作者位次 #3/6ACL 2026Poster· Natural Language Processing
When KV Cache Reuse Fails in Multi-Agent Systems: Cross-Candidate Interaction is Crucial for LLM Judges
Sichu Liang · Zhenglin Wang · Chujiajia · Pengfei Xia · Hui Zang · Deyu Zhou
摘要
Multi-agent LLM systems routinely generate multiple candidate responses that are aggregated by an LLM judge. To reduce the dominant prefill cost in such pipelines, recent work advocates KV cache reuse across partially shared contexts and reports substantial speedups for generation agents. In this...