New paper · SkillAudit: From Fixed-Suite Benchmarking to Skill-Centered Assessment

SAPIENT

Type
Research
Year
2025
Status
active

A conversational recommendation system that combines a learned agent with Monte Carlo tree search for multi-turn planning.

SAPIENT pairs a conversational agent with a planner that explores future dialogue actions through Monte Carlo tree search. The selected plans supervise the conversational agent, creating an iterative self-training loop.

The release includes code and data for evaluating strategic, non-myopic recommendation conversations, plus an efficiency-oriented SAPIENT-e variant.