REVIEW 7 cited by
Debate Helps Supervise Unreliable Experts
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
As AI systems are used to answer more difficult questions and potentially help create new knowledge, judging the truthfulness of their outputs becomes more difficult and more important. How can we supervise unreliable experts, which have access to the truth but may not accurately report it, to give answers that are systematically true and don't just superficially seem true, when the supervisor can't tell the difference between the two on their own? In this work, we show that debate between two unreliable experts can help a non-expert judge more reliably identify the truth. We collect a dataset of human-written debates on hard reading comprehension questions where the judge has not read the source passage, only ever seeing expert arguments and short quotes selectively revealed by 'expert' debaters who have access to the passage. In our debates, one expert argues for the correct answer, and the other for an incorrect answer. Comparing debate to a baseline we call consultancy, where a single expert argues for only one answer which is correct half of the time, we find that debate performs significantly better, with 84% judge accuracy compared to consultancy's 74%. Debates are also more efficient, being 68% of the length of consultancies. By comparing human to AI debaters, we find evidence that with more skilled (in this case, human) debaters, the performance of debate goes up but the performance of consultancy goes down. Our error analysis also supports this trend, with 46% of errors in human debate attributable to mistakes by the honest debater (which should go away with increased skill); whereas 52% of errors in human consultancy are due to debaters obfuscating the relevant evidence from the judge (which should become worse with increased skill). Overall, these results show that debate is a promising approach for supervising increasingly capable but potentially unreliable AI systems.
Forward citations
Cited by 7 Pith papers
-
Human-AI Complementarity: A Goal for Amplified Oversight
Confidence-based routing of fact-verification to humans, plus evidence-only AI assistance, beats either human or AI raters alone: 91.3% hybrid accuracy vs 87.7% for the AI rater.
-
When to Trust Context: Self-Reflective Debates for Context Reliability
SR-DCR uses an asymmetric debate plus self-confidence to gate whether a model follows context or its prior, improving ClashEval accuracy on several models.
-
Helpful Agent Meets Deceptive Judge: Understanding Vulnerabilities in Agentic Workflows
LLM agents frequently switch correct answers after one round of misleading feedback, and the new WAFER-QA benchmark measures this with web-backed critiques.
-
CortexDebate: Debating Sparsely and Equally for Multi-Agent Debate
CortexDebate prunes the multi-agent debate graph every round using a McKinsey-style trust score per directed link, reporting accuracy gains over full-debate baselines on eight datasets with shorter per-agent contexts.
-
AUTOLAW: Enhancing Legal Compliance in Large Language Models via Case Law Generation and Jury-Inspired Deliberation
AutoLaw's verifier-ranked legal-role jury with a similar-case demonstration beats majority voting for violation detection on three law and policy benchmarks.
-
Out of Control -- Why Alignment Needs Formal Control Theory (and an Alignment Control Stack)
A position paper proposing that AI alignment adopt formal optimal control and a ten-layer Alignment Control Stack for organizing and interoperating control interventions.
-
Tournament of Prompts: Evolving LLM Instructions Through Structured Debates and Elo Ratings
DEEVO evolves better LLM prompts by debating outputs and selecting survivors with Elo ratings, without requiring labeled data or a hand-written fitness function.
Discussion (0). Continue with ORCID to comment.