Feedback-driven loops support iterative improvement in large language models, reinforcement learning, and autonomous discovery, yet their gains often diminish under repeated internal feedback. We study why closed-loop knowledge systems saturate and what external information can move them beyond their current attractors.
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 introduces a framework for intrinsic‑extrinsic coupling in learning dynamics, where a learner’s current observations do not solely dictate its future training responses. It formalizes this coupling through a continuation‑conditioned value of a constrained learning‑state intervention and employs an executable finite‑frame classifier‑head to protect current logits while adjusting historical margins. Experiments across CLINC‑derived class‑incremental settings, output distillation with RoBERTa, and SGDW dynamics reveal that coupling can produce both positive and negative interactions, and that coordinated interventions can match or exceed development gains while reducing cross‑entropy loss compared to standard replay.
The paper introduces a framework for intrinsic‑extrinsic coupling in learning dynamics, defining it via a continuation‑conditioned value of a constrained learning‑state intervention and observation‑relative fibers. It presents an executable finite‑frame classifier‑head that protects current logits while repairing historical margins, and distinguishes local admissibility, intervention value, and complete‑policy performance. Experiments on CLINC‑derived class‑incremental tasks, output distillation with RoBERTa, and SGDW dynamics demonstrate that coupling can produce both positive and negative interactions, and that coordinated content controls can match or exceed development gains while guided allocation reduces cross‑entropy loss compared to standard replay.
The paper introduces the concept of Stable but Wrong (SBW), describing situations where a learning process appears stable yet converges to a solution that is systematically biased away from a true objective. Using a minimal strongly convex model, the authors demonstrate that persistent bias in update directions can shift the convergence point from the optimal solution. Experiments across reinforcement learning, supervised learning, and continual fine-tuning of large language models reveal a recurring disconnect between normal optimization behavior and correctness under both static and feedback‑coupled biases, and show that recovery interventions can still modify subsequent learning trajectories.
arXiv:2605. 25582v2 Announce Type: replace Abstract: Reinforcement learning for large language models faces a fundamental trade-off between sample efficiency and asymptotic performance: strictly on-policy methods discard trajectories after a single update, while off-policy reuse introduces distribution mismatch that existing trust-region techniques mitigate primarily by enforcing conservative optimization, often leaving rich training signals underutilized.
arXiv:2607. 04470v1 Announce Type: cross Abstract: Large Language Models (LLMs) offer a natural interface for translating human objectives into reward signals for cooperative multi-agent reinforcement learning (MARL), yet the training-time dynamics of this integration remain poorly understood.
The paper introduces OSOL, a method for mitigating higher‑order interference in multi‑domain reinforcement learning. OSOL selects a focus domain each iteration, uses token‑level footprints from the previous checkpoint to rank rebound risk, and applies an adaptively scaled correction to the GRPO update. Experiments on Qwen3‑30B‑A3B show a 5.7% improvement over the best baseline without higher‑order differentiation.
arXiv:2606. 09929v1 Announce Type: cross Abstract: Physical reservoir computing harnesses nonlinear mechanical dynamics but, by convention, freezes the substrate and trains only a linear readout, presuming the substrate is not usefully trainable.
arXiv:2607. 16858v1 Announce Type: cross Abstract: Across environments with mixed sources of uncertainty, unsupervised reinforcement learning requires intrinsic motivation that does not precommit to a particular direction of surprise.
The paper introduces On‑Policy Warmup (OPW), a teacher‑guided training stage where a student agent learns from a teacher on its own interaction trajectories before switching to reinforcement learning with verifiable rewards (RLVR). OPW differs from traditional imitation by focusing on states generated by the student’s own decisions, including imperfect actions and recovery situations. The authors provide a theoretical link between on‑policy reverse‑KL distillation and trajectory‑level distribution matching, showing that, under a competent teacher and low distillation loss, OPW can lower bound initial verifier success and reduce reward‑discovery complexity, thereby accelerating RLVR performance.
arXiv:2609.35897v1 Announce Type: new Abstract: The pursuit of recursive self-improvement (RSI) toward general intelligence is divided between macro-level language model scaling and the interaction-d...
arXiv:2605. 15219v3 Announce Type: replace Abstract: Can AI systems discover new knowledge through iterative self-improvement, and at what cost?