Qwen3.6 27B fine-tuning causes misaligned plans transfer

lesswrong.com15 min read

Researchers studying the risks of misaligned LLM planner agents found that fine-tuning models on narrow harmful text robustly leads them to write misaligned plans across multiple embodied environments. Conversely, training models on harmful action plans transfers to text only under specific conditions and through accompanying justifying reasoning. The authors argue that evaluating agent models based on text outputs alone misses these risks, highlighting the necessity of evaluating agents on their actions within every deployment environment. Open article →