发现信息

  • 作者:Ruiqi Zhang, Wenhao Gao, Yuki Kurashige
  • 年份:2026
  • 发表源:ACS Omega
  • OpenAlex ID:https://openalex.org/W7155856030
  • DOI:https://doi.org/10.1021/acsomega.6c05744
  • arXiv:待核验
  • 检索主题:agent-memory
  • OpenAlex 引用数快照:1(2026-11-12)

摘要线索

Abstract LLM-based agents for computational chemistry commonly place the model inside an accumulating execution loop, causing context growth proportional to workflow length, unpredictable token costs, and degraded plan reliability at scale. We present a chemistry computation harness that confines LLM invocations to three semantically bounded stages─compound identification, planning, and reporting─while delegating all tool dispatch, parallel scheduling, arithmetic derivation, and artifact routing to a deterministic LangGraph executor that consumes zero LLM tokens during execution; a library of domain skill modules injects specialist quantum chemistry protocols into the planner context at planning time without enlarging the execution context. Across five representative quantum chemistry workflows, the planning, calculator, and reporting calls averaged 9700–25,300 tokens per run, with zero execution-stage tokens in every case─more than 93% fewer tokens (before caching) than loop-based multiagent systems─and, on a high-throughput single-point-energy scaling benchmark, a full-scheme plan representation sustained 100% plan-structural correctness across all 30 runs from N = 10 to 100 compounds of one workflow type, generating a 200-node graph of 100 independent single-point jobs from a single planning call. On the workloads tested, harness design made quantum chemistry automation more robust and its cost more predictable; the model ablation shows that model capability still matters.

核验与筛选

  • 当前状态:待筛选、去重和一手来源核验。
  • OpenAlex 元数据和引用数只用于发现,不作为正式引文或质量判断。
  • 记录正式版本、venue、方法细节、实验协议和局限后,才迁入稳定论文卡片。