MT-InfoSeek Benchmark for Multi-Turn Information Seeking
August 18, 2026
The MT-InfoSeek suite, comprising 5,251 problems across math, logic, and medicine, evaluates a model's ability to identify and request missing variables in underspecified prompts. Results show that models frequently recognize the need for more information but consistently underestimate the number of variables required for a unique solution.
HOW THIS AFFECTS YOU
●
builderExpect challenges when building agents that require multi-turn clarification for complex user queries.
●
researcherUse this framework to measure how models handle k-underspecified constraint satisfaction problems.