Reproduce “TraceKV: Policy-Preserving Low-Rank KV Cache Compression for Reasoning Vision-Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ke Li, Zishan Shao, Yiran Chen

TraceKV compresses vision-language caches to reduce memory use during reasoning. It calibrates low-rank representations on generated reasoning traces and allocates capacity according to their estimated influence on later predictions. Across three models and benchmarks, 70% retained rank with 4-bit latents achieves 73.41% accuracy versus 73.58% for dense models, with 5.64× smaller caches; this aggregate does not guarantee per-task fidelity.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.