MatrixReward proposes a reward mechanism for open-ended generation by constructing rewards from a rollout-by-rubric win-rate matrix. Compared to previous methods, MatrixReward achieves an average score of 63.02, outperforming the strongest baseline by approximately 2.0%.

Researchers propose using Test-Time Training (TTT) layers in a Decision Transformer to improve long-term memory in offline Reinforcement Learning (RL). They analyze the performance of the Decision Titan, a variant of the Decision Transformer with TTT layers, in the X-Maze environment.

Researchers propose Sharpening Tax, a diagnostic metric to measure the loss in test-time scalability after post-training large language models (LLMs). They also present a Bayesian sampler, posterior-tempered group sampling (PTGS), which adapts the sampling temperature per prompt to its difficulty, and show that it pays a smaller Sharpening Tax than a fixed-temperature baseline.

This paper proposes Dependency-Aware Reward Shaping (DARS), a method for assigning step-level credit to reinforcement learning tasks. DARS represents task progress as a dependency graph and assigns rewards based on the graph distance from broken prerequisites. The method integrates with existing agentic training methods and improves success rates in various tasks.

Researchers propose ETHER, an agent that uses Emergent Communication to learn a grounded, artificial language for goal-conditioned reinforcement learning, addressing limitations of Hindsight Experience Replay (HER).

EnvACE is a new agentic reinforcement learning method that internalizes environment dynamics by replacing external environment interaction during training with world rehearsal. This allows for strong and transferable performance across various benchmarks, including FinMCP-Bench.

SETA (Scaling Environments for Terminal Agents) is a framework for generating verifiable terminal environments for reinforcement learning (RL). It consists of two pipelines and a large open-source dataset, SETA-Env, containing over 4,500 environments. SETA was used to train Qwen3-8B and DeepSeek-V4-Flash, achieving state-of-the-art results on Terminal-Bench 2.0.

Researchers proposed Protocol-level Rubrics (ProRubric), a protocol-level aggregation method for rubric-based reinforcement learning that improves appropriateness and maintains coverage. The method groups criteria into dimensions, counting only when all criteria hold, and has shown a 10.8-point increase in appropriateness without losing coverage.

StateTree is a data-driven RL method for long-term dialogue reasoning in large language models. It constructs a tree-structured path-tracing task from dialogues with verifiable ground truth, allowing the model to traverse and compare records across sessions. StateTree generalizes to longer contexts without full-length RL costs and exhibits capabilities such as cross-session retrieval and temporal reasoning.

A paper proposes GRAFT, an off-policy-aware framework that allows reinforcement learning models to learn from each other's experiences, improving performance and reducing the need for costly rollouts.

MetaCtrl is a lightweight controller that adaptively regulates large language models (LLMs) to improve their reasoning accuracy while reducing inference-time generation. It observes the evolving reasoning trace, decides whether to continue, simplify, or conclude reasoning, and is trained using reinforcement learning. MetaCtrl improves the accuracy of LLMs on various benchmarks, including mathematics, science, and code, and can transfer to unseen reasoners without further training.

OptiCom is a unified framework for state-conditioned composition in LLM-driven optimization. It represents LLM-driven optimizers within a shared configuration space and dynamically composes immediate mechanisms through structured Action Packages. Evaluations demonstrate the superiority of OptiCom over 14 configurations, achieving the top score in 23 benchmark groups.

A new method, PACE, is proposed for controlling staleness in asynchronous reinforcement learning (RL) for large language model post-training. PACE improves validation accuracy and reduces GPU time, matching synchronous RL performance.

The paper presents UpliftMem, a method for learning memory retrieval for large language model (LLM) agents. UpliftMem learns which memory sets improve execution without relying on costly outcome feedback, instead using set-level execution uplift relative to the same executor without memory. This approach is evaluated on three benchmarks (ALFWorld, WebShop, and BigCodeBench) and outperforms other baselines.

Researchers extended Reinforcement Learning with Verifiable Rewards to Large Vision-Language Models and found that visual claims are not always supported by images. They introduced a diagnostic to measure the persistence of visual claims and proposed Persistence-Aware Credit Gating to attenuate credit for persistent claims.

This paper proposes BRIDGE, a new algorithm for agentic reinforcement learning that jointly optimizes large language models (LLMs) and retrievers, addressing the information-credit gap in existing ARL methods.

A paper introduces a governance architecture, Global Executive Control (GEC), to address the issue of persistent action in large language models (LLMs) despite diminishing task value. The architecture separates action generation from project-level control, improving success rates and reducing token use.