Research map / Multi-agent systems and games / LLM agents
AgentTailor: A Semantic-Aware LLM-Based Multi-Agent System with Actor-Critic Structure
Peiting Yang, Jiahao Shi, Caiyi Xu, Ming Liu, Yanxia Wu, Rongsheng Li
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AgentTailor, an LLM-based multi-agent system that uses semantic information from execution to optimize communication structure and reduce token cost, with an actor-critic structure.
Problem
LLM-based multi-agent systems incur high communication cost from redundant interactions, since existing methods optimize communication structures without measuring whether exchanged messages contribute to the final decision.
Approach
Uses semantic information observed during the execution stage to optimize multi-agent communication structure, measuring message contribution to the final decision, within an actor-critic framework to reduce token costs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add semantic edge judgment plus virtual execution (AgentTailor) over structural/temporal pruning | AgentDropout · 6-dataset average | accuracy: 90.69 -> 91.36 | Improved |
| use AgentTailor instead of AgentDropout | multi-agent system · total tokens across datasets | total tokens: -21.2% to -61.6% | Improved |
| AgentTailor vs AgentPrune-R on GSM8K | AgentTailor · GSM8K | accuracy: 95.72 -> 93.81 | Hurt |
| add virtual execution via EPN (ablation) | AgentTailor · HumanEval | pass@1: 87.33 -> 90.31 | Improved |
| add virtual execution via EPN (ablation) | AgentTailor · MMLU | accuracy: 84.24 -> 86.14 | Improved |
| add virtual execution via EPN (ablation) | AgentTailor · AQuA/MultiArith/SVAMP | accuracy: AQuA 87.21 -> 87.34 | No clear effect |
| add virtual execution via EPN (ablation) | AgentTailor · AQuA | total tokens: 920946.2 -> 724299.8 | Improved |
| add relative-contribution normalization of edge utility (ablation) | AgentTailor · AQuA | accuracy: 82.05 -> 87.34 | Improved |
| add relative-contribution normalization of edge utility (ablation) | AgentTailor · MMLU | accuracy: 84.82 -> 86.14 | Improved |
| prune high-contribution edges instead of low (reverse pruning) (ablation) | AgentTailor · AQuA | accuracy: 87.34 -> 83.20 | Hurt |
| remove one top-5 delta edge versus one bottom-5 delta edge (frozen graph) (ablation) | AgentTailor · LOEO held-out subset | accuracy: 79.0/81.0/81.5 vs 79.5/81.0/83.0 | No clear effect |
| combine AgentTailor with GPTSwarm | GPTSwarm · AQuA | accuracy: 78.16 -> 86.98 | Improved |
Closest papers by meaning
- Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems (ICLR 2025 · similarity 0.89)
- Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.88)
- AgentNoiseBench: Benchmarking Robustness of Tool-Using LLM Agents Under Noisy Condition (ICML 2026 · similarity 0.88)
- SafeSieve: From Heuristics to Experience in Progressive Pruning for LLM-based Multi-Agent Communication (AAAI 2026 · similarity 0.88)
- HieraMAS: Optimizing Intra-Node LLM Mixtures and Inter-Node Topology for Multi-Agent Systems (ICML 2026 · similarity 0.88)
- GPTSwarm: Language Agents as Optimizable Graphs (ICML 2024 · similarity 0.88)
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.
