When Thinking Hurts: Mitigating Visual Forgetting in Video Reasoning via Frame Repetition
This addresses visual anchor drifting in Video-LLMs, which causes hallucinations and performance degradation, offering a generalizable solution with lower training costs compared to existing methods.
The paper tackles the problem of visual forgetting in video reasoning by proposing FrameRepeat, an automated framework that uses a lightweight repeat scoring module to identify and reinforce important frames, resulting in improved performance across multiple models and datasets.
Recently, Multimodal Large Language Models (MLLMs) have demonstrated significant potential in complex visual tasks through the integration of Chain-of-Thought (CoT) reasoning. However, in Video Question Answering, extended thinking processes do not consistently yield performance gains and may even lead to degradation due to ``visual anchor drifting'', where models increasingly rely on self-generated text, sidelining visual inputs and causing hallucinations. While existing mitigations typically introduce specific mechanisms for the model to re-attend to visual inputs during inference, these approaches often incur prohibitive training costs and suffer from poor generalizability across different architectures. To address this, we propose FrameRepeat, an automated enhancement framework which features a lightweight repeat scoring module that enables Video-LLMs to autonomously identify which frames should be reinforced. We introduce a novel training strategy, Add-One-In (AOI), that uses MLLM output probabilities to generate supervision signals representing repeat gain. This can be used to train a frame scoring network, which guides the frame repetition behavior. Experimental results across multiple models and datasets demonstrate that FrameRepeat is both effective and generalizable in strengthening important visual cues during the reasoning process.