Reinforcement Learning Can Amplify Emergent Misalignment from Harmless Rewards
cs.CL
Submitted: 2026-05-29
Updated: 2026-08-31
Comments: Accepted to EMNLP 2026
Code: https://github.com/lasseruttert/EMA_
License: http://creativecommons.org/licenses/by/4.0/
The gist: Emergent misalignment (EM) is the surprising tendency of language models to become broadly misaligned after fine-tuning on narrowly misaligned examples.
Terminology
Abstract
Emergent misalignment (EM) is the surprising tendency of language models to become broadly misaligned after fine-tuning on narrowly misaligned examples. While EM has been extensively studied in the supervised fine-tuning (SFT) setting, evidence that it also arises from reinforcement learning (RL) is limited to large, closed-source models, leaving the phenomenon expensive to study and difficult to reproduce. We characterize EM from RL in small, off-the-shelf open-weight models along three axes. First, we show that rewarding narrow, overtly misaligned behavior produces substantially higher general-domain misalignment than sample-matched SFT. Second, we show that EM from RL can be induced by reward signals that could plausibly arise naturally, such as unpopular aesthetic preferences or poor rhetorical appeals. Third, we evaluate in-training mitigations developed for SFT-induced EM and find that they broadly transfer, with preventive steering with persona vectors, interleaving safety data and inoculation prompting all performing well.
Sources
- Phantom Transfer: Data Poisoning can Survive Data-Level Defences
- Phi-4 Technical Report
- Conditional misalignment: common interventions can hide emergent misalignment behind contextual triggers
- Unforgotten Safety: Preserving Safety Alignment of Large Language Models with Continual Learning
- Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation
- Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs
- GPT-4o System Card
- Persona Vectors: Monitoring and Controlling Character Traits in Language Models
- A Rank Stabilization Scaling Factor for Fine-Tuning with LoRA
- Thought Crime: Backdoors and Emergent Misalignment in Reasoning Models
- Efficient Memory Management for Large Language Model Serving with PagedAttention
- Subliminal Learning: Language models transmit behavioral traits via hidden signals in data
- Model Spec Midtraining: Improving How Alignment Training Generalizes
- The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models
- Natural Emergent Misalignment from Reward Hacking in Production RL
- School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs
- Learning to summarize user information for personalized reinforcement learning from human feedback
- Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment
- PyTorch: An Imperative Style, High-Performance Deep Learning Library
- Persona Features Control Emergent Misalignment
Related papers
- Exploring Solution Divergence and Its Effect on Large Language Model Problem Solving
- Ishigaki-IDS-Bench: A Benchmark for Generating Information Delivery Specification from BIM Information Requirements
- Subliminal Steering: Stronger Encoding of Hidden Signals
- MedStruct-S: A Benchmark for Key Discovery, Key-Conditioned QA and Semi-Structured Extraction from OCR Clinical Reports
- The End of Transformers? On Challenging Attention and the Rise of Sub-Quadratic Architectures
- Untangling the Mechanisms of Misleading Context in Medical Question Answering