Amortized Off-Policy Evaluation for LLMs

Name: Amortized Off-Policy Evaluation for LLMs
Uploaded: 2026-10-15T23:29:48.106566+00:00
Duration: 951 s
Description: PFN-OPE is a network that estimates off-policy evaluation for Large Language Models by treating it as an in-context prediction problem. It uses pretraining on a pool of tasks to learn how to predict the value shift between logged data and target policies. This method achieves significantly lower error than existing baselines when evaluating LLMs across different reward settings.

PFN-OPE is a network that estimates off-policy evaluation for Large Language Models by treating it as an in-context prediction problem. It uses pretraining on a pool of tasks to learn how to predict the value shift between logged data and target policies. This method achieves significantly lower error than existing baselines when evaluating LLMs across different reward settings.

Open the full episode