The Model Is Its Own Reward - Online Reinforcement Learning of Visual Token Pruning for Multi-Image VLMs
Nitish Shukla et al.; Under review
Research Goal
Visual token pruning rules for multi-image models are usually fixed by hand. This work learns the pruning rule instead, at test time, without labels or external verifiers.
How it works
- Online reinforcement learning. A pruning policy is trained online while the model runs.
- Internal rewards. The reward comes from the frozen model’s own internal signals, so no labels or verifiers are needed.
Status: under review.