TRACE Enhances Multi-Objective Materials Discovery with LLM Agents

Multi-objective materials discovery, often conducted with LLM agents, faces limitations not just in the number of candidates proposed but also in how effectively each costly property evaluation informs subsequent search steps. Existing agents typically store only evaluated candidates and their scores, meaning they know which materials succeeded but lack insight into which specific executable edits caused beneficial property changes. This makes local refinement challenging, especially when objectives conflict and an edit improving one property might degrade another. To address this, researchers propose TRACE (Transition-Aware Residual Control), a novel framework that redefines the basic unit of feedback as evaluated edits. TRACE records each local refinement as a "parent-edit-child" transition, along with the observed property deltas. It then aggregates this transition evidence to estimate reusable edit effects. This allows TRACE to rank future edits based on their predicted ability to reduce a current candidate's remaining constraint violations, while simultaneously avoiding damage to already satisfied objectives. In a controlled comparison, TRACE demonstrated a significant improvement over LLEMA, a state-of-the-art LLM-agent baseline, boosting the macro-average hit rate from 18.13% to 25.96%.