Google Research Proposes Self-Training Method for LLMs to Learn Clarification

sebkrier · x · 2026-08-30

Nick Swan observed that current AI chat apps often act like "Indecisive Dave," jumping to conclusions instead of asking clarifying questions when facing ambiguous queries, leading to contradictory advice after context updates. A new Google Research paper, "Learning to clarify," addresses this with Action-Based Contrastive Self-Training (ACT).

Key Points:

Problem: RLHF-optimized LLMs often over-hedge or implicitly guess user intents rather than asking clarifying questions, lacking multi-turn conversational skills.
Solution: ACT is a quasi-online preference optimization algorithm based on DPO for data-efficient dialogue policy learning in multi-turn conversations.
Validation: Demonstrated efficacy in real-world tasks like tabular-grounded QA and machine reading comprehension.
New Task: Introduced AmbigSQL, a novel task for disambiguating requests in complex SQL code generation.

Original post →

Google Research Proposes Self-Training Method for LLMs to Learn Clarification

More from Apps