Microsoft Research's CLIO agent uses calibrated deference to design better battery molecules
An AI that knows when it's wrong and adapts — improving redox potential by 130mV then self-correcting.
Microsoft Research and University of Michigan researchers have introduced CLIO (Cognitive Loop via In-Situ Optimization), an AI agent that pairs a continuously-updated belief-state graph with a recursive plan-then-act loop. This architecture enables what the team calls "calibrated deference" — the agent's ability to detect when its own tools or assumptions are failing, then adapt its strategy and generate mechanistic hypotheses to guide experimental revision. Unlike standard black-box optimization, CLIO actively reasons about its knowledge gaps and proposes targeted diagnostics.
In a real-world test, CLIO led a three-round human-AI campaign to design an aqueous organic redox flow battery (AORFB) negolyte. Over 17 candidates, it converged on a top phosphonate compound that delivered a 130 mV improvement in redox potential over literature baselines. When electrochemical tests revealed unexpectedly poor reversibility — a regression no property predictor flagged — CLIO generated competing hypotheses, prioritized discriminating tests, traced the failure to phosphonate-potassium ion pairing, and prescribed a sulfonate replacement. The resulting molecule achieved substantially better reversibility while retaining a 90 mV redox potential advantage, successfully closing the design-make-test-redesign loop.
- CLIO uses a belief-state graph and recursive planning to achieve calibrated deference — recognizing when its tools fail and adapting autonomously.
- The agent identified a phosphonate candidate with a 130 mV redox potential improvement over literature baselines.
- After discovering poor reversibility, CLIO diagnosed ion pairing as the cause and proposed a sulfonate replacement that maintained a 90 mV improvement with better performance.
Why It Matters
CLIO's calibrated deference enables AI to self-correct during materials discovery, accelerating the design-test cycle without constant human intervention.