一句话：提出 TraceGate：以可配置策略控制崩溃证据的采集、选择与披露，提升 LLM 辅助程序修复效果并支持成本-收益可复现研究。
该论文关注 LLM 辅助程序修复中失败时刻可观测性的可配置化问题：现有方法要么依赖固定的证据包，要么放任 LLM 无约束地调用工具，导致证据披露与修复成本难以控制。作者提出 TraceGate，一个可复用的研究工件，它在修复流程中充当中介：捕获结构化的崩溃快照，按可配置策略选择证据与预算动作，将所选证据物化后提供给现有修复后端，验证候选补丁，并记录决策与结果。工件还提供后端适配器、基准工作流、评估脚本与日志，以支撑可复现的失败时刻可观测性与成本-成功率权衡研究。在四个 Python 基准、五个修复后端与三种模型规模上的评估中，TraceGate 平均额外修复了 20.5% 的残余失败，即在底层后端耗尽自身修复尝试后仍未解决的任务。
TraceGate is a reusable research artifact for policy-controlled disclosure of crash evidence in LLM-assisted program repair. Rather than relying on fixed evidence bundles or unconstrained LLM-driven tool usage, TraceGate mediates the repair process by capturing structured crash snapshots, selecting evidence and budget actions according to configurable policies, materializing the selected evidence for existing repair backends, validating candidate patches, and logging decisions and outcomes. The artifact provides backend adapters, benchmark workflows, evaluation scripts, and recorded logs to enable reproducible studies of failure-time observability and cost-success trade-offs. In evaluations across four Python benchmarks, five repair backends, and three model sizes, TraceGate recovered an average of 20.5% of residual failures, i.e., tasks that remained unsolved after the underlying backends had exhausted their own repair attempts. Screencast: https://youtu.be/Gcr6A7tVBis Artifact: https://github.com/NicolasSchuler/tracegate Data: https://zenodo.org/records/19678467
@inproceedings{paperbot12223,
title = {TraceGate: A Framework for Policy-Controlled Failure Evidence in LLM-Assisted Program Repair},
author = {Nicolas Schuler and Vincenzo Scotti and Raffaela Mirandola},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3834596}
}