Research map / Large language models / Retrieval and knowledge
Debating with More Persuasive LLMs Leads to More Truthful Answers
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.
Studies whether weaker (non-expert) LLM and human judges can identify correct answers from debates between stronger expert LLMs on QuALITY, comparing debate against consultancy and showing that more persuasive debaters yield more accurate judgments.
Problem
Scalable oversight: can non-expert judges (weaker models or humans) reliably determine the correct answer when stronger expert models hold information they lack, without ground-truth labels?
Approach
Two expert LLMs with full text access argue opposing answers in a multi-round debate with a quote-verification tool, and a non-expert judge who cannot see the text picks the answer; debater persuasiveness is optimised unsupervised via best-of-N and critique-and-refinement using a preference model, and persuasiveness is shown to correlate with judge accuracy in debate but not consultancy.
Key results
- Debate with GPT-4-Turbo experts lifts non-expert LLM judge accuracy to 75.6% versus 48% naive baseline and 92.5% expert judge.
- Non-expert human judges reach 88% accuracy with debate versus 60% naive and 78% with consultancy.
- Debater persuasiveness (Elo from best-of-N and critique-and-refinement) correlates with higher judge accuracy in debate, while consultant persuasiveness lowers judge accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use debate instead of a naive judge | LLM judges on QuALITY · QuALITY | accuracy: 48% -> 76% | Helped |
| Use debate instead of consultancy | LLM judges on QuALITY · QuALITY | accuracy: 54% -> 76% | Helped |
| Use debate instead of a naive judge | Human judges on QuALITY · QuALITY | accuracy: 60% -> 88% | Helped |
| Use debate instead of consultancy | Human judges on QuALITY · QuALITY | accuracy: 78% -> 88% | Helped |
| Increase debater persuasiveness with inference-time optimisation | Debate · QuALITY | judge accuracy | Helped |
| Increase consultant persuasiveness with optimisation | Consultancy · QuALITY | judge accuracy | Hurt |
| Add judge interaction during debate (ablation) | Debate with human judges · QuALITY | accuracy | No clear effect |
| Add judge-debater interaction during debate (ablation) | Debate with LLM judges · QuALITY (T_L) | accuracy | Hurt |
| Use only arguments instead of verified quotes and arguments (ablation) | Debate and consultancy · QuALITY | judge accuracy | Hurt |
| Use quotes only instead of quotes and arguments (ablation) | Debate and consultancy · QuALITY | judge accuracy | Helped |
| Increase debate rounds (ablation) | Debate with weaker LLM judges · QuALITY | judge accuracy | Hurt |
| Increase consultancy rounds (ablation) | Consultancy with LLM judges · QuALITY | judge accuracy | Helped |
Surprising findings
- Increasing consultant persuasiveness reduces judge accuracy because optimisation disproportionately boosts incorrect consultants.
- Interactive debate with human or LLM judges did not improve accuracy over static debate.
Limitations
- Information asymmetry is the only difference between expert and non-expert; stronger models may differ in reasoning.
- Results rely on verified quote tool; without grounded evidence, debaters could fabricate alternative narratives.
- LLM judges show strong positional and verbosity biases and are overconfident.
Experimental setup
Datasets: QuALITY human-to-human debate dataset (Michael et al.)
Benchmarks: QuALITY HARD subset
Metrics: judge accuracy Elo rating win rate calibration (Brier score) performance gap recovered
Baselines: consultancy naive judge expert judge interactive debate
Builds on: debate (Irving et al.) consultancy critique-and-refinement Elo rating
Closest papers by meaning
- Multi-LLM Debate: Framework, Principals, and Interventions (NeurIPS 2024 · similarity 0.91)
- Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.89)
- On scalable oversight with weak LLMs judging strong LLMs (NeurIPS 2024 · similarity 0.89)
- When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning (NeurIPS 2026 · similarity 0.88)
- Aligning Large Language Models for Faithful Integrity Against Opposing Argument (AAAI 2025 · similarity 0.87)
- Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models? (NeurIPS 2025 · similarity 0.87)
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.