Junfeng Wang, Jiawei Liu, Yongchao Xu, Tao Jiang, Jiangbo Ai, Jin Zhang, Chong Wang, Qixing Zhang
Complex visual question answering often requires vision-language models to think with images, i.e., actively acquire local visual evidence by zooming into informative regions of high-resolution images. Our error analysis shows that many failures in this setting originate from incorrect observation actions, where the model fails to localize key regions or obtain crops that contain sufficient evidence for answering. We further observe that bounding-box tokens exhibit substantially higher generation entropy than ordinary language tokens, indicating high uncertainty when the model decides where to look. However, existing reinforcement learning methods based on final-answer rewards provide only trajectory-level feedback, making it difficult to assign precise credit to the observation actions. We propose Observation Policy Optimization (OPO), an RL algorithm for optimizing observation actions in tool-augmented thinking with images. OPO consists of two core modules. First, Uncertainty-Guided Observation Branching (UOB) treats each image zoom-in tool call as an observation event and selectively branches at high-uncertainty, low-redundancy events during rollout, enabling efficient exploration of alternative cropping decisions and their resulting visual evidence. Second, Evidence-Localized Advantage Attribution (ELAA) compares sibling branches using local evidence-sufficiency rewards and assigns credit only to the corresponding observation-action tokens. By decoupling observation-action optimization from trajectory-level outcome supervision, OPO improves visual evidence acquisition and further enhances visual reasoning performance. Experiments show that OPO consistently improves Qwen3-VL models across scales and achieves competitive performance against larger models and tool-augmented baselines. Code is publicly available.
A poor start lowers the ceiling, but does not negate the value of finishing well. In autoregressive video temporal grounding, models systematically exhibit an endpoint performance asymmetry, predicting ends significantly worse than starts. A committed flawed start acts as an immutable physical constraint, irreversibly capping the attainable overlap as a sunk loss. While models should minimize the remaining conditional regret, standard post-training paradigms fail to teach this adaptation. Specifically, teacher-forced SFT suffers from strict prefix exposure bias, and holistic trajectory evaluation in standard GRPO inadvertently penalizes conditionally optimal ends if they follow a poor start. To address this, we introduce FinishStrong, a novel two-stage post-training framework explicitly designed to teach optimal continuations. First, Online-start SFT exposes the end prediction to the model's own generated contexts. Second, Conditional-tree GRPO restructures the sampling topology, centering end rewards exclusively within the same-start branch to mathematically isolate the conditional end quality from the inherited start penalty. FinishStrong-9B achieves a state-of-the-art 61.7\% overall IoU on TimeLens-Bench, remarkably outperforming flagship closed-source models like Gemini-3.8-Flash and Qwen3.8-Max. Crucially, when forced to continue from imperfect starts, it reduces the subsequent end prediction error by 36.5\%, proving its exceptional ability to end well after flawed starts. Code will be open.