LASER: A Corrective Lens for LVLMs via Visual Attention Preservation and Sink Suppression

2026-07-02

AI Summary

This work addresses the issue of visual forgetting in large vision-language models during long-sequence reasoning, which arises from visual attention drift and leads to performance degradation. The authors propose LASER, a post-training framework that mitigates this problem by regulating visual attention trajectories and internal token distributions. Analysis reveals that early attention decay and excessive focus on irrelevant sink tokens are primary causes. To counteract these effects, LASER incorporates a dual-reward mechanism: a visual anchoring reward preserves attention on semantically salient regions, while a sink suppression reward prevents attention collapse into uninformative areas. Extensive experiments across eight benchmarks demonstrate that LASER significantly outperforms strong baselines, validating the efficacy of attention-aware training in alleviating visual forgetting.

Abstract

Large vision-language models (LVLMs) exhibit strong reasoning ability but suffer from visual forgetting during long-horizon decoding, where attention progressively drifts away from visual evidence. Existing methods largely treat this issue as a late-stage attention decay problem or attempt to mitigate it through heuristic reminders or post-hoc attention lifting. Through systematic empirical analysis, we find that performance degradation under visual forgetting is largely driven by two overlooked factors: early-stage attention decay disrupts evidence acquisition, and attention concentration on a subset of task-irrelevant visual sink tokens. Motivated by these insights, we propose LASER, a post-training framework that regulates both the visual attention trajectory and intra-visual token attention distribution during reasoning. Technically, LASER introduces two complementary rewards: a Visual Grounding Reward, which encourages the model to maintain attention on semantically salient visual tokens throughout decoding, and a Sink Suppression Reward, which penalizes excessive attention concentration on visual sink tokens. Together, these rewards preserve early-stage grounding while preventing attention collapse onto uninformative regions. Extensive experiments on eight benchmark datasets demonstrate that LASER consistently outperforms strong baselines, validating attention-aware training as an effective remedy for visual forgetting.

Problem

• visual forgetting
• attention drift
• vision-language models
• visual grounding
• attention collapse

Innovation

• visual forgetting
• attention preservation
• sink suppression
• vision-language models
• post-training alignment