Research map / Multi-agent systems and games / LLM agents
AgentCDM: Enhancing Multi-Agent Collaborative Decision-Making via ACH-Inspired Structured Reasoning
Read the paperCheck your idea against 108,967 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 AgentCDM, a two-stage RL training framework (ACH-scaffolded GRPO followed by scaffold removal with soft ACH rewards) that trains a decision agent to aggregate multi-agent answers on MCQA benchmarks.
Problem
Collaborative decision-making in LLM-based multi-agent systems relies on either dictatorial single-agent decisions or shallow voting, both vulnerable to cognitive biases; the decision step is underexplored.
Approach
Injects an Analysis of Competing Hypotheses (ACH) structured reasoning protocol into a decision agent, trains it with rule-based RL (format, accuracy, ACH pattern rewards), then gradually removes the scaffold via soft semantic ACH rewards and cosine-annealed prompt sampling.
Key results
- AgentCDM reaches 65.2% average accuracy on MMLU-Pro versus 58.0% for Qwen-7B-R1 under the ACH protocol and 47.9% average for single agents.
- On the All Average metric across the three benchmarks, AgentCDM achieves 77.5 versus 74.5 for Qwen-7B-R1 (ACH) and 65.9 for single agents.
- Training on MMLU-Pro and testing on MMLU yields 80.8, above the 78.5 from training directly on MMLU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ACH-trained two-stage training (AgentCDM) to decision agent vs. single agent | Multiple base models (average over rows) · MMLU_PRO | average accuracy: +17.3 over single agent average (48.0 -> 65.2 reported as average gain) | Helped |
| Replace unstructured Informed Dictatorial decider with AgentCDM | Decision agent · All datasets | average accuracy: 77.5 vs 66.0 (+11.5) | Helped |
| Apply ACH protocol zero-shot (Homogeneous decider, no training) instead of voting | Base models as decider · All datasets | average accuracy: 67.9(+2.0) vs voting ~65.8-66.0 | Helped |
| Use voting-based aggregation (Plurality, Borda, IRV, Minimax, etc.) instead of single-agent answer | Multiple base models · MMLU | average accuracy: roughly -0.3 to +0.6 average change | No clear effect |
| Stage 2 removed (Scaffolding-Only) vs full two-stage training (ablation) | Qwen-7B-Base decider · MMLU_PRO | accuracy: 65.2 full vs 49.4 scaffolding-only (+15.8) | Helped |
| Stage 1 removed (Exploration-Only) vs full two-stage training (ablation) | Qwen-7B-Base decider · MMLU_PRO | accuracy: 65.2 full vs 52.1 exploration-only (+13.1) | Helped |
| Stage 1 removed (Exploration-Only) vs full two-stage training (ablation) | Qwen-7B-Base decider · MMLU | accuracy: 78.5 full vs 71.5 (+7.0) | Helped |
| Stage 2 removed (Scaffolding-Only) vs full two-stage training (ablation) | Qwen-7B-Base decider · MMLU | accuracy: 78.5 full vs 68.1 (+10.4) | Helped |
| Full two-stage training vs vanilla prompt (no training) (ablation) | Qwen-7B-Base decider · MMLU_PRO | accuracy: 65.2 vs 52.5 (+12.7) | Helped |
| Train on MMLU_PRO instead of MMLU or ARC-Challenge (cross-dataset transfer) | AgentCDM decider · MMLU and ARC-Challenge test sets | accuracy: MMLU 80.8 vs 78.5 (MMLU-trained); ARC 94.0 vs 89.5 (ARC-trained) | Helped |
| Increase number of execution agents on MMLU-PRO with weaker models (Mistral-7B) | Voting/decision with weak agents · MMLU_PRO | accuracy | Hurt |
| Increase number of execution agents on MMLU-PRO with capable models (GPT-4) | Decision with strong agents · MMLU_PRO | accuracy | Helped |
Surprising findings
- Directly removing the ACH protocol in the second stage causes training collapse, requiring a smoother transition.
- Training on the harder MMLU-Pro generalized better to MMLU and ARC-Challenge than training on those datasets directly.
Limitations
- Effectiveness depends on quality and diversity of execution-agent hypotheses.
- Designed only for cooperative agents; adversarial or noisy agents are not addressed.
Experimental setup
Datasets: MMLU MMLU-Pro ARC-Challenge
Benchmarks: MMLU MMLU-Pro ARC-Challenge
Metrics: accuracy
Baselines: Informed Dictatorial Plurality Borda Count Bucklin IRV Minimax Ranked Pairs Qwen-7B-Instruct (ACH) Qwen-7B-R1 (ACH) Homogeneous (ACH)
Builds on: DeepSeek-R1 GRPO GEDI
Closest papers by meaning
- Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs (ICML 2026 · similarity 0.90)
- Unlocking the Power of Multi-Agent LLM for Reasoning: From Lazy Agents to Deliberation (ICLR 2026 · similarity 0.89)
- Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.87)
- Agentick: A Unified Benchmark for General Sequential Decision-Making Agents (NeurIPS 2026 · similarity 0.87)
- From Debate to Equilibrium: Belief‑Driven Multi‑Agent LLM Reasoning via Bayesian Nash Equilibrium (ICML 2025 · similarity 0.87)
- MARSHAL: Incentivizing Multi-Agent Reasoning via Self-Play with Strategic LLMs (ICLR 2026 · similarity 0.86)
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.