Multi-objective materials discovery often relies on large language model (LLM) agents to propose candidates, yet it is constrained by both the number of proposals and how effectively each costly property evaluation informs the next search step. Existing agents mainly store evaluated candidates and their scores, so they can identify successful materials but cannot trace which concrete edits caused the observed property changes. This hampers local refinement when objectives compete—an edit that improves one property may damage another already satisfied constraint.
TRACE (Transition-Aware Residual Control) treats each local edit as the fundamental feedback unit. It records every parent‑edit‑child transition along with the observed property deltas $\Delta$, then aggregates these transition evidences to estimate reusable edit effects. Based on these estimates, TRACE ranks future edits by their predicted ability to reduce the remaining constraint violations of the current candidate while minimizing damage to already satisfied objectives.
In a controlled comparison against LLEMA, the state‑of‑the‑art LLM‑agent baseline, TRACE raises the macro‑average hit rate from 18.13% to 25.96%, demonstrating a clear performance gain.
Blogger's Review: TRACE leverages edit‑level feedback to achieve fine‑grained understanding of property changes, offering a more efficient search strategy for multi‑objective optimization. Its promising results merit further exploration in real‑world materials discovery pipelines.