Stepwise preference

One step, scored, with the correction a researcher would make.

• —`current_step` is the action under evaluation; `history` is everything before it.
• —`correction` names the error type and the earliest step that went wrong.
• —`preferred_insertion_point` says where the better action belonged.

Samples

• Analogical RL Representation Comparisonview
• Augmentor Policy Handoffsview
• Can a decoding method that measures novelty in a …view
• CRL Importance Samplingview
• How close to optimality are current state-of-the-…view
• Is there a characteristic geometric structure tha…view
• Long-Horizon Offline GCRL Value Errorsview
• RL For Novel LLM Capabilitiesview