SEE elicits self-evaluation in base LLMs using 160 examples and minimal training
Self-Evaluation Is Already There: Eliciting Latent Judge Calibration in Base LLMs with Minimal Data
Hugging FaceTrending 3
Summary
The authors find that base LLMs can predict external judges' quality scores above chance with few-shot prompting. They introduce Self-Evaluation Elicitation (SEE), which improves calibration using only 160 unique examples—roughly 31 times fewer than reinforcement learning baselines—while preserving answer quality. The elicited self-evaluation transfers across judges the model was never trained against.