RELAY: Scaling Historical Build Restoration through Rule Reuse and Agentic Repair
Source
arxiv.org
Author
Olivier Nourry, Yoshiki Higo
Date
Why it matters
Shows that having an LLM derive reusable deterministic rules from recurring failures matches per-case agents with fewer build runs. Agents still fix cases the rules miss.
Key takeaways · AI-distilled
Of 14,159 OSS-Fuzz cases, the authors found 69.7% no longer complete reconstruction and verification, because base images and packages change independently of the source. Most failures share a small number of recurring symptoms.
Relay has an LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition → derive and validate deterministic repair rules from representative failures, then reuses those rules without model calls. Rule derivation could be fully automated, and rules restored as many builds as per-case agents with fewer build executions.
Agents still restored cases rules could not, so the best result came from applying rules first and escalating leftovers to an AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition →: more builds restored than either alone, at lower cost than agents alone.
Terms in this piece · Glossary
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
articleSupporting Industrial Test-Failure Analysis with LLM-Based Systems: An Experience ReportEric Jansson, Per Strandberg, Thomas S\"orensen, Eduard Paul Enoiu, Wasif Afzal
articleCode Transformation Rule Synthesis using LLMs: Potential and LimitsAxel Allain, Aymeric Blot, Djamel Eddine Khelladi, Mathieu Acher
articleORCA: Observability-Grounded Program Repair for Microservice IncidentsYuanchen Gao, Yifang Tian, Yiran Li, Charles Zhang, Hans-Arno Jacobsen