ChronoStitch: Training-Free Composition of Visual KV Memories for Long-Horizon Temporal Reasoning

2026-07-21

AI Summary

This work addresses the challenge of cross-segment temporal reasoning in long-form video question answering, where existing methods suffer from temporal phase conflicts in visual key-value (KV) caches due to local positional encodings. To overcome this limitation without requiring retraining, the authors propose an efficient approach that reconstructs the KV cache using a global triaxial multimodal RoPE coordinate system, preserving spatiotemporal structural consistency and avoiding spatial-temporal confusion caused by one-dimensional reindexing. Additionally, the method selectively recomputes high-bias visual tokens to compensate for missing contextual information. Notably, this is the first work to employ three-dimensional RoPE for cross-segment temporal alignment, achieving significant performance gains over baselines on Qwen2.5-VL-3B and TempCompass, with improved event ordering accuracy and 3.3× faster inference compared to full re-prefilling.

Abstract

Long-video question answering requires a model to preserve visual evidence over time without repeatedly reprocessing the same video. A practical approach is to store the vision-language model's internal key-value (KV) cache for each video chunk and retrieve that state at query time. However, independently cached video chunks do not compose correctly: every chunk is prefilled from local rotary position zero, so naive concatenation collides temporal phases and removes the global order required for questions about what happened first, how often events occurred, or what changed across the video. This paper presents ChronoStitch, a training-free method for composing independently stored visual KV memories. The method first re-bases stored post-rotary keys onto a global three-axis multimodal RoPE coordinate system that preserves time, height, and width structure. We show why a one-dimensional scalar re-indexing is geometrically inconsistent for visual tokens because it turns spatial order within a frame into false temporal displacement. We then address the residual content gap left by positional repair: later chunks were originally encoded without attending to earlier chunks. ChronoStitch therefore selectively recomputes a small fraction of high-deviation later-chunk visual tokens while allowing them to attend over the composed cache. On Qwen2.5-VL-3B and the temporal split of TempCompass, ChronoStitch outperforms naive composition and position-only variants, improving event-ordering accuracy while running 3.3x faster than full joint re-prefilling.

Problem

• long-video question answering
• temporal reasoning
• KV cache composition
• visual memory
• time-order preservation

Innovation

• ChronoStitch
• KV cache composition
• multimodal RoPE
• long-horizon temporal reasoning
• training-free adaptation