AI Paper Atlas

Research map / Large language models / Agents, code and math

AgentVocab: Structure-Aware Vocabulary Adaptation for Efficient LLM Agents

Kai Bian, Haosi Mo, Xuebo Liu, Shuangyong Song, Jing Li, Yongxiang Li, Min Zhang, Xuelong Li

ICML 2026 · Large language models › Agents, code and math · approach: Agent instruction-following benchmarks

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
adapt vocabulary with AgentVocab (800 structural + 200 content tokens)Vanilla SFT Qwen2.5-7B-Instruct · tau-bench overalllatency (s): 0.209 -> 0.172Improved
adapt vocabulary with AgentVocabVanilla SFT Qwen2.5-7B-Instruct · tau2-bench overalllatency (s): 0.375 -> 0.302Improved
adapt vocabulary with AgentVocabVanilla SFT Qwen2.5-7B-Instruct · tau-bench overallaccuracy: 22.42% -> 24.65%Improved
adapt vocabulary with AgentVocabVanilla SFT Qwen2.5-7B-Instruct · tau2-bench overallaccuracy: 23.79% -> 23.67%No clear effect
adapt vocabulary with AgentVocabVanilla SFT Qwen2.5-7B-Instruct · tau2-bench Telecomaccuracy: 24.56% -> 17.25%Hurt
add structural tokens only (800) (ablation)Vanilla SFT · tau2-bench overalllatency (s): 0.375 -> 0.312Improved
add content tokens only (800) (ablation)Vanilla SFT · tau2-bench overalllatency (s): 0.375 -> 0.360Mixed
increase structural budget from 800 to 1,000 (ablation)AgentVocab structural-only · tau2-bench overallaccuracy: 23.42% -> 23.17%No clear effect
reduce structural budget to 200 (ablation)AgentVocab structural-only · tau2-bench overallaccuracy: 23.42% -> 21.00%Hurt
add 200 content tokens to 800 structural (ablation)AgentVocab · tau2-bench overallaccuracy: 23.42% -> 23.67%Improved
increase content budget to 800 on top of 800 structural (ablation)AgentVocab · tau2-bench overallaccuracy: 23.67% -> 20.20%Hurt
fine-tune with SFT on tool tracesQwen2.5-7B-Instruct · tau-bench overallaccuracy: 13.94% -> 22.42%Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy