The Value-of-Information method asks a clarifying question only when the expected benefit exceeds the user's effort, beating tuned rules across four benchmarks.
ORIGINAL PAPER
Value of Information: A Framework for Human–Agent Communication
Agents learned when one more question was worth asking
The method matched or beat the best tuned baseline in 18 of 20 conditions.
Utility improved by up to 1.36 points in high communication-cost settings.
Open-ended question generation, real user time-cost estimates, conversion and error rates, and performance under miscalibrated beliefs.
What the paper supports
Supported findings
The method matched or beat the best tuned baseline in 18 of 20 conditions.
p. 2, contributionsUtility improved by up to 1.36 points in high communication-cost settings.
AbstractLimitations
Questions and actions are predefined; the paper tests when to ask, not how to generate the best question.
The communication-cost model is linear and does not capture real human cognitive burden.
Belief calibration is weaker in medical diagnosis, where uncertainty is more complex.
Companies with a traceable link
Alibaba Group9988 HK · BABA UN
Supports LongBetter clarification policy could reduce costly mistakes in merchant and commerce workflows.
ThematicTencent700 HK
Supports LongThe mechanism could improve user-facing agent completion without excessive questioning.
ThematicPDD HoldingsPDD UW
Supports LongCommerce agents could benefit if clarification improves task completion and conversion.
Thematic