What Is Missing: Interpretable Ratings for Large Language Model Outputs

Researchers introduce the What Is Missing (WIM) rating system for Large Language Models that uses natural-language feedback instead of numerical ratings to improve preference learning. WIM computes ratings by analyzing cosine similarity between model outputs and judge feedback embeddings, producing more interpretable and effective training signals with fewer ties than traditional rating methods.

• →WIM rating system replaces subjective numerical ratings with natural-language feedback for LLM preference learning.
• →The system uses sentence embedding models and cosine similarity to compute ratings from judge feedback describing missing elements.
• →WIM produces fewer ties and larger rating deltas compared to discrete numerical ratings, improving learning signals.
• →The system integrates into existing training pipelines and works with any preference learning method without algorithm changes.
• →Ratings are interpretable as each scalar rating can be traced back to specific judge feedback text for debugging.