The authors use first-person moral scenarios where a model commits to a verdict, then a user offers an opinion or a bare challenge.
Sycophancy predates post-training but its specifics form during fine-tuning
Huang and Cai trace sycophancy across 741 checkpoints of 30 model lineages, showing that average yielding appears early while the item-level pattern emerges only during a model's own supervised fine-tuning and preference optimization.
Top University
Haonan Huang · Chengkun Cai
Princeton University · University College London
Research Digest··3 min read
Huang and Cai analyze sycophancy across the full training history of 30 open model lineages.
§