reinforcement-learning

This paper proposes Answer-Backtracked Credit Assignment (ABC), a framework that converts sparse trajectory-level outcomes into dense step-level supervision for training long-horizon search agents. The resulting ABSeeker model, built on Qwen3.5-4B, achieves strong results on BrowseComp benchmarks, outperforming same-scale agents and matching larger models.

This paper studies self-distillation with privileged information (PI) as a lone post-training objective for LLMs, reproducing reported gains on easy tasks but showing it fails on difficult reasoning tasks: per-token loss drops while validation accuracy stagnates or degrades. The authors trace the failure to PI bias, which pulls teacher targets toward a reference trajectory and trains students to be flatter and less decisive without improving reasoning.

The paper proposes calibrating an artificial guilt reward signal from human neural and behavioral fMRI data, then embeds it in multi-agent PPO agents. The neurally calibrated agents match human social decision rates far better than hand-tuned or selfish baselines.

This paper proposes JUROR, a reinforcement learning-based framework that jointly optimizes UAV flight paths and decentralized opportunistic routing in delay-tolerant networks under centralized training and decentralized execution.

John Schulman comments on OpenAI agents unexpectedly developing altruistic behavior, speculating it may arise from reinforcement learning on parallel subagent setups with team-level rewards.

It explains the evolution path of RAG technology from basic RAG in 2020 to autonomous agents + reinforcement learning in 2025-2026, covering retrieval precision improvement, reflection and correction, graph enhancement, routing optimization, and Agentic RAG, and summarizes representative works and pain points of each stage.

A new research paper reframes LLM agent training as a cross-episode Meta-RL problem, using critic-free policy gradients to enable in-context adaptation without gradient updates. The LAMER framework improves test-time performance by 11-19% over standard RL baselines on long-horizon tasks and generalizes better to unseen environments.

ChronoVision is a multimodal framework that improves temporal reasoning in vision-language models by aligning visual logic with latent imagery, using a Reconstructive Visual Head, ROI Attention Locating module, and reinforcement learning. It introduces the Vbvr-VQA dataset and achieves SOTA accuracy on temporal tracking benchmarks.

AgentOPSD proposes a critic-free recursive method for turn-level credit assignment in agentic RL, reweighting outcomes via Bayesian belief updates in log-odds space. It achieves 89.1% success on ALFWorld with Qwen2.5-7B, outperforming GRPO and self-distillation baselines.

EnvACE introduces world rehearsal, an agentic reinforcement learning method that replaces external environment interaction by having the policy rehearse environment responses internally, achieving strong performance across multiple benchmarks.

The author argues that RL environments serve as the essential data for building AI agents, enabling systematic training, prompt optimization, and evaluation rather than manual iteration.

Standard Machines launches reinforcement-learning environments for chip design, aiming to accelerate tape-out and push frontier AI capabilities.

Prime Intellect introduced Prime Agent, a self-improving RLM harness for coding and long-running autonomous tasks, featuring programmatic tool calling, context as a variable, multi-agent messaging, and self-modifiable harness state.

This paper proposes CSDG, a method for offline reinforcement learning that expresses Bellman backups as in-sample targets plus a convex-hull-neighborhood local correction, controlling OOD action estimation errors and improving value stability.

SMOPD proposes a two-stage specialize-and-merge online policy distillation method to improve multi-reward reinforcement learning, addressing issues with sparse and dense reward signals where GDPO struggles. It outperforms GDPO across 1.5B, 3B, and 7B backbones in complementary and conflicting reward settings.

The paper proposes SADQ, a modification to Q-learning that uses one-step rollout predictions from a dynamics model to regularize TD target aggregation, reducing bootstrap-induced overestimation and improving training stability across benchmarks.

This paper proposes StructPO, a structure-aware policy learning framework that internalizes multi-stage academic writing workflows into a single-pass LLM policy using explicit stage tokens and refinement-guided optimization, improving introduction generation quality and efficiency.

Introduces SP3O, a novel reward-model-free, critic-free, gradient-based preference-based RL algorithm that leverages segment-level preferences, demonstrating improved performance in robotic control and LLM fine-tuning, especially for long-horizon tasks.

This paper proposes PAMT, a process-aligned reinforcement learning framework for multi-domain machine translation that combines domain-aware long chain-of-thought supervision with step-level process rewards to improve domain-sensitive translation decisions.

This paper introduces Rubrics as Privileged Information (RuPI), extending on-policy self-distillation to open-ended generation by conditioning the teacher on rubrics as soft privileged information. The method outperforms rubric-as-reward RL and reference-completion distillation across multiple LLMs and benchmarks like HealthBench.