# Teaching Models to Verbalize Reward Hacking in Chain-of-Thought Reasoning

arXiv: https://arxiv.org/abs/2506.22777 · HTML: https://mlsanity.com/paper/2506.22777/

## Key facts

- Cited by 21 papers in the map (data through 14 Oct 2026).
- Placed in Reinforcement learning for reasoning, one of 17 directions in LLM reasoning.
- Builds on 29 of its 37 references that are in the map.
- First posted on arXiv 28 Jun 2025.

## Placed in

- LLM reasoning / [Reinforcement learning for reasoning](https://mlsanity.com/lr/reinforcement-learning-for-reasoning/)
- LLM alignment / RLHF / preference optimisation / [Reward hacking and misspecification](https://mlsanity.com/alignment/reward-hacking-and-misspecification/)

## Builds on (most cited first, 29 of 29)

- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [Chain-of-Thought Prompting Elicits Reasoning in Large Language Models](https://mlsanity.com/paper/2201.11903/) (2022)
- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [Towards A Rigorous Science of Interpretable Machine Learning](https://mlsanity.com/paper/1702.08608/) (2017)
- [Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting](https://mlsanity.com/paper/2305.04388/) (2023)
- [Measuring Faithfulness in Chain-of-Thought Reasoning](https://mlsanity.com/paper/2307.13702/) (2023)
- [Towards Faithfully Interpretable NLP Systems: How should we define and evaluate faithfulness?](https://mlsanity.com/paper/2004.03685/) (2020)
- [Reasoning Models Don't Always Say What They Think](https://mlsanity.com/paper/2505.05410/) (2025)
- [Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation](https://mlsanity.com/paper/2503.11926/) (2025)
- [Alignment faking in large language models](https://mlsanity.com/paper/2412.14093/) (2024)
- [Faithful Chain-of-Thought Reasoning](https://mlsanity.com/paper/2301.13379/) (2023)
- [Frontier Models are Capable of In-context Scheming](https://mlsanity.com/paper/2412.04984/) (2024)
- [Measuring Progress on Scalable Oversight for Large Language Models](https://mlsanity.com/paper/2211.03540/) (2022)
- [Evaluating Explainable AI: Which Algorithmic Explanations Help Users Predict Model Behavior?](https://mlsanity.com/paper/2005.01831/) (2020)
- [Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models](https://mlsanity.com/paper/2406.10162/) (2024)
- [Chain-of-Thought Reasoning In The Wild Is Not Always Faithful](https://mlsanity.com/paper/2503.08679/) (2025)
- [Thought Anchors: Which LLM Reasoning Steps Matter?](https://mlsanity.com/paper/2506.19143/) (2025)
- [Question Decomposition Improves the Faithfulness of Model-Generated Reasoning](https://mlsanity.com/paper/2307.11768/) (2023)
- [Do Models Explain Themselves? Counterfactual Simulatability of Natural Language Explanations](https://mlsanity.com/paper/2307.08678/) (2023)
- [Prover-Verifier Games improve legibility of LLM outputs](https://mlsanity.com/paper/2407.13692/) (2024)
- [Are DeepSeek R1 And Other Reasoning Models More Faithful?](https://mlsanity.com/paper/2501.08156/) (2025)
- [Preventing Language Models From Hiding Their Reasoning](https://mlsanity.com/paper/2310.18512/) (2023)
- [Thought Crime: Backdoors and Emergent Misalignment in Reasoning Models](https://mlsanity.com/paper/2506.13206/) (2025)
- [CoT Red-Handed: Stress Testing Chain-of-Thought Monitoring](https://mlsanity.com/paper/2505.23575/) (2025)
- [Bias-Augmented Consistency Training Reduces Biased Reasoning in Chain-of-Thought](https://mlsanity.com/paper/2403.05518/) (2024)
- [Large language models can learn and generalize steganographic chain-of-thought under process supervision](https://mlsanity.com/paper/2506.01926/) (2025)
- [Towards Consistent Natural-Language Explanations via Explanation-Consistency Finetuning](https://mlsanity.com/paper/2401.13986/) (2024)
- [Deductive Closure Training of Language Models for Coherence, Accuracy, and Updatability](https://mlsanity.com/paper/2401.08574/) (2024)
- [MONA: Myopic Optimization with Non-myopic Approval Can Mitigate Multi-step Reward Hacking](https://mlsanity.com/paper/2501.13011/) (2025)

## Built on by (most cited first, 21 of 21)

- [Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges](https://mlsanity.com/paper/2604.13602/) (2026)
- [Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use](https://mlsanity.com/paper/2605.02964/) (2026)
- [Is It Thinking or Cheating? Detecting Implicit Reward Hacking by Measuring Reasoning Effort](https://mlsanity.com/paper/2510.01367/) (2025)
- [Measuring Chain-of-Thought Monitorability Through Faithfulness and Verbosity](https://mlsanity.com/paper/2510.27378/) (2025)
- [Counterfactual Simulation Training for Chain-of-Thought Faithfulness](https://mlsanity.com/paper/2602.20710/) (2026)
- [Spilling the Beans: Teaching LLMs to Self-Report Their Hidden Objectives](https://mlsanity.com/paper/2511.06626/) (2025)
- [MONICA: Real-Time Monitoring and Calibration of Chain-of-Thought Sycophancy in Large Reasoning Models](https://mlsanity.com/paper/2511.06419/) (2025)
- [Investigating CoT Monitorability in Large Reasoning Models](https://mlsanity.com/paper/2511.08525/) (2025)
- [Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failures](https://mlsanity.com/paper/2601.19928/) (2026)
- [Detecting and Suppressing Reward Hacking with Gradient Fingerprints](https://mlsanity.com/paper/2604.16242/) (2026)
- [AI Must not be Fully Autonomous](https://mlsanity.com/paper/2507.23330/) (2025)
- [Faithfulness as Information Flow: Evaluating and Training Faithful Chain-of-Thought Reasoning](https://mlsanity.com/paper/2605.24286/) (2026)
- [Chain-of-Thought Monitoring Can Be Unreliable in Implicit-Influence Settings](https://mlsanity.com/paper/2608.04735/) (2026)
- [Would this change your answer? Evaluating Explanations of LLM Behavior In The Wild with Counterfactual Experiments](https://mlsanity.com/paper/2608.16747/) (2026)
- [The Horcrux: Mechanistically Interpretable Task Decomposition for Detecting and Mitigating Reward Hacking in Embodied AI Systems](https://mlsanity.com/paper/2511.17869/) (2025)
- [Finding RELIEF: Shaping Reasoning Behavior without Reasoning Supervision via Belief Engineering](https://mlsanity.com/paper/2601.13752/) (2026)
- [Directional Alignment Mitigates Reward Hacking in Reinforcement Learning for Language Models](https://mlsanity.com/paper/2605.25189/) (2026)
- [Process-Constituted Intelligence: A Shared Criterion for Humans and Machines](https://mlsanity.com/paper/2608.16213/) (2026)
- [A Removal Based Approach to Improve LLM Faithfulness at Test-Time](https://mlsanity.com/paper/2609.04343/) (2026)
- [Training LLMs to Verbalize Evaluation Awareness](https://mlsanity.com/paper/2609.36316/) (2026)
- [Reward Hacking as a Transmissible Strategy in Multi-Agent Coding Systems](https://mlsanity.com/paper/2610.12822/) (2026)

## Abstract (arXiv)

Language models trained with reinforcement learning (RL) can engage in reward hacking--the exploitation of unintended strategies for high reward--without revealing this behavior in their chain-of-thought reasoning. This makes the detection of reward hacking difficult, posing risks for high-stakes applications. We propose verbalization fine-tuning (VFT), a pre-RL fine-tuning intervention that trains models to explicitly acknowledge when they are influenced by prompt cues--hints which point to incorrect answers (e.g., "a Stanford professor thinks the answer is A"). To evaluate VFT, we subsequently train models with RL on environments where held-out prompt cues signal which incorrect answers will receive high reward, incentivizing models to exploit these cues instead of reasoning correctly. We measure how often models exploit these cues without verbalizing it. After RL, only 6% of the VFT-trained model's responses consist of undetected reward hacks. In comparison, when we perform RL without VFT, the rate of undetected reward hacks goes up to 88%; with a debiasing baseline intervention, this increases further to 99%. VFT achieves this by substantially increasing how often models verbalize the influence of cues, from 8% to 43% after VFT, and up to 94% after RL. Baselines remain low even after RL (11% and 1%). Our results show that teaching models to explicitly verbalize reward hacking behavior before RL significantly improves their detection, offering a practical path toward more transparent and safe AI systems.