Research map / Multi-agent systems and games / LLM agents

Strategic Planning and Rationalizing on Trees Make LLMs Better Debaters

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 0 citations (OpenAlex)

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 TreeDebater, an Oxford-style competitive debate system that plans with a Rehearsal Tree (anticipated attack/defense scoring), a Debate Flow Tree (debate state tracking), a speech time controller, and simulated audience feedback, outperforming Agent4Debate in human evaluation.

Problem

Competitive debate under strict time limits with no objective reward signal: LLM debaters struggle to choose which arguments to pursue and to plan across back-and-forth exchanges.

Approach

Model debate planning with two trees: a Rehearsal Tree that pre-computes tree-structured attacks and defenses and scores claim strength via minimax-like lookahead, and a Debate Flow Tree that tracks the live debate state to produce candidate actions; combined with a TTS-based speech time controller and retrieval-augmented simulated audience feedback.

Key results

• Stage-level head-to-head human evaluation: TreeDebater averages 3.69 vs 3.54 for Agent4Debate with Gemini and 4.01 vs 3.47 with DeepSeek (about +15.6% persuasiveness with DeepSeek).
• End-to-end Oxford-style debate: opinion shift win of 0.46 vs 0.13 for Agent4Debate with Gemini and 0.40 vs 0.30 with DeepSeek.
• Time validity of closing statements reaches 100% for TreeDebater versus 13.5% (Gemini) and 5.8% (DeepSeek) for Agent4Debate.

Surprising findings

• Agent4Debate with Gemini often outputs reviewer feedback as the final statement, failing format validity in 23% of debates.
• Audiences often judged a debate on personal belief in the motion rather than strategy when both sides performed reasonably well.

Limitations

• Human evaluation scale is small (13 motions for DeepSeek/Gemini end-to-end, 120 stage-level comparisons) and subjective.
• Stage-level annotator agreement is moderate at 60.7%.

Experimental setup

Datasets: Kialo PanelBench

Metrics: persuasiveness score win rate opinion shift win format and time validity

Baselines: Agent4Debate

Builds on: Agent4Debate Kialo FastSpeech

Closest papers by meaning

• Multi-LLM Debate: Framework, Principals, and Interventions (NeurIPS 2024 · similarity 0.89)
• When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning (NeurIPS 2026 · similarity 0.87)
• Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models? (NeurIPS 2025 · similarity 0.86)
• Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.86)
• Debating with More Persuasive LLMs Leads to More Truthful Answers (ICML 2024 · similarity 0.86)
• Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate (ICML 2026 · similarity 0.86)