# Safety Training Modulates Harmful Misalignment Under On-Policy RL, But Direction Depends on Environment Design

arXiv: https://arxiv.org/abs/2604.12500 · HTML: https://mlsanity.com/paper/2604.12500/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 25 of its 42 references that are in the map.
- First posted on arXiv 14 Apr 2026.

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

- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [Universal and Transferable Adversarial Attacks on Aligned Language Models](https://mlsanity.com/paper/2307.15043/) (2023)
- [Constitutional AI: Harmlessness from AI Feedback](https://mlsanity.com/paper/2212.08073/) (2022)
- [Jailbroken: How Does LLM Safety Training Fail?](https://mlsanity.com/paper/2307.02483/) (2023)
- [Towards Understanding Sycophancy in Language Models](https://mlsanity.com/paper/2310.13548/) (2023)
- [Discovering Language Model Behaviors with Model-Written Evaluations](https://mlsanity.com/paper/2212.09251/) (2022)
- [Large Language Models](https://mlsanity.com/paper/2307.05782/) (2023)
- [CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models](https://mlsanity.com/paper/2010.00133/) (2020)
- [BBQ: A Hand-Built Bias Benchmark for Question Answering](https://mlsanity.com/paper/2110.08193/) (2021)
- [Tree of Attacks: Jailbreaking Black-Box LLMs Automatically](https://mlsanity.com/paper/2312.02119/) (2023)
- [RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback](https://mlsanity.com/paper/2309.00267/) (2023)
- [Bias in Bios: A Case Study of Semantic Representation Bias in a High-Stakes Setting](https://mlsanity.com/paper/1901.09451/) (2019)
- [The Effects of Reward Misspecification: Mapping and Mitigating Misaligned Models](https://mlsanity.com/paper/2201.03544/) (2022)
- [Do the Rewards Justify the Means? Measuring Trade-Offs Between Rewards and Ethical Behavior in the MACHIAVELLI Benchmark](https://mlsanity.com/paper/2304.03279/) (2023)
- [Evaluating and Mitigating Discrimination in Language Model Decisions](https://mlsanity.com/paper/2312.03689/) (2023)
- [Language Models Learn to Mislead Humans via RLHF](https://mlsanity.com/paper/2409.12822/) (2024)
- [AI Safety Gridworlds](https://mlsanity.com/paper/1711.09883/) (2017)
- [Safetywashing: Do AI Safety Benchmarks Actually Measure Safety Progress?](https://mlsanity.com/paper/2407.21792/) (2024)
- [SimulatorArena: Are User Simulators Reliable Proxies for Multi-Turn Evaluation of AI Assistants?](https://mlsanity.com/paper/2510.05444/) (2025)
- [On Targeted Manipulation and Deception when Optimizing LLMs for User Feedback](https://mlsanity.com/paper/2411.02306/) (2024)
- [Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time](https://mlsanity.com/paper/2510.04340/) (2025)
- [Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignment](https://mlsanity.com/paper/2510.05024/) (2025)

## Abstract (arXiv)

Specification gaming under Reinforcement Learning (RL) is known to cause LLMs to develop sycophantic, manipulative, or deceptive behavior, yet the conditions under which this occurs remain unclear. We train 11 instruction-tuned LLMs (0.5B-14B) with on-policy RL across 3 environments and find that model size acts as a safety buffer in some environments but enables greater harmful exploitation in others. Controlled ablations trace this reversal to environment-specific features such as role framing and implicit gameability cues. We further show that most safety benchmarks do not predict RL-induced misalignment, except in the case of Sycophancy scores when the exploit relies on inferring the user's preference. Finally, we find that on-policy RL preserves a safety buffer inherent in the model's own generation distribution, one that is bypassed during off-policy settings.