SIGMA is a hierarchical framework for cooperative multi‑agent reinforcement learning that addresses structured noise effects—local correlations in noise-induced decision impacts among agents with strong task dependencies. It groups agents into adaptive local structures using density‑based clustering, aggregates intra‑group representations to smooth deviations, and then applies inter‑group attention to integrate information while respecting heterogeneous contributions. Experiments on noisy‑observation StarCraft II tasks confirm that SIGMA improves robustness to observation noise without sacrificing performance in clean environments.
By Li Mingqian
AffectOmni is a reinforcement‑learning‑trained framework that enhances multimodal large language models for affective reasoning in social and art‑related scenes. It introduces People Focus and Temporal Order rewards to prioritize people‑centric cues and structured reasoning, and uses within‑group comparative scoring for more discriminative rewards. A Thinking Summarizer converts rationales into executable evidence instructions, which are grounded into pixel‑level regions via SAM3, enabling external auditability.
By Yibo Wang, Rui Yang, Jisheng Dang, Bimei Wang, Yitao Wu, Pengfei Cao, Wencan Zhang, Hong Peng, Bin Hu, Tat-Seng Chua
The paper introduces ASIL, an Agent‑Software Interaction Layer that replaces traditional screenshot‑and‑click interfaces with structured JSON observations and code‑executable semantic actions. ASIL is implemented across 15 applications and evaluated on 300 single‑application and 80 multi‑application tasks, achieving over 80% success with fewer than five actions per task. The structured interface also improves training efficiency, boosting performance of Qwen models from 58–66% to 72–80% with small‑scale supervised fine‑tuning and further gains with on‑policy reinforcement learning.
By Rui Xie, Lu Chen
DeepPlanner is an end-to-end reinforcement learning framework designed to enhance the planning capabilities of deep research agents. It introduces an entropy-based advantage shaping mechanism that allocates larger updates to high-entropy planning tokens and selectively upweights sample-level advantages during planning-intensive rollouts. Experiments on seven deep research benchmarks show that DeepPlanner improves planning quality and achieves state‑of‑the‑art results with a lower training budget.
By Wei Fan, Wenlin Yao, Zheng Li, Feng Yao, Xin Liu, Liang Qiu, Qingyu Yin, Yangqiu Song, Bing Yin
The paper introduces Planning Diffusion Policy Optimization (PDPO), an offline‑to‑online reinforcement‑learning framework that employs a diffusion policy to produce short‑horizon action chunks for robot crowd navigation. PDPO is pretrained on collision‑avoidance demonstrations and fine‑tuned online with PPO, generating five‑step action sequences applied in a receding‑horizon manner. The authors also identify a benchmark artifact where agents can leave the valid domain without explicit boundary constraints, and they mitigate this by treating boundary violations as collisions, leading to improved success rates over strong baselines.
By Wendong Li, Jochen Garcke
SPEAR (Symbolic Process Evaluation and Alignment Reward) is a training‑free, plug‑and‑play reward method for on‑policy distillation in reinforcement learning. It converts natural‑language reasoning traces into domain‑adaptive symbolic milestones and uses the longest common subsequence to align student exploration with teacher milestones, producing a dense, order‑aware reward that enforces logical consistency without an external neural verifier. Experiments on math, science, and commonsense reasoning tasks show that SPEAR effectively bridges the reasoning gap between student and teacher models through sequence‑level distillation with efficient dense process rewards.
By Zhuochun Li, Yuelyu Ji, Yiming Zeng, Daqing He
The paper introduces Private Best-of-N (PrivBoN), a method that adds calibrated Gumbel noise to reward scores during inference-time alignment, achieving both ε-differential privacy and KL-regularized alignment. When the privacy budget exceeds a critical threshold ε*, the noise becomes regret-optimal, matching the theoretical alignment skyline. The authors also propose Private Inference-Time Pessimism (PrivITP), which uses χ^2-regularized rejection sampling and a two-phase Gaussian mechanism to provide ex-post (ε,δ)-DP with a privacy cost independent of the number of responses, and demonstrate that both methods outperform standard Best-of-N across multiple models and datasets.
By Ishi Jain, Nandini Bhattad, Sayak Ray Chowdhury
PACIFIER is a graph‑learning framework that generates intervention sequences to reduce opinion polarization in online social networks modeled by the Friedkin‑Johnsen framework. It uses history‑aware node representations and both greedy and reinforcement‑learning strategies to score actions, supporting multiple moderation types and continuous opinions. Trained on small synthetic graphs, PACIFIER transfers effectively to large real‑world Twitter networks, outperforming baselines by up to 35.3% and achieving near‑oracle performance while being significantly faster.
By Mingkai Liao, Zijun Zhang, Mingyang Zhou
The paper introduces a closed‑loop framework using autotelic reinforcement learning to explore and manipulate complex systems, specifically Lenia, a continuous cellular automaton. An agent called CARL autonomously samples diverse goals and learns a goal‑conditioned policy that intervenes with minimal, local perturbations. CARL demonstrates three key abilities: discovering stable solitons more efficiently than heuristic baselines, steering existing solitons with few interventions, and enabling humans to guide solitons through maze environments in real time via high‑level commands. The agents generalize zero‑shot to out‑of‑distribution conditions, suggesting a path toward artificial experimentalist agents that can discover and control emergent phenomena.
By Marko Cvjetko, Benedikt Hartl, Michael Levin, Cl\'ement Moulin-Frier, Pierre-Yves Oudeyer
AirLLM introduces a hierarchical diffusion policy framework that adapts Low‑Rank Adaptation (LoRA) rank configurations for remote fine‑tuning of large language models over wireless channels. The system uses a Proximal Policy Optimization agent to make coarse decisions based on wireless and linguistic cues, then refines these decisions with Denoising Diffusion Implicit Models to produce task‑ and channel‑specific rank vectors. Experiments across different signal‑to‑noise ratios show that AirLLM improves fine‑tuning performance while substantially lowering transmission costs.
By Shiyi Yang, Xiaoxue Yu, Rongpeng Li, Jianhang Zhu, Zhifeng Zhao, Honggang Zhang
RubricRM introduces a pairwise generative reward modeling framework that generates an input‑specific rubric—comprising evaluation dimensions, weights, and scoring criteria—to score candidate images. The method is trained in two stages: supervised fine‑tuning to learn the rubric‑based scoring paradigm and GRPO to refine dimension‑level rewards. Experiments on text‑to‑image generation and instruction‑based image editing benchmarks demonstrate that RubricRM outperforms existing specialized reward models and competes with strong proprietary MLLM judges while using smaller backbones.
By Zijian Kan, Wei Wang, Long Luo, Bing Zhao, Xuan Ren, Weixu Qiao, Wenbo Li, Hu Wei, Lin Qu
PercepCap is a video captioning framework that explicitly models spatio‑temporal perception before generating captions. It follows a perceive‑describe chain, first producing a perception trace of object trajectories and temporal events, then generating the final caption conditioned on that trace. The method uses a two‑stage training strategy—supervised fine‑tuning followed by perception‑grounded reinforcement learning—and builds caption‑aligned perception data to ensure the perception trace and caption refer to the same objects and events.
By Yifan Xu, Zihao Wang, Zhixiao Wang, Jiaming Zhang, Yichun Yang, Desen Meng, Yuanxing Zhang, Pengfei Wan, Limin Wang
The paper introduces PATH, a curriculum‑learning framework that actively refines reinforcement learning by exploiting the directed acyclic curriculum graph formed by prerequisite relations among environments. PATH first expands coverage by sampling diverse curriculum paths and then reallocates training toward regions that remain unmastered. Experiments across diverse environments demonstrate that explicitly leveraging the graph structure leads to strong robustness and generalization.
By Zhenya Liu, Yuxin Chen
The paper investigates the Direct Preference Optimization (DPO) objective used for aligning language models, revealing that its coefficient β simultaneously controls both the inverse preference-noise scale and the optimization dynamics. This entanglement causes non‑monotonic policy deviation with respect to β and makes loss values incomparable across different β settings. The authors propose a centered‑softplus reformulation that decouples these effects, allowing independent tuning of the noise scale and learning‑rate, and provides a smooth β←0 limit that reduces to a linear preference‑margin objective.
By Ivan Kruzhilov
Behavior2Trip introduces a new task—Behavior‑Aware Travel Planning—where user preferences are inferred from past behavior trajectories rather than explicit instructions. The benchmark contains 11,400 instances from a major Chinese travel platform, each with nearly 40 recorded behaviors across 14 attributes and 5 preference dimensions. A reinforcement‑learning agent, B2T‑Agent, leverages these trajectories, external retrieval tools, and internal memory, outperforming GPT‑4.1 and other baselines on the dataset.
By Zihao Cheng, Yingyu Shan, Hongru Wang, Zeming Liu, Xinyi Wang, Xiangrong Zhu, Yuhang Guo, Wei Lin, Yunhong Wang
Think3D introduces a framework that endows Vision‑Language Models with interactive 3D chain‑of‑thought reasoning by integrating 3D manipulation tools for active spatial exploration. The approach improves performance on benchmarks such as BLINK Multi‑view, MindCube‑1K, and VSI‑Bench‑Tiny for proprietary models like GPT‑4.1 and Gemini 2.5 Pro, and a reinforcement‑learning variant, Think3D‑RL, enables open‑weight models such as Qwen3‑VL‑4B to autonomously learn effective 3D exploration strategies, yielding tool‑use patterns comparable to stronger models and turning a performance drop on MindCube‑1K into a substantial improvement.
By Zaibin Zhang, Yuhan Wu, Lianjie Jia, Yifan Wang, Zhongbo Zhang, Yijiang Li, Binghao Ran, Fuxi Zhang, Zhuohan Sun, Yizhuang Peng, Zhenfei Yin, Lijun Wang, Huchuan Lu
The paper introduces the Agentic Compositional Generalization hypothesis, suggesting that reinforcement learning (RL) primarily refines high‑level decision‑making behaviors that orchestrate pre‑trained low‑level skills, rather than teaching new domain‑specific skills from scratch. It proposes River, a training recipe that enhances reward quality by filtering low‑quality synthetic environments and adding process‑level behavior regularization. Using River, RL‑trained agents outperform other open‑source 8B models on four terminal‑agent benchmarks, achieving significant gains with fewer than 30% of the training environments.
By Yihang Yao, Bo Pang, Xuan Phi Nguyen, Ding Zhao, Shafiq Joty, Semih Yavuz
The paper introduces Test‑Time Policy Optimization (TTPO), a method that enables large language models to improve mathematical reasoning during inference without relying on ground‑truth labels. TTPO uses majority‑vote pseudo‑labels to guide an asymmetric objective: agreeing rollouts are distilled via On‑Policy Self‑Distillation, while disagreeing rollouts are penalized with Grouped Reinforcement Learning, with token‑level selection refining both branches. Experiments show that TTPO matches label‑supervised OPSD on five benchmarks, boosts Qwen3‑1.7B from 38.0 % to 45.2 % in test‑time training, and achieves significant gains without explicit reasoning steps, demonstrating strong cross‑task generalization.
We establish a global finite-sample guarantee for synchronous quantile temporal-difference learning (QTD) in tabular distributional reinforcement learning. The proof separates two stability mechanisms.
The paper examines the computational challenges of training Reasoning Language Models (RLMs) using reinforcement learning with verifiable rewards (RLVR) and similar post‑training methods. It provides a compute‑centric analysis of popular RL algorithms such as PPO and GRPO, and introduces a taxonomy of intra‑ and inter‑model parallelism strategies—including traditional and novel techniques—to improve scalability and cost‑efficiency. The authors also evaluate existing RLM frameworks and offer practical guidelines and research directions for building high‑performance, scalable RLMs.