Unsupervised Elicitation of Language Models

arxiv:2506.10139

Original source · Paper JSON

Post-training · 2025-06-11

Cites (5) · Cited by (2)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3

Contents

• Unsupervised Elicitation of Language Models
• Abstract
• 1 Introduction
• 2 Methodology
• 2.1 Problem Statement
• 2.2 Scoring Function
• 2.3 Our Algorithm
• 3 Experiment Setup
• 3.1 Datasets
• 3.2 Baselines
• 4 Experiments
• 4.1 Eliciting Capabilities on Common NLP Tasks
• 4.2 Unsupervised Elicitation Fails when Concepts are not Salient
• 4.3 Eliciting Superhuman Capabilities
• 4.4 Training an Assistant Chatbot without Supervision
• 5 Ablations
• 6 Discussion
• Ethics Statement
• Appendix
• Appendix A The Use of LLMs
• Appendix B Related Work
• Appendix C Illustrative Example of Scoring Function
• Appendix D Additional Implementation Details
• D.1 Hyperparameters
• D.2 Data Statistics
• Appendix E Additional Baselines
• E.1 Distillation
• E.2 CCS
• Appendix F Compute Costs
• Appendix G Scalable ICM
• Appendix H Evaluation Results of Claude 4 Sonnet Reward Models
• Appendix I Human Annotation
• Appendix J Discussion: Data Contamination
• Instructions for reporting errors