The authors ran recursive self-distillation (RSD) on safety-tuned open-weight models: Qwen3 14B and Phi-4 14B.
Recursive self-distillation on benign outputs can progressively misalign safe models
Starting from safety-tuned Qwen3 and Phi-4 models, repeated distillation on the model's own safe generations with harmless cues increases unsafe responses to harmful queries.
Top University
Karim Abdel Sadek · Federico Di Gennaro · Fanny Yang
University of California, Berkeley · ETH Zurich
Research Digest··3 min read
Abdel Sadek et al.
§