TRACE: Transition-Aware Residual Control for Multi-Objective Materials Discovery

A new framework named TRACE—Transition‑Aware Residual Control—has been introduced to improve multi‑objective materials discovery using large language model (LLM) agents. The system treats each evaluated edit to a candidate material as a discrete feedback unit, recording the parent material, the specific edit applied, and the resulting child material together with the observed changes in material properties. By aggregating evidence from these edit transitions, TRACE can estimate the reusable effect of particular modifications and prioritize future edits that are most likely to reduce remaining constraint violations without compromising objectives that have already been satisfied. In a controlled comparison against the leading LLM‑agent baseline LLEMA, TRACE raised the macro‑average hit rate from 18.13 % to 25.96 %, demonstrating a significant boost in the efficiency of discovering viable material candidates.

The central challenge addressed by TRACE is the difficulty of local refinement when multiple material properties must be optimized simultaneously. Existing agents typically store only the evaluated candidates and their overall scores, lacking information about which specific edits caused the observed property changes. This gap makes it hard to navigate trade‑offs, as an edit that improves one property can inadvertently degrade another. TRACE’s transition‑aware approach fills this void by capturing the causal link between an edit and its property deltas, allowing the system to learn which edits are generally beneficial across different contexts and to avoid those that tend to cause detrimental side effects, thereby enabling more strategic, constraint‑aware exploration of the material space.

Beyond the performance gains, TRACE’s methodology could reshape how autonomous agents conduct iterative scientific searches, emphasizing the importance of granular, edit‑level feedback rather than coarse candidate‑level outcomes. By providing a mechanism to reuse learned edit effects, the framework promises to reduce the number of costly property evaluations required to converge on optimal solutions, potentially accelerating research cycles in materials science and other domains where multi‑objective optimization is critical. The development aligns with broader community values of openness and collaborative advancement, as exemplified by platforms like arXivLabs that support the sharing of innovative tools and methodologies across the scientific ecosystem.

⚡ Effects Interpreter