The paper introduces Outcome-Conditioned Visual Credit (OVC), a plug‑in framework for multimodal reinforcement learning that separates visual association measurement from token‑level credit assignment. OVC uses Relative Visual Encoding to create a context‑relative visual coordinate and Outcome‑Conditioned Evidence Calibration to adjust token contributions based on outcome‑discriminative evidence. When applied to GSPO-token, OVC improves mean accuracy by 1.86–2.39 percentage points across six benchmarks without requiring additional reward models or inference‑time modules.
By Yuhang Han, Dingning Liu, Yujie Chen, Yaojie Zhang, Jingwei Song, Zequn Yang, Xuanming Shang, Li Ding, Zhijiang Guo, Linfeng Zhang
The paper demonstrates that reinforcement learning with verifiable rewards (RLVR) can improve vision‑language benchmark performance even when models are trained without visual input. When real images are introduced at test time, models trained blind recover about half of the performance gain at 3B parameters and nearly four‑fifths at 7B, but extended real‑image training can erode grounding while benchmark gains persist. The authors propose a visual resolvability rule and show that requiring visual evidence for correct answers leads to significant improvements in target discovery and generalization to unseen question types, while controls confirm that the gains stem from actual visual grounding rather than artifacts.
By Haocun Ye, Xinlong Jiang, Qile Chen, Bingyu Wang, Teng Zhang, Shubai Chen, Tingyu Wu, Zhenkun Zheng, Yiqiang Chen
arXiv:2608.21595v1 Announce Type: new
Abstract: Reinforcement learning with verifiable rewards (RLVR) improves the reasoning ability of vision-language models (VLMs), and diversifying the rollouts wi...
By Michael Jerge, Joseph Pelczar, Justin Downes
The paper introduces Perception and Reasoning Attribution (PARA), a method that quantifies how much of a vision‑language model’s errors can be fixed by better perception versus stronger reasoning. PARA reveals that reinforcement learning (RL) benefits more from improving reasoning than perception, and motivates a new group‑relative attention distillation objective called GRAD. GRAD directly guides weaker rollouts toward the visual attention patterns of the best rollouts, consistently boosting performance across multiple model scales and benchmarks.
By Zeyuan Yang, Zhuohao Chen, Jiaben Chen, Yao Zhang, Maohao Shen, Chuang Gan
The paper introduces RECAP, a training method that enhances vision‑language models by incorporating temporally localized reactivation of image evidence—called visual revisit—into the reinforcement learning process. Visual revisit peaks are shown to predict correctness and causally improve performance; masking them hurts accuracy more than masking non‑peaks, while adding revisit peaks to failed reasoning partially restores correct answers. RECAP applies revisit‑conditioned advantage estimation, dynamic discounting, and gated value prediction to propagate final rewards through visually important reasoning steps, achieving significant gains on HallusionBench and other benchmarks without extra rewards or inference‑time changes.
By Zhaolu Kang, Tailong Luo, Chenxin Li, Yingjie He, Zhenyu Yu, Fengyu Zhou, Jiachen Qian, Lei Wei, Shuang Chen, Jiachen Li, Eric Hanchen Jiang, Rongchao Zhang, Zhengtao Yao, Hoi Leong Lee, Guansu Wang, Kaiyue Zhou
The paper introduces EASE, a method that enhances multimodal reinforcement learning with verifiable rewards (RLVR) by adding visual‑evidence process supervision. EASE transforms annotated evidence regions into smoothed visual‑token targets and uses them to guide attention during RL training, but only on high‑reward trajectories. Experiments on Qwen2.5‑VL‑7B, Qwen3‑VL‑4B, and Qwen3‑VL‑8B show that EASE improves average scores over DAPO by 2.5 to 3.1 points across perception, hallucination, visual math, and multimodal reasoning benchmarks, and diagnostics confirm better alignment of visual attention with annotated evidence.
By Ruina Hu, Chen Wang, Lai Wei, Jionghao Bai, Bin Yu, Weiran Huang, Kai Wang, Yue Wang
The paper introduces Activation Replay, a training‑free method that improves reasoning in post‑trained large multimodal models (LMMs) by replaying low‑entropy activations from the base model’s input context. It shows that Reinforcement Learning with Verifiable Rewards (RLVR) shifts low‑entropy activations and that modulating these activations enhances reasoning across tasks such as mathematics, visual agents, and video reasoning. Experiments demonstrate that Activation Replay outperforms alternatives like high‑entropy replay or direct cross‑model intervention, boosting Pass@K and broadening RLVR’s reasoning coverage.
By Yun Xing, Xiaobin Hu, Qingdong He, Jiangning Zhang, Shuicheng Yan, Shijian Lu, Yu-Gang Jiang
SPLIT-RL introduces a staged post‑training method that separates visual reasoning (VR) and language reasoning (LR) into distinct phases, allowing each phase to be optimized with a phase‑specific reward. The approach includes Claim‑Level Advantage (CLA‑GRPO), which breaks VR rollouts into atomic visual claims and assigns fine‑grained advantages based on visual‑type group formation. Experiments on Qwen3‑VL and InternVL models show that SPLIT‑RL improves average accuracy by 1.4–6.1 points over GRPO and enhances both VR and LR capabilities.
By Raja Kumar, Rajat Koner, Ritwick Chaudhry, Zhuowei Li, Nishant Sankaran, Yifan Xing
arXiv:2609.40360v1 Announce Type: cross
Abstract: Reinforcement learning with verifiable rewards (RLVR) has improved the reasoning capabilities of large language models (LLMs), yet their predictions...
By Junshu Pan, Zhizhang Fu, Shulin Huang, Yiran Ding, Zifan Cheng, Wenqi Shao, Qiaosheng Zhang, Yue Zhang
The paper investigates how reinforcement learning can unintentionally obscure the chain‑of‑thought (CoT) reasoning in vision‑language models, making their internal reasoning less traceable. By analyzing activation patterns, the authors show that template‑associated activations become less distinguishable during RL and that targeted interventions can mitigate this effect. They introduce TAME, a method that uses sparse autoencoders to suppress these problematic activations while still encouraging accurate behavior, achieving significant gains in CoT monitorability across multiple datasets and model families.
By Xutao Mao, Jianing Zhu, Jinman Zhao, Tongliang Liu, Xiaowen Chu, Cong Wang, Bo Han
arXiv:2609.13308v1 Announce Type: cross
Abstract: A companion evaluation found that naming the target part in a manipulation prompt increased action accuracy by 0.32-0.63 across eight vision-language...
By Sarthak Sattigeri
The paper investigates how reinforcement learning with verifiable rewards can improve multimodal large language models (LLMs) by separating perception and reasoning stages. Experiments on algorithmic visual puzzles show that perception, not reasoning, limits performance, and replacing images with textual descriptions boosts Claude models by over 20 points. Reward designs that encourage visual grounding during reasoning, such as GRPO for Qwen-2.5-VL-7B, yield structured reasoning and a 5.56‑point gain, though benefits vary across categories and trade‑offs exist between cross‑domain transfer and in‑domain accuracy.
By Omar Sharif, Eftekhar Hossain, Nikhil Singh, Patrick Ng