Position Rebinding Cache Reuse: Replay-Free Visual Revisiting for Interleaved Multimodal Reasoning

2026-06-25

AI Summary

Existing visual revisiting methods rely on repeated forward passes, while directly reusing historical key-value (KV) caches leads to attention distortion or even decoding collapse due to outdated positional bindings. This work is the first to reveal the critical impact of positional binding in KV caches on multimodal reasoning and proposes a replay-free cache reuse mechanism. By preserving the original visual spatial structure and dynamically rebinding cache entries to coordinates compatible with the current decoding position, the method enables efficient and accurate reuse of visual evidence. Evaluated across multiple multimodal benchmarks, it achieves an average accuracy improvement of 5% and reduces visual revisiting computational overhead by orders of magnitude—matching or surpassing the performance of conventional replay-based strategies.

Abstract

Interleaved multimodal reasoning improves visual grounding by revisiting visual evidence during multi-step generation, yet existing methods typically rely on token replay, repeatedly forwarding selected visual tokens. A natural shortcut is to reuse the historical visual key-value (KV) cache directly. However, we identify a critical failure mode of this strategy: cached visual keys are already bound to their original positional context. Such stale positional binding distorts attention under later decoding contexts and can trigger severe autoregressive decoding collapse. This failure suggests that effective cache reuse requires reconstructing visual evidence under positions compatible with the current decoding state, rather than directly copying position-bound historical cache entries. To this end, we propose Position Rebinding Cache Reuse (PRCR), a cache-level framework for replay-free visual revisiting. PRCR stores raw visual KV cache together with their original spatial coordinates, then reassigns position-compatible coordinates to select entries and rebinds their keys before injecting the reconstructed cache into the active decoder cache. This design reuses historical visual evidence while preserving textual positional continuity and relative visual structure. Experiments across multiple multimodal reasoning benchmarks show that PRCR achieves replay-level or better performance, improving average accuracy by 5 percent and reducing visual-revisiting computation by up to tens of thousands of times.

Problem

• visual revisiting
• KV cache reuse
• positional binding
• multimodal reasoning
• autoregressive decoding

Innovation

• Position Rebinding
• KV Cache Reuse
• Visual Revisiting
• Multimodal Reasoning
• Replay-Free