The paper introduces Stepwise Think-Critique (STC), an end‑to‑end trainable framework that lets a single large language model (LLM) interleave reasoning with inline, step‑level critique. STC uses reinforcement learning to jointly reward reasoning accuracy and critique consistency, achieving a 7.2% improvement in Pass@1 on five mathematical reasoning benchmarks and a 67.4% step‑level critique F1 score, outperforming several external reward models. This approach aligns LLM behavior more closely with human critical thinking by embedding self‑evaluation directly into the reasoning process.
By Jiaqi Xu, Cuiling Lan, Xuejin Chen, Yan Lu
The paper argues that in multi‑turn agentic reinforcement learning, credit assignment should be viewed as a coverage problem rather than a targeting problem. It introduces verifier information density (V_d) as a structural metric, showing that terminal‑state verifiers operate in a low‑V_d regime where targeting fails. Experiments on tau^2‑bench, BFCL, and ToolACE‑2‑8B demonstrate that uniformly distributing reward across all turns outperforms sparse, targeted rewards, and that full chain coverage is necessary for optimal performance.
By Chenyu Zhou, Qiliang Jiang, Shuning Wu, Xu Zhou
The paper introduces Potential-Guided Policy Optimization (PGPO), a method for multi-turn agentic tasks that improves credit assignment by estimating empirical state potentials from anchor-state-group return statistics. PGPO derives action advantages from potential differences between adjacent states, enabling cross-trajectory credit propagation and finer-grained step-level credit assignment, especially within failed trajectories. Experiments on ALFWorld and WebShop demonstrate strong performance compared to recent group-based reinforcement learning methods, with negligible training overhead.
By Yuyao Zheng, Haipeng Sun, Junwei Bao, Lemao Liu, Hongfei Jiang, Yang Song, Dejing Dou
The paper presents an online algorithm that achieves the same $0.401$ approximation factor for maximizing nonnegative, non-monotone DR-submodular functions over compact convex down-closed subsets of the $d$-dimensional unit cube as the best known offline construction. In the full-information value-oracle model, the algorithm attains this factor with sublinear regret, using $O(dT^{1/4})$ oracle calls per round and $O(T^{3/4})$ regret, and offers flexible batching trade-offs. Under a positive-anchor condition, a randomized blocking strategy preserves the $0.401$ factor while achieving $O(T^{5/6})$ one-point bandit regret.
By Vaneet Aggarwal, Yiyang Lu
The study examines how the effort expended by large reasoning models (LRMs) compares to that of humans during abductive reasoning tasks. By analyzing reaction times and reasoning traces, the authors find that LRMs and humans exhibit similar patterns of effort and error types. They also demonstrate that decoding strategies allowing models to explore multiple reasoning paths further align the models’ reasoning costs with human effort.
By Henry Arthur
The paper introduces Discriminative World Models for Web Agents, proposing a predicted-state matching objective that trains world models to produce representations that can distinguish the true resulting state from those of alternative actions. Using a branching dataset from WebArena Go-Browse, the authors demonstrate that this approach outperforms traditional supervised next-state prediction on a held‑out benchmark and improves action ranking on WebPRMBench. Additionally, employing the discriminative world model for test‑time action selection boosts end‑to‑end task success on WebArena‑Lite.
By Kelvin Li, Dhruv Pendharkar, Anish Pahilajani, Chuyi Shang, Leon Oks, Leonid Karlinsky, Rogerio Feris, Trevor Darrell, Roei Herzig
Cliff is a reward‑shaping method for reinforcement learning with verifiable rewards that identifies the first mistake in a language model’s reasoning process using an off‑the‑shelf teacher. By splitting each rollout into a correct prefix and an incorrect suffix, Cliff assigns token‑level advantages—positive for correct tokens and negative for the rest—providing fine‑grained supervision. Across 12 scenarios, Cliff outperforms on‑policy distillation by 15% and standard GRPO by 7%, even when the teacher is only modestly capable.
By Peixuan Han, Runhui Wang, Ketan Ramaneti, Jie Hao, Gerald Friedland, Chris Kong
IDEA is a training‑free, input‑dependent steering method for large language models that matches activations to cluster‑specific directions aligned with a target concept. It clusters positive and negative activation supports per attention head, solves an optimal‑matching problem to create a pool of cluster‑conditional directions, and selects the best match for each input at inference time. This approach preserves the input’s original representation while improving the truth × info rate on TruthfulQA by an average of 9.9% (up to 23.5%) over input‑independent baselines.
By Zheng Wang, Muchen Li, Renjie Liao, Yan Leng
The paper introduces Exchange Policy Optimization (EPO), a framework for semi‑infinite safe reinforcement learning that handles infinitely many constraints by iteratively solving finite subproblems. EPO expands or deletes constraints based on tolerance violations and Lagrange multipliers, maintaining computational tractability while converging to an optimal policy with bounded safety violations. The authors prove finite convergence, provide iteration bounds, and quantify the suboptimality gap under mild assumptions.
By Jiaming Zhang, Yujie Yang, Haoning Wang, Liping Zhang, Shengbo Eben Li
The paper presents CycleGRPO, a reinforcement learning framework that unifies region understanding and localization for multimodal large language models (MLLMs). By treating the MLLM as both actor and critic, the method generates region captions and immediately grounds them back into spatial coordinates, using a token‑level cycle‑consistency reward that obviates the need for textual ground truths. Experiments on SAMTok demonstrate that CycleGRPO can bootstrap region captioning, VQA, grounded dialogue, and referring segmentation simultaneously, achieving consistent performance gains without task‑specific fine‑tuning.
By Xin Zhang, Haochen Wang, Yikang Zhou, Zhuochen Wang, Xiangtai Li, Robby T. Tan
CroCo introduces cross‑lingual contrastive preference tuning on self‑generations, extending prior English‑only methods to 14 high‑ and low‑resource languages. A reward model trained solely on English preferences, applied to a multilingual base, yields effective within‑language rankings and improves performance in both monolingual and multilingual settings without catastrophic forgetting. The approach requires on‑policy data; off‑policy responses and online preference optimization offer limited gains, yet on structured tasks CroCo matches or surpasses the base model in most languages, and on open‑ended generation it wins 28/30 judge evaluations across 15 languages.
By Mike Zhang, Ali Basirat, Desmond Elliott
PRISM (Proactive Risk Intelligence and Safety Management) is an agentic multi-model architecture designed to shift autonomous transportation safety from reactive crash avoidance to proactive, continuous risk management. It uses inverse crash‑probability modeling to transform binary crash classifiers into dynamic safety scores, and runs three specialized models—trajectory kinematics, environmental risk, and VRU interaction—coordinated by a reinforcement‑learning reasoning layer. Across 1,296 naturalistic driving scenarios, PRISM achieved a mean safety score of 68/100, classified 77.6% of situations as advisory, and flagged 3.8% as near‑misses, with 11% requiring intervention or emergency response, highlighting trajectory risk and VRU proximity as key safety factors.
By Joyjit Roy, Samaresh Kumar Singh, Sushanta Das
WMLLM is a self‑evolving optimization‑agent framework that uses a predict‑then‑act world modeling approach. The agent first predicts promising optimization directions and then generates candidates, refining both its world model and strategy through multi‑turn refinement, population‑based search, and reinforcement learning. Experiments on black‑box tasks, particularly multi‑objective molecular optimization, demonstrate that WMLLM improves sample efficiency and achieves state‑of‑the‑art results within a limited evaluation budget.
By Zhongzheng Li, Qingsong Ran, Shikun Feng, Nian Ran, Wenhao Li, Xiaoyuan Zhang, Yue Wang, Xiaoguang Zhao
The paper introduces SPACE, a method for enabling large language model agents to emit variable-length action chunks in long-horizon tasks. By distilling chunk-boundary supervision from programmatic skills derived from successful trajectories, SPACE overcomes the tendency of agents to either act one step at a time or commit to overly long sequences. Experiments on ALFWorld and ScienceWorld demonstrate that SPACE raises success rates by 7.0%–31.3% and cuts LLM decision rounds by up to 78.9%.
By Yanting Yang, Can Jin, Jinman Zhao, Jiahao Wu, Yang Zhou, Zhepeng Wang, Zhendong Wang, Mu Zhou, Dimitris N. Metaxas
The paper introduces Document-Mediated Reinforcement Learning (DMRL), a framework that treats skill document optimization in advertising recommendation as a sequence of structured editing actions. An upper-level agent edits the skill documents while a frozen lower-level task agent evaluates the edits via A/B testing. DMRL incorporates Dual-Relative Policy Optimization for robust advantage estimation and a Long-term Reward Predictor that models population heterogeneity to estimate long-term outcomes, achieving superior performance on a large-scale short‑video ads platform.
By Wei Zhang, Hongji Li, Song Sun, Peng Yu, Xue Yang, Lei Zhao, Peng Jiang
Discriminative World Models for Web Agents proposes a new training objective called predicted‑state matching, which forces a world model to produce representations that can distinguish the true resulting web state from those produced by alternative actions. The authors train these models on a branching dataset from WebArena Go‑Browse, where each decision point includes multiple actions and their outcomes. Experiments show that models trained with predicted‑state matching outperform those trained with standard supervised next‑state prediction on a held‑out benchmark, improve PRM‑style action ranking on WebPRMBench, and enhance end‑to‑end task success on WebArena‑Lite when used for test‑time action selection.
Cliff is a reward‑shaping method for reinforcement learning with verifiable rewards that identifies the first mistake in a language model’s reasoning process using an off‑the‑shelf teacher. It splits each rollout into a correct prefix and an incorrect suffix, assigning positive token‑level advantages to the prefix and negative feedback to the suffix. Across 12 scenarios, Cliff improves reasoning performance, outperforming on‑policy distillation by 15% and standard GRPO by 7%, even with modest teachers.
SafeEvolve is an experience-driven framework that co‑evolves a harness and policy to improve safety alignment for LLM‑based agents. It uses on‑policy trajectory safety evidence to update safety prompts and hierarchical skills, producing auditable harness artifacts. The policy is trained via a two‑stage SFT‑RL pipeline that bootstraps with the evolved harness and then refines behavior through verifier‑decomposed rewards, yielding a better safety‑utility tradeoff on benchmarks such as AgentDojo.
The study couples the Met Office Unified Model with distributed reinforcement learning agents that apply bounded temperature corrections across 70 vertical levels. Training on ten nudged forecasts and evaluating on a non‑nudged run, the learned policy remains numerically stable and improves forecast accuracy, reducing Z₅₀₀ MAE by up to 45.8% in tropical bands and MSLP error by up to 27.3% in certain latitudes. This experiment demonstrates the feasibility of online RL for bias correction in operational weather models.
The paper introduces Multi-Teacher Self-Distillation Policy Optimization (MT‑SDPO), an on‑policy distillation method that combines multiple frozen teachers into a single student model. MT‑SDPO uses self‑anchors, answer‑verified eligibility, and privileged distillation to identify the most reliable teacher for each sample rather than relying on domain labels. Experiments on five students from three model families show that MT‑SDPO improves the weakest domain of Qwen3‑8B by 14.79 points and reduces its domain gap by 74.7%, achieving a better balance than matching a single teacher per domain.