What's this about?
CoRT replays counterfactual token choices — "what if the model had picked differently here" — to produce fine-grained, rubric-based training signal at the token level. It was the day's top-upvoted paper.
Why it matters
Token-level, rubric-guided signal (78 upvotes) is a more precise training lever than a single end-of-sequence reward, and precision is what makes RL on language models stable.