# Language Models Learn to Mislead Humans via RLHF

arXiv: https://arxiv.org/abs/2409.12822 · HTML: https://mlsanity.com/paper/2409.12822/

## Key facts

- Cited by 124 papers in the map (data through 14 Oct 2026).
- Placed in Process reward models for reasoning, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 25 of its 35 references that are in the map.
- First posted on arXiv 19 Sep 2024.

## Placed in

- LLM alignment / RLHF / preference optimisation / [Process reward models for reasoning](https://mlsanity.com/alignment/process-reward-models-for-reasoning/)

## Builds on (most cited first, 25 of 25)

- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [Proximal Policy Optimization Algorithms](https://mlsanity.com/paper/1707.06347/) (2017)
- [Llama 2: Open Foundation and Fine-Tuned Chat Models](https://mlsanity.com/paper/2307.09288/) (2023)
- [Deep reinforcement learning from human preferences](https://mlsanity.com/paper/1706.03741/) (2017)
- [DeepSeek-Coder: When the Large Language Model Meets Programming -- The Rise of Code Intelligence](https://mlsanity.com/paper/2401.14196/) (2024)
- [Towards Understanding Sycophancy in Language Models](https://mlsanity.com/paper/2310.13548/) (2023)
- [Measuring Coding Challenge Competence With APPS](https://mlsanity.com/paper/2105.09938/) (2021)
- [Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback](https://mlsanity.com/paper/2307.15217/) (2023)
- [Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training](https://mlsanity.com/paper/2401.05566/) (2024)
- [Fine-Grained Human Feedback Gives Better Rewards for Language Model Training](https://mlsanity.com/paper/2306.01693/) (2023)
- [Self-critiquing models for assisting human evaluators](https://mlsanity.com/paper/2206.05802/) (2022)
- [A Deep Reinforced Model for Abstractive Summarization](https://mlsanity.com/paper/1705.04304/) (2017)
- [A Long Way to Go: Investigating Length Correlations in RLHF](https://mlsanity.com/paper/2310.03716/) (2023)
- [Debating with More Persuasive LLMs Leads to More Truthful Answers](https://mlsanity.com/paper/2402.06782/) (2024)
- [The False Promise of Imitating Proprietary LLMs](https://mlsanity.com/paper/2305.15717/) (2023)
- [An Overview of Catastrophic AI Risks](https://mlsanity.com/paper/2306.12001/) (2023)
- [QuALITY: Question Answering with Long Input Texts, Yes!](https://mlsanity.com/paper/2112.08608/) (2021)
- [Do Users Write More Insecure Code with AI Assistants?](https://mlsanity.com/paper/2211.03622/) (2022)
- [Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models](https://mlsanity.com/paper/2406.10162/) (2024)
- [LLM Critics Help Catch LLM Bugs](https://mlsanity.com/paper/2407.00215/) (2024)
- [Human Feedback is not Gold Standard](https://mlsanity.com/paper/2309.16349/) (2023)
- [Debate Helps Supervise Unreliable Experts](https://mlsanity.com/paper/2311.08702/) (2023)
- [Two-Turn Debate Doesn't Help Humans Answer Hard Reading Comprehension Questions](https://mlsanity.com/paper/2210.10860/) (2022)
- [Learning Task Decomposition to Assist Humans in Competitive Programming](https://mlsanity.com/paper/2406.04604/) (2024)
- [Non-Programmers Can Label Programs Indirectly via Active Examples: A Case Study with Text-to-SQL](https://mlsanity.com/paper/2205.12422/) (2022)

## Built on by (most cited first, 30 of 124)

- [VLM-R1: A Stable and Generalizable R1-style Large Vision-Language Model](https://mlsanity.com/paper/2504.07615/) (2025)
- [Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation](https://mlsanity.com/paper/2503.11926/) (2025)
- [The Landscape of Agentic Reinforcement Learning for LLMs: A Survey](https://mlsanity.com/paper/2509.02547/) (2025)
- [ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models](https://mlsanity.com/paper/2505.24864/) (2025)
- [Large Language Models Often Know When They Are Being Evaluated](https://mlsanity.com/paper/2505.23836/) (2025)
- [Sycophancy in Large Language Models: Causes and Mitigations](https://mlsanity.com/paper/2411.15287/) (2024)
- [Reward Shaping to Mitigate Reward Hacking in RLHF](https://mlsanity.com/paper/2502.18770/) (2025)
- [A Comprehensive Survey of Reward Models: Taxonomy, Applications, Challenges, and Future](https://mlsanity.com/paper/2504.12328/) (2025)
- [Which of These Best Describes Multiple Choice Evaluation with LLMs? A) Forced B) Flawed C) Fixable D) All of the Above](https://mlsanity.com/paper/2502.14127/) (2025)
- [Scaling of Search and Learning: A Roadmap to Reproduce o1 from Reinforcement Learning Perspective](https://mlsanity.com/paper/2412.14135/) (2024)
- [Sample, Scrutinize and Scale: Effective Inference-Time Search by Scaling Verification](https://mlsanity.com/paper/2502.01839/) (2025)
- [On Targeted Manipulation and Deception when Optimizing LLMs for User Feedback](https://mlsanity.com/paper/2411.02306/) (2024)
- [Adaptive Deployment of Untrusted LLMs Reduces Distributed Threats](https://mlsanity.com/paper/2411.17693/) (2024)
- [Nemotron-CrossThink: Scaling Self-Learning beyond Math Reasoning](https://mlsanity.com/paper/2504.13941/) (2025)
- [Mitigating Deceptive Alignment via Self-Monitoring](https://mlsanity.com/paper/2505.18807/) (2025)
- [Step-wise Adaptive Integration of Supervised Fine-tuning and Reinforcement Learning for Task-Specific LLMs](https://mlsanity.com/paper/2505.13026/) (2025)
- [Pairwise or Pointwise? Evaluating Feedback Protocols for Bias in LLM-Based Evaluation](https://mlsanity.com/paper/2504.14716/) (2025)
- [Sycophancy Hides Linearly in the Attention Heads](https://mlsanity.com/paper/2601.16644/) (2026)
- [Institutional AI: Governing LLM Collusion in Multi-Agent Cournot Markets via Public Governance Graphs](https://mlsanity.com/paper/2601.11369/) (2026)
- [MONA: Myopic Optimization with Non-myopic Approval Can Mitigate Multi-step Reward Hacking](https://mlsanity.com/paper/2501.13011/) (2025)
- [Towards Hierarchical Multi-Step Reward Models for Enhanced Reasoning in Large Language Models](https://mlsanity.com/paper/2503.13551/) (2025)
- [Training Language Models to Win Debates with Self-Play Improves Judge Accuracy](https://mlsanity.com/paper/2409.16636/) (2024)
- [Search, Verify and Feedback: Towards Next Generation Post-training Paradigm of Foundation Models via Verifier Engineering](https://mlsanity.com/paper/2411.11504/) (2024)
- [An alignment safety case sketch based on debate](https://mlsanity.com/paper/2505.03989/) (2025)
- [Lies, Damned Lies, and Distributional Language Statistics: Persuasion and Deception with Large Language Models](https://mlsanity.com/paper/2412.17128/) (2024)
- [RLHS: Mitigating Misalignment in RLHF with Hindsight Simulation](https://mlsanity.com/paper/2501.08617/) (2025)
- [Technological folie `a deux: Feedback Loops Between AI Chatbots and Mental Illness](https://mlsanity.com/paper/2507.19218/) (2025)
- [Recontextualization Mitigates Specification Gaming without Modifying the Specification](https://mlsanity.com/paper/2512.19027/) (2025)
- [Reverse Question Answering: Can an LLM Write a Question so Hard (or Bad) that it Can't Answer?](https://mlsanity.com/paper/2410.15512/) (2024)
- [Persuasion Dynamics in LLMs: Investigating Robustness and Adaptability in Knowledge and Safety with DuET-PD](https://mlsanity.com/paper/2508.17450/) (2025)

## Abstract (arXiv)

Language models (LMs) can produce errors that are hard to detect for humans, especially when the task is complex. RLHF, the most popular post-training method, may exacerbate this problem: to achieve higher rewards, LMs might get better at convincing humans that they are right even when they are wrong. We study this phenomenon under a standard RLHF pipeline, calling it "U-SOPHISTRY" since it is Unintended by model developers. Specifically, we ask time-constrained (e.g., 3-10 minutes) human subjects to evaluate the correctness of model outputs and calculate humans' accuracy against gold labels. On a question-answering task (QuALITY) and programming task (APPS), RLHF makes LMs better at convincing our subjects but not at completing the task correctly. RLHF also makes the model harder to evaluate: our subjects' false positive rate increases by 24.1% on QuALITY and 18.3% on APPS. Finally, we show that probing, a state-of-the-art approach for detecting Intended Sophistry (e.g. backdoored LMs), does not generalize to U-SOPHISTRY. Our results highlight an important failure mode of RLHF and call for more research in assisting humans to align them.