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

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, long

Summary

ChronoStitch is a training-free method designed to compose independently stored visual Key-Value (KV) memories for long-horizon temporal reasoning in vision-language models. The approach addresses two key issues: positional errors from overlapping local multimodal Rotary Position Embedding (mRoPE) phases and content gaps where later video chunks were encoded without attending to earlier ones. ChronoStitch first re-bases stored post-rotary keys onto a global three-axis mRoPE coordinate system, preserving time, height, and width structure. Second, it selectively recomputes a small fraction (e.g., 35%) of high-deviation later-chunk visual tokens, allowing them to attend over the composed cache. Evaluated on Qwen2.5-VL-3B and the temporal split of TempCompass, ChronoStitch achieved 61.5% overall accuracy and 65.8% event ordering accuracy, outperforming naive composition and position-only variants. It also demonstrated a 3.3x speedup compared to full joint re-prefilling, making it efficient for query-time video memory.

Key takeaway

For Machine Learning Engineers building long-video question answering systems, you should consider ChronoStitch to efficiently extend temporal reasoning capabilities. This method allows you to reuse independently cached visual KV memories, avoiding costly full re-prefilling while significantly improving event-ordering accuracy. Implement its two-stage process of three-axis key re-basing and selective token recomputation to achieve a 3.3x query-time speedup and enhance temporal coherence in your models.

Key insights

Composing visual KV caches for long videos requires both positional re-basing and selective content repair.

Principles

• Scalar re-indexing is geometrically inconsistent for visual tokens.
• Positions are necessary but not sufficient for temporal reasoning.
• Selective recomputation repairs cross-chunk attention gaps efficiently.

Method

ChronoStitch re-bases stored post-rotary keys to global three-axis mRoPE coordinates, then selectively recomputes high-deviation later-chunk visual tokens against the composed cache.

In practice

• Implement three-axis mRoPE re-basing for visual token positional correction.
• Use key deviation as a proxy for identifying tokens needing recomputation.
• Recompute approximately 25-35% of later-chunk visual tokens for temporal coherence.

Topics

• Video Understanding
• KV Cache
• Vision-Language Models
• Rotary Position Embedding
• Temporal Reasoning
• Long-Context Inference

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

Related on AIssential

• ChronoStitch: Training-Free Composition of Visual KV Memories for Long-Horizon Temporal Reasoning — ChronoStitch enables training-free composition of visual KV memories for long-horizon temporal reasoning by re-basing keys and selective recomputation.
• 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.
• Reroute, Don't Remove: Recoverable Visual Token Routing for Vision-Language Models — Visual token reduction in VLMs should prioritize recoverable routing over irreversible pruning due to dynamic token importance.
• REAL: REtrieval-reAsoning and Logic-constructed Attention Behaviors for Long-Context KV Cache Compression — LLM KV cache compression can be robustly optimized by distinguishing and allocating resources based on "surface memorization" and "logic construction" attention behaviors.
• 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.
• Seeing Time: Benchmarking Chronological Reasoning and Shortcut Biases in Vision-Language Models — VLMs often use superficial visual cues like color filters as shortcuts, failing genuine chronological reasoning.

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 cs.CV updates on arXiv.org.