives can degrade capabilities already present in the base model.
The paper presents a finite‑model synthesis that integrates operational state abstraction with optimal control within a stability‑evidence‑revision (SER) framework, termed "Poincaré meets Bellman." It distinguishes qualitative dynamics for reusable action‑response structures from dynamic programming that governs acquisition, retention, reuse, merging, and forgetting, and introduces a Bellman recursion over the joint law of hidden state and deployed memory. The authors derive explicit retention rules, demonstrate how factor sharing and informative observations improve identification, and verify coding and retention calculations through finite enumerations.
The paper demonstrates that functional compatibility—how well new learning can coexist with behavior that must be preserved—is a causal determinant of persistent neural learning. By deliberately altering compatibility between matched neural states, the authors show that persistent learning can be measured under a common retention requirement across different learning directions, architectures, and data modalities. The study reveals that learning rules vary in how efficiently they exploit compatibility, and that retention constraints and finite updates limit what can be stored, with nonlinear geometry ultimately preventing full compatibility realization.
Revelation Control studies how to price interventions that reveal hidden state only when the revealed distinctions can alter a consequential decision, while separately accounting for any useful progress the intervention itself creates. The authors develop a framework for learning systems that defines decision‑sufficient revelation, revelation depth, and a cost‑adjusted factorization criterion, and they provide a target‑independent protocol for model‑specific instantiation. Experiments on Qwen2.5‑7B and Mistral‑7B‑v0.3 show that deeper future‑learning probes have positive decision value and that productive reuse yields strict equal‑compute utility advantages, supporting a structural transfer of the decision theory and evaluation protocol across architectures.
arXiv:2607. 05609v1 Announce Type: cross Abstract: The Continual Learning (CL) literature has long been driven by the goal of mitigating catastrophic forgetting.
arXiv:2608. 07228v1 Announce Type: new Abstract: When a reinforcement learning agent cannot observe the full state, we usually blame its policies: it cannot see enough to represent a good one.
arXiv:2510. 21978v2 Announce Type: replace-cross Abstract: Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning and has become a standard post-training paradigm for contemporary language and vision-language models.
arXiv:2603.25579v2 Announce Type: replace-cross Abstract: A key capability of modern neural networks is their capacity to simultaneously learn underlying rules and memorize specific facts or exceptio...
arXiv:2608. 16409v1 Announce Type: new Abstract: Today, a neural system is almost always used in two phases -- trained, then deployed -- and in that regime it freezes twice: training ends, and the topology itself was never a degree of freedom.
arXiv:2606. 14536v1 Announce Type: new Abstract: Safe reinforcement learning (RL) aims to learn policies that optimize rewards while satisfying constraints.
arXiv:2607. 14185v1 Announce Type: cross Abstract: Feedback-driven loops support iterative improvement in large language models, reinforcement learning, and autonomous discovery, yet their gains often diminish under repeated internal feedback.
The paper demonstrates that learned simulators can fail in two distinct ways when conditions change: long‑horizon drift due to accumulated errors and incorrect responses to interventions on physical parameters. By adding a symplectic integrator to preserve conservative dynamics, rollouts remain stable for up to 100× the training horizon, while encoding physical coupling via explicit linear factorization allows the model to generalize to unseen signs of that coupling. The study shows that stability and counterfactual generalization arise from separate structural choices, enabling designers to impose each property independently.