Gaze-RL redistributes reward credit inside multimodal reasoning rollouts
Researchers propose a reinforcement learning method that infers where a trajectory-level reward should land within a model's own response, using internal cross-modal signals instead of external annotators or extra modules.
In plain words
In reinforcement learning for image and video reasoning, a reward is usually given to the whole answer, even though visual evidence appears only at a few spots inside it. Gaze-RL is a proposed method that decides which parts of the answer deserve more credit by reading the model's own internal computations. It needs no extra labels, perception module, or architecture changes.
What was announced
The authors frame the problem as one of reward allocation: outcome rewards are assigned to complete generated trajectories, while visual evidence is used only at sparse and uneven positions inside the response. Earlier work made multimodal RL more token-sensitive using grounding signals or external token-level weighting. Gaze-RL instead asks whether the same trajectory-level reward can be redistributed using the model's own multimodal computation, without an auxiliary annotator, perception module, or process label.
Technically, the method redistributes completion-level advantage through what the paper calls a latent reward landing field, computed from internal cross-modal geometry and local predictive uncertainty. Positions that are both visually anchored and decision-bearing receive larger update weight. The field stays mean-one, so the total credit of each rollout is preserved and only its positionwise allocation changes. The authors report improvements for both Qwen2.5-VL and InternVL3.5 across a broad range of image and video benchmarks.
What it does not show
- The source is only the paper's abstract, so no benchmark numbers, experimental setup, or independent verification of the reported improvements are available.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.