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

2026-07-21 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision and Pattern Recognition, Image and Video Processing · Depth: Expert, quick

Summary

ChronoStitch is a training-free method designed for composing independently stored visual key-value (KV) memories, addressing a critical challenge in long-video question answering. Traditional approaches that cache video chunks independently fail to preserve global temporal order due to local rotary position zero prefilling, leading to incorrect temporal phase collisions. ChronoStitch resolves this by re-basing post-rotary keys onto a global three-axis multimodal RoPE coordinate system, which maintains time, height, and width structure. It also selectively recomputes a small fraction of high-deviation later-chunk visual tokens to bridge content gaps. Tested on Qwen2.5-VL-3B and the temporal split of TempCompass, ChronoStitch improves event-ordering accuracy and runs 3.3x faster than full joint re-prefilling.

Key takeaway

For Machine Learning Engineers developing long-video question answering systems, ChronoStitch offers a robust, training-free solution to a common problem. If you are struggling with preserving global temporal order or inefficient KV cache composition in vision-language models, consider implementing this method. It significantly improves event-ordering accuracy and runs 3.3x faster than full joint re-prefilling, directly addressing limitations of naive chunk concatenation.

Key insights

ChronoStitch enables training-free composition of visual KV memories for long-horizon temporal reasoning by re-basing keys and selective recomputation.

Principles

• Naive KV cache concatenation fails temporal order.
• Visual tokens need 3-axis RoPE coordinates.
• Selective recomputation addresses content gaps.

Method

Re-base stored post-rotary keys onto a global three-axis multimodal RoPE coordinate system, then selectively recompute high-deviation later-chunk visual tokens.

In practice

• Improve long-video QA accuracy.
• Accelerate KV cache composition 3.3x.
• Preserve global temporal order in video.

Topics

• ChronoStitch
• Visual KV Memory
• Temporal Reasoning
• Long-Video QA
• Rotary Position Embedding
• Vision-Language Models
• Computer Vision

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• ChronoStitch: Training-Free Composition of Visual KV Memories for Long-Horizon Temporal Reasoning — Composing visual KV caches for long videos requires both positional re-basing and selective content repair.
• TimeProVe: Propose, then Verify for Efficient Long Video Temporal Reasoning in Activities of Daily Living — TimeProVe efficiently reasons in long videos by proposing hypotheses with lightweight LLMs and verifying them with targeted VLM calls.
• Moment-KV: Momentum-Based Decode-Time KV Cache Compression for Long Generation — Moment-KV improves LLM long-generation fidelity by using momentum-driven attention aggregation for decoding-time KV cache compression.
• ChronoQG: Towards a Temporally Expressive and Hop-Bounded Benchmark for Temporal Knowledge Graph Question Generation — Existing KGQG methods fail to generate temporally faithful questions, necessitating ChronoQG, a new benchmark for TKGQG.
• Fractional Decay KV-Cache: Ownership-Aware Memory Management for Improved Inference Relevancy in Dialog Systems — FD-KVC uses dual-channel scoring and adaptive learning to manage KV-cache, improving dialog relevancy and topic adaptation.
• Visual Latents Know More Than They Say: Unsilencing Latent Reasoning in MLLMs — Visual latents in MLLMs can be "silenced" by training objectives, but their reasoning capacity can be unleashed at inference time.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.