ContextLift: Breaking the Spurious Correlation Behind Conditional Emergent Misalignment
Finetuning a language model on narrowly misaligned data can make it broadly misaligned.
ProofPaper ↗
Key points
- Interleaving generic safety data is the most effective known in-training defense, but the misalignment returns as soon as a prompt resembles the finetuning context.
- We argue that this conditional misalignment is a spurious correlation: in the training mix, the context of the finetuning domain occurs only with misaligned answers, so the model keys its behavior on the context rather than on the request.
- We propose ContextLift, a counterfactual augmentation that breaks this correlation without any model, labels or knowledge of the domain: half of the interleaved safety rows are duplicated with a context sentence lifted verbatim from a random training prompt, their aligned answer unchanged.
Sources (1)
- [1]ContextLift: Breaking the Spurious Correlation Behind Conditional Emergent MisalignmentarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 10:18 AM
Finetuning a language model on narrowly misaligned data can make it broadly misaligned.
Interleaving generic safety data is the most effective known in-training defense, but the misalignment returns as soon as a prompt resembles the finetuning context.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 15, 2026Federated Art Chase: Privacy Leakage and Data Poisoning in Federated Fine-Tuning
- Oct 15, 2026Reinforcement Learning-driven Diffusion Models: Language, Vision, and Beyond
- Oct 14, 2026mistralai/Timestral-209M
- Oct 14, 2026ReGRA: Residual-Gradient Adaptation for Complementary Low-Rank Fine-Tuning
- Oct 14, 2026Active Generative Simulation: Meta-Learning What to Simulate for Sim-and-Real Policy Co-Training
- Oct 14, 2026The Design Space of Tri-Modal Masked Diffusion Models
