Debugging the Debuggers: Failure-Anchored Structured Recovery for Software Engineering Agents

ASE 2026 · CCF A

Chenyu Zhao, Shenglin Zhang^*, Yihang Lin, Wenwei Gu, Zhimin Chen, Yongqian Sun, Dan Pei, Chetan Bansal, Saravan Rajmohan, Minghua Ma

[Paper]

• Introduces PROBE, a failure-anchored recovery pipeline that turns execution telemetry into localized diagnoses and grounded guidance for failed software-engineering agents.
• Evaluates structured recovery across repository repair, enterprise workflow recovery, and AIOps mitigation.

Multimodal Monitoring and Automated RCA for AgentOps

Core Leader Apr. 2026 – Present

• Lead the design of a low-intrusion observability pipeline that organizes agent execution traces into aligned Trace–Span timelines.
• Develop reflection-guided root-cause analysis for instruction drift, hallucinations, and tool-use failures, with verified diagnoses feeding prompt- and tool-level repair.
• National Undergraduate Innovation Training Program project, selected as a municipal-level project.

ICSE 2027 Build-Bench Challenge Platform and Preparation

Team Member Jul. 2026 – Present

[Code] [Paper]

• Contribute to challenge preparation and platform development for LLM agents that diagnose and repair real package-build failures arising during cross-architecture migration.
• Help implement a tool-augmented repair workflow that spans package inspection, failure localization, source and configuration editing, build submission, and iterative diagnosis from updated build logs.
• Support a reproducible evaluation protocol across x86_64, Arm64, and RISC-V environments, where a repair is accepted only after a clean executable build on the target architecture.

BlueCare: An AI Memory Companion for Older Adults

Team Member Mar. 2026 – Jul. 2026

[Code]

• Conducted field interviews with older adults to identify user needs and guide feature and interaction design.
• Designed a memory-centered AI companion that combines personalized retrieval, multimodal life-story organization, and conversational interaction to support everyday reminiscence and companionship.
• Developed a privacy-aware memory system that organizes conversations, photographs, relationships, and life events into a structured long-term memory base for personalized recall and story generation.
• Combined local data persistence with speech-enabled DeepSeek dialogue, while exploring offline and on-device capabilities to improve accessibility and protect sensitive personal data in care settings.

Knowledge Graph-Enhanced RAG System

Team Member Sep. 2025 – Feb. 2026

[Code]

• Built a knowledge graph-augmented question answering pipeline that models frontend technology concepts as a weighted directed graph and retrieves entity-centric relational context to ground LLM-generated responses.
• Designed a graph-based retrieval and recommendation mechanism that traverses incoming and outgoing relations around matched entities, constructs structured knowledge triples for generation, and ranks related concepts by edge weights for relevance-aware recommendation.
• Integrated graph retrieval with DeepSeek-based generation to support knowledge-grounded QA and graph-aware learning path planning, where neighboring concepts are used to generate structured prerequisite, core, and next-step learning recommendations.