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

TRACE：面向多目标材料发现的转换感知残差控制

Abstract: Multi-objective materials discovery with LLM agents is often limited not only by how many candidates can be proposed, but by how effectively each costly property evaluation informs the next search step. Existing agents mainly store evaluated candidates and their scores, so they know which materials succeeded but not which executable edits caused useful property changes. This makes local refinement difficult when objectives compete and an edit that improves one property may damage another.

摘要： 利用大语言模型（LLM）智能体进行多目标材料发现时，其局限性往往不仅在于能提出多少候选材料，还在于每一次昂贵的属性评估能多有效地指导下一步搜索。现有的智能体主要存储已评估的候选材料及其分数，因此它们只知道哪些材料成功了，却不知道哪些可执行的编辑操作导致了有用的属性变化。当目标之间存在竞争，且改善某一属性的编辑可能会损害另一属性时，这种局限性使得局部优化变得十分困难。

We propose TRACE, a transition-aware residual control framework that treats evaluated edits as the basic unit of feedback. TRACE records each local refinement as a parent-edit-child transition with observed property deltas, aggregates transition evidence to estimate reusable edit effects, and ranks future edits by their predicted ability to reduce the current candidate’s remaining constraint violations while avoiding damage to already satisfied objectives.

我们提出了 TRACE，这是一个转换感知残差控制框架，它将已评估的编辑操作视为反馈的基本单位。TRACE 将每一次局部优化记录为“父代-编辑-子代”的转换，并附带观察到的属性增量；通过聚合转换证据来估计可复用的编辑效果，并根据预测的减少当前候选材料剩余约束违规的能力对未来的编辑进行排序，同时避免对已满足的目标造成损害。

In a controlled same-backbone comparison, TRACE improves over LLEMA, the state-of-the-art LLM-agent baseline, raising macro-average hit rate from 18.13% to 25.96%.

在受控的相同骨干模型对比实验中，TRACE 优于当前最先进的 LLM 智能体基准 LLEMA，将宏观平均命中率从 18.13% 提升至 25.96%。