Research map / Large language models / Instruction and fine-tuning
Spilling the Beans: Teaching LLMs to Self-Report Their Hidden Objectives
Chloe Li, Mary Phuong, Daniel Tan
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 |
|---|---|---|---|
| add self-report fine-tuning (SRFT) to the model | GPT-4.1 interrogation · Hidden objective detection (4 stealth tasks) | F1 score: 0 -> 0.98 | Improved |
| add SRFT to the model | GPT-4.1 interrogation · Hidden objective elicitation | % details elicited: 0% -> 28-100% | Improved |
| train on turn-two admissions only, not on generating false answers on-policy (ablation) | SRFT · Hidden objective detection and elicitation | detection F1 and % details elicited: zero elicitation on all tasks | Hurt |
| add a system prompt telling the model to give false answers in turn-two-only training (ablation) | SRFT (turn2-only) · Hidden objective detection and elicitation | auditing performance | Improved |
| align admissions with ground truth rather than model belief on HLE hard questions (ablation) | SRFT · Hidden objective detection and elicitation | auditing performance | Hurt |
| align admissions with model belief rather than ground truth (ablation) | SRFT · Hidden objective detection and elicitation | auditing performance | Improved |
| add 10% random flipping of turn-two admissions to training data (ablation) | SRFT · Hidden objective detection and elicitation | auditing performance | Hurt |
| mix SRFT data 50/50 with general instruction tuning data (ablation) | SRFT · Hidden objective detection and elicitation | auditing performance: 50/50 token-balanced mixture | Improved |
| add a decoy objective to the agent during interrogation | SRFT · Hidden objective elicitation (4 tasks) | % details elicited: unchanged on 3 out of 4 tasks | No clear effect |
| add a decoy objective to the agent during interrogation | Prefilled response attack on Claude Sonnet 4 · Hidden objective elicitation | % details elicited: drops by 40-48% in 3 of 4 tasks | Hurt |
| apply SRFT to the model | GPT-4.1 · Instruction hierarchy compliance | system instruction compliance rate (%): 72 -> 61 | Hurt |
| apply SRFT to the model | GPT-4.1 · GPQA-diamond, AIME 2025, HLE | accuracy | No clear effect |
Closest papers by meaning
- Among Us: A Sandbox for Measuring and Detecting Agentic Deception (NeurIPS 2025 · similarity 0.87)
- OpenDeception: Learning Deception and Trust in Human–AI Interaction via Multi-Agent Simulation (ICML 2026 · similarity 0.86)
- How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions (ICLR 2024 · similarity 0.86)
- Preference Learning with Lie Detectors can Induce Honesty or Evasion (NeurIPS 2025 · similarity 0.86)
- Agent-ToM: Learning to Monitor Autonomous LLM Agents via Theory-of-Mind Reasoning (NeurIPS 2026 · similarity 0.86)
- Tell me about yourself: LLMs are aware of their learned behaviors (ICLR 2025 · similarity 0.85)
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.
