Research map / Large language models / Language, safety and interpretability
Truthfulness Despite Weak Supervision: Evaluating and Training LLMs Using Peer Prediction
Tianyi Alex Qiu, Micah Carroll, Cameron Allen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with peer prediction DPO reward (0.135B expert) instead of LLM-as-a-Judge reward | Llama-3.1-8B deceptively finetuned · Free-form QA (37079 questions, 85 domains) | ground-truth accuracy | Improved |
| widen participant-to-expert size gap (smaller expert) (ablation) | peer prediction evaluation · Honest vs deceptive detection (Llama-3.1-8B, Gemma2-2B, Gemma2-27B participants) | cross-entropy of honesty regression | Improved |
| use LLM-as-a-Judge when judge is 5-20x smaller than deceptive participant | LLM-as-a-Judge · Honest vs deceptive detection | honesty prediction: worse than random | Hurt |
| increase number of participants (ablation) | peer prediction · Llama-3.1-8B participants, SmolLM-360M expert | honesty prediction | Improved |
| increase number of experts (ensemble) (ablation) | peer prediction · Honest vs deceptive detection | honesty regression R^2 surplus | Improved |
| ensemble weight exponent alpha deviating from -1 (ablation) | expert ensemble · Honest vs deceptive detection | surplus | Hurt |
| increase expert size from 0.5B to 7B (ablation) | peer prediction · Correctness prediction (pooled) | Pearson correlation: 0.1947 -> -0.1595 | Hurt |
| increase judge size from 135M to 7B (ablation) | LLM-as-a-Judge · Correctness prediction (pooled) | Pearson correlation: 0.0352 -> 0.2379 | Improved |
| make deceptive participants half or more of population (ablation) | peer prediction · 4 Llama-3.1-8B participants, Mistral-7B expert | honest vs deceptive scores | Hurt |
| use peer prediction with a small SmolLM-360M expert under majority deception (ablation) | peer prediction · 4 Llama-3.1-8B participants, SmolLM-360M expert | deception detection | Improved |
Closest papers by meaning
- Incentivizing Truthful Language Models via Peer Elicitation Games (NeurIPS 2025 · similarity 0.92)
- Evaluating LLM-contaminated Crowdsourcing Data Without Ground Truth (NeurIPS 2025 · similarity 0.88)
- LieCraft: A Multi-Agent Framework for Evaluating Deceptive Capabilities in Language Models (AAAI 2026 · similarity 0.88)
- Stochastically Dominant Peer Prediction (NeurIPS 2025 · similarity 0.87)
- Carrot and Stick: Eliciting Comparison Data and Beyond (NeurIPS 2024 · similarity 0.87)
- HonestLLM: Toward an Honest and Helpful Large Language Model (NeurIPS 2024 · 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.
