The paper reviews how reinforcement learning (RL) can refine diffusion models (DMs) after pretraining, addressing limitations in reasoning, exploration, and alignment. It surveys RL-driven DMs across language, multimodal, robotic, and scientific domains, focusing on trajectory optimization and reward engineering. Experiments highlight the benefits of various RL post‑training strategies and identify key bottlenecks—reward fragility, exploration‑exploitation trade‑offs, and training instability—while proposing a roadmap for future research and a public repository for updates.
By Zhiyuan Ma, Jiabao Wei, Tianyu Wang, Hanqing Wang, Jiadong Pan, Haowen Xu, Liangliang Zhao, Ruixun Liu, Yuhan Wang, Yuwei Zhang, Liang Feng, Junhao Xiao, Kaiyan Zhang, Xuekai Zhu, Daizong Liu, Biqing Qi, Yihao Liu, Ganqu Cui, Ning Ding, Linfeng Zhang, Jianjun Li, Kun He, Xiang Bai, Bowen Zhou
OS-Helix is an end‑to‑end multimodal agent that unifies graphical user interface (GUI) actions and code execution into a single action space. It is trained on high‑quality hybrid data generated through application‑artifact composition and a decoupled instruction/evaluator pipeline, and further refined with online reinforcement learning using a step‑balanced GSPO algorithm. The resulting models achieve significant performance gains over hybrid baselines on the OSWorld‑Verified benchmark and demonstrate cross‑platform transfer to WindowsAgentArena.
By Bowen Yang, Kaiming Jin, Zichen Ding, Zhenyu Wu, Zehao Li, Zhoumianze Liu, Kanzhi Cheng, Fangzhi Xu, Yanan Sun, Qiushi Sun, Yu Qiao
The paper introduces a domain‑specific debugging agent for repairing near‑miss accelerator kernels, arguing that debugging is more efficient than regenerating kernels from scratch. It tackles knowledge scarcity, integrity, and cost control through pattern retrieval, anti‑cheat detection, and convergence safeguards. Empirical results show the debug agent achieves a 66.7% Pass@1 rate and uses 92.8% fewer tokens per success compared to traditional regeneration methods.
By Yansong Sun, Shenxiu Wu, Siyuan Chen, Runlin Hou, Junhao Qiu, Junming Cao, Shudi Shao, Zhichao Lu, Qingfu Zhang
The paper introduces Modality Self-Play (MSP), a training framework that enables multimodal large language models to learn cross‑modal interactions without requiring paired non‑text data. MSP uses proxy‑mediated textual descriptions of unseen modalities as semantic bridges, allowing a model to generate and align these proxies with observed modalities. Experiments show that MSP achieves state‑of‑the‑art or competitive performance on several multimodal benchmarks and demonstrates effective zero‑shot modality composition through learned bridges and anchor‑supporting proxy descriptions.
By Robert Moseley, Md Kaykobad Reza, Ameya Patil, Edward Ayrapetian, Salman Asif
The study evaluates six deep reinforcement learning (DRL) approaches for high‑performance computing (HPC) job scheduling, comparing Proximal Policy Optimization (PPO), Advantage Actor‑Critic (A2C), and Deep Q‑Network (DQN) with and without action masking. Across two real Slurm workload traces—one homogeneous and one heterogeneous—masked DRL treatments complete evaluation, while unmasked ones are truncated. On the heterogeneous trace, MaskablePPO outperforms baseline heuristics in all three metrics, achieving waiting times within 10% of the best heuristics and a 22% lower maximum bounded slowdown than Last‑Come, First‑Served (LCFS).
By Justin M. Cheney, Clement N. Nyirenda
All-in-One Memory (AiO-Mem) is a reinforcement learning framework that unifies memory retrieval, writing, and stopping decisions into a single reward space. It compiles diverse histories into traceable multi‑view events, uses a unified policy actor to generate interaction trajectories, and employs a trajectory‑conditioned process reward model trained with supervised fine‑tuning and DPO. Experiments on long‑context memory benchmarks show AiO‑Mem consistently improves answer accuracy compared to static memory systems and other RL‑trained agents.
By Bingbing Wang, Jingjie Lin, Weiming Qiao, Jing Li, Ruifeng Xu
ARGUS is a neuro‑symbolic, agenda‑driven framework designed to improve the reliability of user simulators in interactive environments. It separates state tracking from natural language generation by using a dynamic agenda stack and symbolic operations to enforce goal progression and constraint‑aligned interactions. The authors distill the model into compact versions via evaluator‑guided SFT and online RL, and evaluate its behavioral fidelity with a new Composite Behavioral Score metric on $ au^3$‑Bench and MultiWOZ, showing significant performance gains.
By Hang Zheng, Nan Jiang, Zichen Zhu, Zixu Yang, Zhiyang Tang, Shuai Fan, Lu Chen, Xiaobao Wu, Kai Yu
The paper introduces TIR-EC, a tool‑integrated reinforcement learning framework designed to improve large language models (LLMs) for emergency department prediction tasks. TIR-EC exposes groups of electronic health record (EHR) features as callable tools and trains the LLM using rewards from a gradient‑boosted tree teacher, while a reverse curriculum gradually removes prompt‑based skills to embed task‑specific clinical knowledge into the model weights. Experiments on three clinical prediction tasks show that a 4B TIR‑EC agent achieves the highest macro F1 among compared models, surpasses conventional machine‑learning baselines, and competes with much larger frontier LLMs, while also selecting more useful evidence with fewer tool calls.
By Pengfei Hu, Chenwei Wu, Qiao Jin, Jun Deng, Zhiyong Lu, Liyue Shen, Yue Ning
QREF is a plug‑and‑play system that improves DeepSearch agents by refining search queries to match the current reasoning intent and filtering evidence to keep only what is relevant. The Query Refiner rewrites queries for complementary coverage across each search batch, while the Evidence Filter removes distracting retrieval results. Experiments show that QREF boosts DeepSearch performance on four benchmarks, achieving significant Pass@1 gains and dramatically reducing ineffective query batches without changing the agents’ architectures or retraining their reasoning policies.
By Yujia Chen, Yuhui Wang, Junzhe Wang, Shihan Dou, Di Liang, Yujiong Shen, Yuhao Zhou, Zhiheng Xi, Ming Zhang, Senjie Jin, Jiayi Chen, Dingwei Zhu, Yajie Yang, Yuan Hua, Chenyao Lu, Tao Gui, Qi Zhang, Xuanjing Huang
DeciCAD introduces a domain‑native harness for CAD agents that defines semantic CAD actions, decision‑oriented observations, environment feedback, and persistent state transitions. Using this harness, the authors develop a learning pipeline that extracts causal decision supervision from interaction trajectories and refines the policy through supervised fine‑tuning and geometry‑grounded online reinforcement learning. Across four CAD benchmarks, DeciCAD outperforms a general‑purpose coding harness, achieving higher F1 and IoU scores while reducing model‑output tokens by 45–88% and improving geometry availability from 66.5% to 73.5%.
By Yufeng Peng, Yixin Zhang, Ziyi Liu, Zhikun Xu, Jiarong Yang, Sijie Wang, Wenjie Feng, Qiyu Kang
The paper investigates why reinforcement learning with self‑generated labels—specifically majority‑vote rewards—can recover most of the performance gains of training with verifiable rewards. It argues that the key factor is batch‑level updates: a prompt’s own gradient contributes less than 1% to its output change, so only coherent label errors across prompts influence learning. Experiments on five models show that majority‑vote updates align almost perfectly with gold updates, with misdirection disappearing as batch size grows and incoherent errors behaving like random noise.
By Dylan P. Wilson, Chengsong Huang, Jiaxin Huang
The paper introduces HOW TO TRAIN YOUR LITTLE DRAGON, a method for teaching a small language model to generate and refine GPU optimization scripts. Using a compact, lossless representation of compiler traces, the authors train the model on 108,594 examples derived from TVM MetaSchedule, first with supervised fine‑tuning and then with reinforcement learning that rewards correctness and performance. After training, an 8‑billion‑parameter model achieves over 91% correctness in both construction and refinement tasks and outperforms several much larger LLMs on kernel performance.
By Annabelle Sujun Tang, Wei Tang, Songyao Jiang, Tianhui Li, Shengzhe Xu, Jiin Woo, Kaan Ozkara, Youngsuk Park, Zhen Jia, Hadi Esmaeilzadeh
The paper introduces a Bellman-consistent trajectory reinforcement learning approach for large language model (LLM) arithmetic reasoning, called Trajectory Bellman Residual Minimization (TBRM). By converting sparse terminal rewards into dense token-level signals and employing a Reinforce Leave-One-Out (RLOO) strategy to reduce gradient variance, the method improves learning stability. On the Countdown task, TBRM-RLOO achieves Pass@1 scores of 0.55–0.64 and a state-of-the-art Pass@16 accuracy of 76.0–84.0%, surpassing several baseline methods.
By Lianfa Li
MAI-Gym is an open, all-in-one platform that unifies GUI agent evaluation, data collection, and online reinforcement learning training across mobile, computer, and web environments. It supports ten established benchmarks, scales to 20,000 parallel environment instances, and enables automated synthesis of diverse GUI applications and tasks with execution-validated verifiers. Using this platform, the authors trained a 9B-parameter agent, MAI-Gym-9B, achieving state‑of‑the‑art performance on MobileWorld and OSWorld-Verified, and demonstrated that training on a mocked train‑ticketing app improves real‑world task success from 28.8% to 48.1% on a physical phone.
By Panrong Tong, Quyu Kong, Pengxiang Li, Tianyu Xia, Hanzhang Zhou, Lei Wang, Chenglin Cai, Long Li, Long Chen, Jianan Zhang, Zhengwen Zeng, Xu Zhang, Gongjie Zhang, Yue Wang, Steven Hoi
Rubric-based evaluation of large language models (LLMs) often assigns binary judgments to multiple criteria, but the certainty of these judgments varies widely. Existing aggregation methods treat all criteria as equally certain, potentially misleading reinforcement learning signals. The authors propose Uncertainty-aware Rubric Weighting (URW), a training-free framework that uses conformal calibration to quantify and weight criterion-level uncertainty, improving preference accuracy and downstream RL alignment.
By Licheng Pan, Haocheng Yang, Haoxuan Li, Yuzhou Chen, Yichen Sun, Yuan Lu, Xiaoxi Li, Hao Wang, Zhixuan Chu, Kui Ren
The paper introduces an open midtraining recipe that primes base models for agentic post‑training, filling a gap between pretraining and full fine‑tuning. It trains on 256K context with hundreds of billions of tokens drawn from natural software‑engineering data and a small share of agentic traces. Applying this midtraining to Nemotron 3.5 Lightning and Super models yields measurable gains on SWE‑bench and DeepSWE benchmarks after identical full SFT, and improves robustness across agent harnesses.
By Soumye Singhal, Joseph Jennings, Ka Shun Shum, Alexander Bukharin, Khushi Bhardwaj, Zirui Liu, Venkat Srinivasan, Jonathan Li, Yash Jain, Kezhi Kong, Markus Kliegl, Olivier Delalleau, Sanjeev Satheesh, Mostofa Patwary
The paper introduces Subspace Tuning (SST), a fine‑tuning technique that restricts weight updates to the bottom singular subspace of model parameters, thereby protecting the principal directions that are most prone to catastrophic forgetting. SST is optimizer‑agnostic, works with existing fine‑tuning recipes such as SFT and self‑distillation, and includes dense gradient estimation and a refresh mechanism to enhance learning. Experiments show that SST can preserve up to 99 % of the weight subspace, significantly reduce KL drift, and nearly eliminate forgetting in large‑scale language models while maintaining strong target‑task performance.
By Arnav Chavan, Kusha Sareen, Rishabh Tiwari, Abhranil Chandra, Johan Obando-Ceron, Kurt Keutzer, Scott Niekum, Inderjit S Dhillon, Devvrit Khatri
The paper proposes a framework for aligning AI agents with deployment requirements through a process called calibration, which avoids the need for source imitation or universal superiority. It introduces three families of requirements—basic capability, technical precision, and user context—alongside a versioned mapping to manage changes. The authors outline four instruments (benchmarks, examples, skills, and role-conditioned cost/utility) that provide tunable reward channels within a fixed acceptance contract, and describe methods for adaptation, qualification, and evaluation, noting that confirmatory validation remains outstanding.
By Yaxiao Liu (PwC China AI Center), Pengbo Liu (PwC China AI Center), Yiwen Liu (PwC China AI Center), Yihua Guan (PwC China AI Center), Jiaxing Song (Tsinghua University), Xingjian Li (PwC China AI Center)
The paper introduces Future‑L1, an interleaved latent visual reasoning framework that alternates between language tokens and continuous latent visual spans during autoregressive decoding for video event prediction. By training on a curated dataset (Future‑L1‑50K) and optimizing latent trajectories with a latent‑aware RL objective (LA‑DAPO), Future‑L1 achieves state‑of‑the‑art results on FutureBench and TwiFF‑Bench, outperforming prior models by significant margins. The work demonstrates that preserving intermediate visual semantics in latent space, rather than verbalizing every reasoning step, leads to more accurate and visually grounded future predictions.
By Tianxiang Jiang, Linquan Wu, Sheng Xia, Songze Li, Ziang Yan, Haoyu Yang, Yu Qiao, Yi Wang
V‑SHARE is an advantage‑based value‑learning algorithm that blends cross‑action value sharing with an importance‑weighted Bellman‑residual correction. The paper presents the first full theoretical analysis of such an algorithm, offering both asymptotic convergence and finite‑sample guarantees. It also introduces schedules that balance deterministic contraction against importance‑weighting noise, and a two‑timescale analysis that eliminates a residual‑weight threshold while maintaining convergence.
By Kintan Saha, Hamed Masoudi, Xinyi Sheng, Dominik Baumann, Joni Pajarinen, Arsenii Mustafin