# Honesty to Subterfuge: In-Context Reinforcement Learning Can Make Honest Models Reward Hack

arXiv: https://arxiv.org/abs/2410.06491 · HTML: https://mlsanity.com/paper/2410.06491/

## Key facts

- Cited by 8 papers in the map (data through 14 Oct 2026).
- Placed in Reward hacking and misspecification, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 18 of its 26 references that are in the map.
- First posted on arXiv 9 Oct 2024.

## Placed in

- LLM alignment / RLHF / preference optimisation / [Reward hacking and misspecification](https://mlsanity.com/alignment/reward-hacking-and-misspecification/)

## Builds on (most cited first, 18 of 18)

- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [Reflexion: Language Agents with Verbal Reinforcement Learning](https://mlsanity.com/paper/2303.11366/) (2023)
- [Self-Refine: Iterative Refinement with Self-Feedback](https://mlsanity.com/paper/2303.17651/) (2023)
- [The Effects of Reward Misspecification: Mapping and Mitigating Misaligned Models](https://mlsanity.com/paper/2201.03544/) (2022)
- [Principle-Driven Self-Alignment of Language Models from Scratch with Minimal Human Supervision](https://mlsanity.com/paper/2305.03047/) (2023)
- [The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning](https://mlsanity.com/paper/2312.01552/) (2023)
- [Can language models learn from explanations in context?](https://mlsanity.com/paper/2204.02329/) (2022)
- [Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models](https://mlsanity.com/paper/2406.10162/) (2024)
- [Training Language Models with Language Feedback at Scale](https://mlsanity.com/paper/2303.16755/) (2023)
- [ISR-LLM: Iterative Self-Refined Large Language Model for Long-Horizon Sequential Task Planning](https://mlsanity.com/paper/2308.13724/) (2023)
- [LLM2LLM: Boosting LLMs with Novel Iterative Data Enhancement](https://mlsanity.com/paper/2403.15042/) (2024)
- [Investigating the Effectiveness of Task-Agnostic Prefix Prompt for Instruction Following](https://mlsanity.com/paper/2302.14691/) (2023)
- [Generative AI for Synthetic Data Generation: Methods, Challenges and the Future](https://mlsanity.com/paper/2403.04190/) (2024)
- [Reflection-Tuning: Data Recycling Improves LLM Instruction-Tuning](https://mlsanity.com/paper/2310.11716/) (2023)
- [Spontaneous Reward Hacking in Iterative Self-Refinement](https://mlsanity.com/paper/2407.04549/) (2024)
- [Using Interactive Feedback to Improve the Accuracy and Explainability of Question Answering Systems Post-Deployment](https://mlsanity.com/paper/2204.03025/) (2022)
- [MetaReflection: Learning Instructions for Language Agents using Past Reflections](https://mlsanity.com/paper/2405.13009/) (2024)
- [Self-Judge: Selective Instruction Following with Alignment Self-Evaluation](https://mlsanity.com/paper/2409.00935/) (2024)

## Built on by (most cited first, 8 of 8)

- [A Comprehensive Survey of Reward Models: Taxonomy, Applications, Challenges, and Future](https://mlsanity.com/paper/2504.12328/) (2025)
- [Writing-Zero: Bridge the Gap Between Non-verifiable Tasks and Verifiable Rewards](https://mlsanity.com/paper/2506.00103/) (2025)
- [Open Rubric System: Scaling Reinforcement Learning with Pairwise Adaptive Rubric](https://mlsanity.com/paper/2602.14069/) (2026)
- [Autonomous Agents on Blockchains: Standards, Execution Models, and Trust Boundaries](https://mlsanity.com/paper/2601.04583/) (2026)
- [ScholarEval: Research Idea Evaluation Grounded in Literature](https://mlsanity.com/paper/2510.16234/) (2025)
- [The Secret Agenda: LLMs Strategically Lie and Our Current Safety Tools Are Blind](https://mlsanity.com/paper/2509.20393/) (2025)
- [Specification Self-Correction: Mitigating In-Context Reward Hacking Through Test-Time Refinement](https://mlsanity.com/paper/2507.18742/) (2025)
- [Bridging Online and Offline RL: Contextual Bandit Learning for Multi-Turn Code Generation](https://mlsanity.com/paper/2602.03806/) (2026)

## Abstract (arXiv)

Previous work has shown that training "helpful-only" LLMs with reinforcement learning on a curriculum of gameable environments can lead models to generalize to egregious specification gaming, such as editing their own reward function or modifying task checklists to appear more successful. We show that gpt-4o, gpt-4o-mini, o1-preview, and o1-mini - frontier models trained to be helpful, harmless, and honest - can engage in specification gaming without training on a curriculum of tasks, purely from in-context iterative reflection (which we call in-context reinforcement learning, "ICRL"). We also show that using ICRL to generate highly-rewarded outputs for expert iteration (compared to the standard expert iteration reinforcement learning algorithm) may increase gpt-4o-mini's propensity to learn specification-gaming policies, generalizing (in very rare cases) to the most egregious strategy where gpt-4o-mini edits its own reward function. Our results point toward the strong ability of in-context reflection to discover rare specification-gaming strategies that models might not exhibit zero-shot or with normal training, highlighting the need for caution when relying on alignment of LLMs in zero-shot settings.