The paper introduces MoEvolve, a bidirectional distillation framework that leverages increased activation budgets in Mixture-of-Experts (MoE) models to create a stronger internal teacher. By first restoring a high‑k model and then transferring its enhanced capabilities back to the original low‑k configuration, MoEvolve improves performance across various MoE models and benchmarks, achieving up to 22.2 domain‑average points and outperforming same‑budget RLVR baselines by up to 6.5 points. The resulting low‑k models often surpass their activation‑expanded teachers.
By Haoyu Chen, Jun Bai, Yang Liu, Jianfei Zhang, Wenge Rong
The paper introduces LM2LM, a method that trains large language models (LLMs) to propose improvements for other LLMs using reinforcement learning (RL). It employs a frozen research world model (RWM) to predict the gains of proposals, providing a cheaper surrogate reward, and periodically validates high‑scoring proposals to keep feedback reliable. Experiments across eight LM development environments show that LM2LM enables a Qwen3.8‑27B model to outperform GPT‑6‑sol on held‑out tasks while reducing data‑generation cost by 33%, and it generalizes to unseen model families.
By Minhua Lin, Zijun Wang, Zewen Liu, Zhaotian Weng, Bing He, Zhan Shi, Benoit Dumoulin, Xiang Zhang, Suhang Wang, Hanqing Lu
R$^2$ Flow is a recursive self‑improvement framework for large‑language‑model agents that alternates policy learning, independent verification, and versioned skill‑library updates on a shared‑state orchestration graph. By merging histories that differ only in the order of independent steps, it pools evidence across equivalent executions and uses flow‑share readouts, signed utility ranks, verifier evidence, and residual‑variance plateaus to decide when and which skills to edit. Experiments on question answering, mathematical reasoning, interactive decision making, and code generation show that R$^2$ Flow improves task accuracy and library‑edit precision compared to heuristic orchestration, reinforcement learning, and other skill‑evolution baselines, and it transfers across different executors.
By Mingda Zhang, Qiang Huang, Yanjin Li, Zijia Wang, Qika Lin, Xiaoying Tang, Tiesunlong Shen
The paper introduces EDIC, a framework for per‑instruction compliance in large language models that interact with external data. It decomposes the task into identifying, classifying, and acting on each embedded directive, using a scene‑grounded pipeline and a generative judge trained with GRPO to attribute compliance at the instruction level. Experiments on two backbones and five benchmarks show that EDIC improves safety and utility, achieving up to a 7.7% increase in harmful‑rejection and a 12.5% boost in Arena‑Hard‑Auto utility on Qwen3‑4B.
By Zesheng Shi, Xue Li, Chenrui Dai, Jing Li
SpeechCritic learns a diagnostic speech judge from only about 300 human-labeled comparisons, calibrating a frontier audio-language model with non-binding hints that select acoustic measurements agreeing with human judgments. The approach improves dimension-level agreement by 6.3 points and reduces mismatch in Tie rates by 10.4 points, and training a 7B judge with SFT, OPD, and RL yields distinct behaviors, with RL producing more specific rationales. The method works across language pairs, demonstrating a path from limited human preferences to a diagnostic speech judge.
By Mingyue Huo, Shivam Mehta, Bhavin Jawade, Yinghong Lan, Haoqi Li
The paper introduces a framework for reinforcement learning in Markov decision processes where policies are ordered by the information they provide, allowing data from one policy to estimate the performance of others. An epoch‑based algorithm is developed for infinite‑horizon average‑cost MDPs with continuous state and action spaces, achieving regret that scales with the width of the policy class’s information order. The authors apply the framework to lost‑sales inventory, dual‑sourcing, and state‑dependent queueing systems, demonstrating improved regret bounds and empirical performance without requiring global convexity or specific arrival‑rate structures.
By Zhongjun Zhang, Shipra Agrawal, Ilan Lobel, Sean R. Sinclair, Christina Lee Yu
arXiv:2608. 12973v3 Announce Type: replace-cross Abstract: In this paper, we study online statistical inference for quantile-based distributional policy evaluation, which aims to characterize the distribution of discounted cumulative rewards under a given policy.
By Zijie Cheng, Yang Peng, Zhihua Zhang
RoutePack is a replay‑guided planner that coordinates expert placement and rollout‑trajectory scheduling for Mixture‑of‑Experts reinforcement learning models. By incrementally adjusting expert placement while preserving residency and then rescheduling trajectories into token‑capped micro‑batches, it jointly balances dense‑side work and layer‑wise expert peaks under fixed capacity constraints. Evaluations on 7.9B and 124B hybrid‑attention MoE models across NVIDIA H20‑3e configurations show throughput gains of 4.44‑11.85% over a DP‑local first‑fit‑decreasing baseline and 1.33‑5.28% over expert rerouting alone.
By Yibo Shen, Xudong Han, Xiaowei Zhu, Gen Li, Zhenxuan Pan
The paper introduces Grounded Credit Assignment (GCA) for reinforcement learning in multi‑turn search agents. GCA separates credit for evidence retrieval and answer correctness, correcting the bias in GRPO that favors evidence‑retrieving rollouts only proportionally to accuracy gains. Applying GCA to Search‑R1 boosts average EM by 5.7–6.3 points and increases the proportion of evidence‑supported correct answers from 76.9% to 93.4%.
By Qirui Liu, Hao Chen, Yi Wen, Weijie Shi, Yao Zhao, Xianhe Chen, Hao Zhao, Xing Tang, Weihong Luo, Yuyu Luo
Process Reward Models (PRMs) are crucial for scaling large language model reasoning at test time, but they suffer from two types of uncertainty: PRM uncertainty from limited supervision and imperfect training, and policy uncertainty from mismatched data collection policies. The paper introduces BayesianPRM, which treats process rewards as random variables, discretizes the reward space with an ensemble, and applies two belief updates—reliability‑aware updating for PRM uncertainty and conservatism‑aware calibration for policy uncertainty—to form a Bayesian belief whose expectation yields a more accurate reward estimate. Experiments on multimodal reasoning benchmarks show that BayesianPRM improves reward accuracy and test‑time scaling performance.
By Zhongjian Qiao, Zhixiang Zhou, Jiafei Lyu, Peisong Wang, Qiaosheng Zhang, Siyang Gao, Shuang Qiu
DiffusionRVGE presents a reinforcement learning approach that improves diffusion models by estimating value gradients through regression of centered rewards on centered final denoised images, avoiding costly reward derivatives and backpropagation through the sampling trajectory. The method adapts rollout bandwidth and query weights based on reward continuity, enabling efficient training across multiple rewards. Experiments show that a single SD3.5‑Medium model trained with DiffusionRVGE matches or surpasses DiffusionNFT on most metrics while using roughly one‑fifth of the training compute, and achieves faster reward convergence in single‑reward settings.
By Jiayuan Wang, Tongyao Pang
The paper introduces a cost‑aware controller called certified predictive advice gating that decides when to request language‑model advice in reinforcement learning. By forecasting advisor responses, evaluating parsed actions, and reusing cached advice, the gate only requests advice when a calibrated lower bound on the value exceeds the cost. Experiments on BabyAI tasks show that the gate improves return while reducing advice calls by over 97 % compared to always asking, though matched‑budget comparisons do not show a clear advantage over an early‑query schedule.
By Ibne Farabi Shihab, Md Najmus Swaqeeb, Abu Sa-Adat Mohamed Moon-Im Al Ahsan, Joyanta Jyoti Mondal
CastFlow is a role‑specialized agentic forecasting framework that combines general‑purpose tool reasoning with specialized temporal forecasting. It uses an Experience Memory to guide a frozen LLM in acquiring evidence via a multi‑view toolkit, then a lightweight specialized LLM refines predictions through evidence‑grounded reasoning. The two‑stage training strategy—supervised fine‑tuning followed by reinforcement learning with verifiable rewards—optimizes the specialized model based on final forecasting outcomes, achieving superior performance on nine out of ten real‑world benchmarks.
By Bokai Pan, Mingyue Cheng, Xiaoyu Tao, Zhiding Liu, Shuo Yu, Yuchong Wu, Qi Liu, Defu Lian, Enhong Chen
The paper introduces EquivAug, a data augmentation framework that applies symbolic equivalence transformations to chain-of-thought (CoT) solutions, generating diverse yet provably correct reasoning paths for supervised fine-tuning (SFT). By converting CoT solutions into formal symbolic representations, applying equivalence transformations, and reconstructing natural-language explanations, EquivAug produces training data that varies in reasoning steps rather than surface-level lexical differences. Experiments on GSM8K and FOLIO show that EquivAug consistently achieves the highest reasoning path diversity after SFT and the strongest accuracy after GRPO, establishing a positive correlation between CoT diversity during SFT and downstream reinforcement learning with verifiable rewards (RLVR) performance.
By Xiaoou Liu, Tiejin Chen, Hanyang Chen, Yuan Ling, Hua Wei
NestedVLA proposes a new approach for consolidating skills in Vision‑Language‑Action models by learning skills independently and then generating task‑specific parameters via a nested hypernetwork, rather than aggregating parameters. The method keeps a frozen base VLA and task‑specific compressed VLM residuals, enabling flexible, interference‑mitigated skill reuse. Experiments on MetaWorld, ManiSkill, and CALVIN show that NestedVLA outperforms the best model‑merging baseline by 41 pp on average and approaches per‑task fine‑tuning performance.
By Guodong Du, Wanyu Lin
The study investigates whether five large language‑model agents adapt their contributions in a collective‑risk dilemma game. Across ten rounds, most models maintained a roughly fixed contribution level regardless of changing risk or partner behavior, and many continued to contribute even when it offered no payoff advantage. The results reveal that cooperative self‑play does not necessarily indicate strategic adaptability, and that agents’ responses to incentives, partner actions, and population composition vary significantly.
By Trung-Kiet Huynh, Dao-Sy Duy-Minh, Chi-Nguyen Tran, Phu-Hoa Pham, Phu-Quy Nguyen-Lam, Le Hong Trang, The Anh Han
ECHOES is a method for long‑horizon GUI agents that compresses visual history into a textual state using a Compression‑Think‑Action format. It employs a privileged teacher model that provides token‑level on‑policy distillation to guide the student’s textual rollouts, while a Perception‑Memory Fidelity Gap‑guided framework balances distillation with task reward optimization. The approach is evaluated on four GUI benchmarks and a new ECHOESBMK diagnostic benchmark, showing improved task performance, memory fidelity, and a favorable trade‑off between performance and inference cost, with positive zero‑shot transfer to unseen planning tasks.
By Tengda Guo (Tsinghua University), Jiawang Hu (StepFun), Yifei Tao (Nanyang Technological University), Yin Wang (StepFun), Haotian Hu (Zhejiang Leapmotor Technology Co., Ltd), Fanyi Wang (Honor), Liujuan Tang (StepFun)
GlobeBench is a new benchmark for evaluating text-based world models, comprising 1,632 tasks across five domains and nineteen simulation skills. It tests whether models can predict an environment’s response to an action based on the full agent trajectory, requiring understanding of state changes, environment rules, and action effects on related and unrelated objects. The benchmark compares predictions to execution-derived references using a hybrid architecture of deterministic gates and a semantic judge, and finds that GPT‑6 Astra tops the 20 tested models with a 49.57% success rate, while accuracy drops sharply when actions affect many dependent objects, leaving 631 tasks unsolved by any model.
By Nicholas Saban, Charvi Bannur, Mariya Vasileva, Darshan Deshpande
R$^2$Orch introduces a Recursive Self-Play Orchestration framework that uses an inner self‑play loop to improve an orchestrator on self‑selected tasks and an outer loop that refines how the inner loop improves. The method employs Recursive Learnability to reward task selection that yields diverse outcomes as skills are removed, and Skill Withdrawal to prune skills when the orchestrator performs equally well without them. Experiments on twelve datasets demonstrate that R$^2$Orch outperforms existing baselines in question answering, math, code, and interactive tasks.
By Mingda Zhang, Yun Huang, Qiang Huang, Yanjin Li, Rui Qian, Qichao Zhao, Tiesunlong Shen, Na Zhao, Jian Wang, Xiaoying Tang
Reward Engine is a method for creating a reward function for open‑ended tool agents by first determining what success looks like and then optimizing toward it. It builds a RewardPack of test cases and rubrics from initial trajectories, which is frozen and used to guide the agent’s output selection and can also serve as a reinforcement‑learning reward. Experiments on WorkBuddy Bench Office and Claw‑Eval show that the constructed reward aligns with benchmark graders on most candidate pairs and improves three‑run consistency from 45.0% to 54.0% and from 78.4% to 83.7%.
By Litong Feng, Ji Qi, Yiran Cheng, Shaowu Pan, Xinyu Zhang, Qijun Li