Research map / Multi-agent systems and games / LLM agents
Analyze–Compose–Execute: A Dynamic Dialogue Framework for Multi-Agent Debate
Wenyuan Gu, Haowen Wang, Jiale Han, Xiang Li, Zhixuan Wu, Hongru Xiao, Bo Cheng
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 a dynamic dialogue framework for multi-agent debate that adapts discussion acts across rounds instead of following fixed pipelines.
Problem
Current multi-agent debate (MAD) methods constrain LLM agents to fixed discussion pipelines that repeat the same discussion act for a predetermined number of rounds, limiting effectiveness and adaptivity.
Approach
Replaces fixed MAD pipelines with a dynamic dialogue framework (Analyze, Compose, Execute) in which agents choose and sequence discussion acts adaptively; details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add dynamic Analyze-Compose-Execute debate instead of fixed pipelines | MAD with GPT-3.5-turbo · BBH 23 tasks avg | accuracy (%): 52.5 -> 61.0 (best baseline RTC vs ACE) | Improved |
| add Analyze module (ablation) | ACE · BBH Penguins in a Table | accuracy (%): 60.3 -> 70.1 | Improved |
| add Analyze module (ablation) | ACE · BBH Tracking Shuffled Objects | accuracy (%): 23.1 -> 37.1 | Improved |
| add guided Compose instead of random act selection (ablation) | ACE · BBH 4 tasks (TSO, DU, PIA, GS) | accuracy (%): 49.7 -> 70.1 on PIA | Improved |
| add separate Execute module (ablation) | ACE · BBH 4 tasks (TSO, DU, PIA, GS) | accuracy (%): 67.4 -> 71.0 on DU | Improved |
| change number of discussion rounds (ablation) | PAR, RPD, RTC, TAR baselines · BBH 5 tasks | accuracy (%) | Mixed |
| use GPT-4 instead of GPT-3.5-turbo backbone | ACE · BBH Tracking Shuffled Objects | accuracy (%): 37.1 -> 54.2 | Improved |
| apply ACE on Llama-2-13B vs baselines | ACE · BBH Penguins in a Table | accuracy (%): 58.4 -> 69.5 (RTC vs ACE) | Improved |
| apply ACE on commonsense recall tasks | ACE · BBH Hyperbaton / Movie Recommendation | accuracy (%): 77.6 vs 77.0 on Hyperbaton (TAR vs ACE) | No clear effect |
| apply ACE on supplementary datasets | ACE · MMLU, StrategyQA, Common MT | accuracy / COMET | Improved |
Experimental setup
Builds on: multi-agent debate
Closest papers by meaning
- Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.91)
- Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate (ICML 2026 · similarity 0.89)
- Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models? (NeurIPS 2025 · similarity 0.89)
- Multi-LLM Debate: Framework, Principals, and Interventions (NeurIPS 2024 · similarity 0.89)
- Key Decision-Makers in Multi-Agent Debates: Who Holds the Power? (AAAI 2026 · similarity 0.88)
- iMAD: Intelligent Multi-Agent Debate for Efficient and Accurate LLM Inference (AAAI 2026 · 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.
