Teaching AI agents to ask better questions with a world model

What changed. A Monte Carlo inference strategy helped smaller language models ask more informative questions in Collaborative Battleship; MIT reports Llama 4 Scout improved from 8% to 82% versus humans at roughly 1% of GPT-5's cost.

Technique / discovery. Particle-style Monte Carlo inference, question planning, and code-based answer verification.

Why it matters. Explicit belief tracking and simulation can outperform simply scaling the answer model when the task requires active information gathering.

Application. Diagnostic agents, scientific discovery, and interactive search

Evidence maturity. Published paper (source tier A)

Identified bottleneck. Toy environments may overstate transfer to open-ended scientific or medical inquiry.

Caveat / evidence note. Strong controlled result in a simplified environment; external validity remains the main question.

Review status. Reviewed. User requested: Yes.

• MIT CSAIL (opens in a new tab)

A reproduction tutorial is not yet available for this entry. The closest reviewed material is Reliability, uncertainty, and evaluation and AI paradigms and knowledge representation.