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

Researchers introduce Kamera, a training-free method that enables efficient reuse of cached key-value pairs in multimodal AI models regardless of position in the context window. By storing small low-rank conditioning patches alongside position-free chunks, the system maintains accuracy for complex multi-hop reasoning tasks while reducing computational overhead—particularly benefiting video and vision-heavy applications.

Kamera addresses a fundamental inefficiency in how multimodal large language models handle repeated content examination. When AI agents process videos or UI screenshots multiple times as their reasoning evolves and context windows shift, traditional prefix caches force complete re-encoding because they only work at fixed positions. This represents significant computational waste in systems that repeatedly reference the same visual content. The innovation lies in identifying exactly what gets lost during naive KV reuse: cross-chunk conditioning signals that enable multi-hop reasoning across different parts of the context.

The breakthrough demonstrates that direct chunk readout recovers automatically through standard state-merge operations, but a diffuse low-rank residue remains in deeper layers—invisible to single-hop retrieval yet critical for complex reasoning chains. Rather than accepting this accuracy loss, Kamera repairs it with minimal overhead through training-free low-rank conditioning patches stored with each position-free chunk. This unified approach works across multiple attention mechanisms (MLA, GQA, MHA) and enables three previously expensive operations: reordering cached content, sliding-window survival through pure rotation, and chunk rehydration without re-encoding.

The practical impact emerges in production environments using frameworks like SGLang, where the method reconstructs re-prefill KV values to within bf16 floating-point precision. Multimodal agents running vision-heavy or video-intensive workloads benefit most, as redundancy in these modalities amplifies the recompute savings. For developers building long-context multimodal systems, this represents an immediate efficiency gain—reducing memory footprint and latency without architectural changes or model retraining. The solution scales across different backbone models, suggesting broad adoption potential in production AI systems handling complex multi-modal reasoning.

• →Kamera enables position-invariant KV cache reuse using small low-rank conditioning patches, eliminating expensive re-encoding of repeated visual content.
• →The method maintains full accuracy on multi-hop reasoning tasks while halving memory footprint through training-free conditioning signals stored in deeper model layers.
• →Three window operations become computationally cheap: content reordering, sliding-window survival via RoPE rotation, and chunk rehydration without re-encoding.
• →Production testing across six backbone models shows reconstructed KV values match original precision within bf16 rounding error margins.
• →Vision and video streams show strongest gains, making the solution most impactful for multimodal agents processing repeated visual content in long contexts.