Kamera: Unified Position-Invariant Multimodal KV Cache for Training-Free Reuse

2026-06-22

AI Summary

Existing KV caching mechanisms incur substantial redundant computation in multimodal agents during sliding-window and iterative reasoning, as they repeatedly re-encode identical visual content due to reliance on fixed positional encodings. This work is the first to uncover the fundamental cause of performance degradation when applying position-agnostic KV reuse in multi-hop reasoning and introduces a training-free, position-agnostic multimodal KV caching method. By leveraging low-rank conditional patch recovery to reconstruct cross-block dependencies and integrating RoPE-based rotary repositioning for efficient reuse at arbitrary positions, the approach seamlessly supports cache block reordering, sliding-window retention, and retrieval. Compatible with MLA, GQA, and MHA architectures, it achieves near-bf16 reconstruction fidelity with minimal memory overhead, significantly restoring multi-hop accuracy—validated across six backbone models on MM-NIAH and two-page document question answering benchmarks.

Abstract

Multimodal agents repeatedly re-examine the same video frames, UI screenshots, and rendered artifacts as their context window slides and reasoning iterates, yet every look-back re-encodes from scratch, because prefix caches serve reuse only at a fixed leading position. We show this recompute is avoidable, and identify exactly what naive KV reuse loses: the cross-chunk conditioning a chunk absorbs from its neighbours. This loss is asymmetric. The direct readout of a cached chunk is recovered exactly and for free by the standard state-merge. What remains is a diffuse, low-rank residue concentrated in deep layers, invisible to single-hop retrieval but precisely what multi-hop reasoning binds on. Blind reuse therefore leaves single-hop recall intact while halving multi-hop accuracy; this is the failure mode prior position-independent caches, designed for single-context or single-image reuse, do not address. We repair it with a small, training-free low-rank conditioning patch stored alongside each position-free chunk. Reuse reduces to one operator across MLA, GQA, and MHA: exact RoPE re-rotation to any target position, plus the patch that restores cross-chunk binding. This makes three window operations cheap: reorder (one patch serves every ordering of a cached set), sliding-window survival (surviving chunks relocate via rotation only, zero re-encode), and recall (an evicted chunk is rehydrated by its patch, never re-encoded). A rank-m patch recovers full task accuracy on cross-chunk-binding benchmarks, MM-NIAH across two attention families and two-page doc-QA, at a fraction of the KV footprint, and reconstructs re-prefill KV to within bf16 rounding in a production SGLang kernel across six backbones. The conditioning signal is strongest in redundant vision and video streams, making our solution most impactful where multimodal agents spend their recompute budget.

Problem

• multimodal agents
• KV cache reuse
• position-invariant
• multi-hop reasoning
• cross-chunk conditioning

Innovation

• position-invariant KV cache
• training-free reuse
• cross-chunk conditioning
• low-rank patch
• multimodal reasoning