作者位次 #4/6ACL 2026Poster· Natural Language Processing
Beyond Accuracy: Unveiling Inefficiency Patterns in Tool-Integrated Reasoning
Qisheng Su · Shiting Huang · Zhen Fang · Ziyan Chen · Zehui Chen · Feng Zhao
摘要
In real-world Tool-Integrated Reasoning (TIR) scenarios, a major source of inefficiency is that the toolcalls create pauses between LLM requests and cause KV-cache eviction. Also, the long, unfiltered response returned by external tools inflates the KV-cache, so each decode step spends more time...