ChronoStitch trains-free video chunk composition runs 3.3x faster

Long-video QA models face a trade-off: storing each chunk's KV cache avoids reprocessing the entire video, but naively concatenating cached chunks causes temporal phase collision because every chunk starts from local rotary position zero. This destroys global order needed for questions about event sequence, frequency, or change over time. ChronoStitch solves this with a training-free approach that first re-bases stored post-rotary keys onto a global three-axis multimodal RoPE coordinate system, preserving time, height, and width structure. The authors show that one-dimensional scalar re-indexing is geometrically inconsistent for visual tokens, as it turns spatial order within a frame into false temporal displacement. To address the content gap left by positional repair (later chunks were encoded without attending to earlier ones), ChronoStitch selectively recomputes a small fraction of high-deviation later-chunk visual tokens while allowing them to attend over the composed cache.

Evaluated on Qwen2.5-VL-3B and the temporal split of TempCompass, ChronoStitch outperforms naive composition and position-only variants. It achieves significant gains in event-ordering accuracy while running 3.3x faster than full joint re-prefilling, making it practical for real-time applications. The method requires no additional training or model modifications, enabling easy integration into existing vision-language models. This work addresses a critical bottleneck in long-horizon video understanding, paving the way for efficient temporal reasoning in domains like surveillance, sports analysis, and video search.