Intel / article

EviRCA: Decoupling Evidence Extraction from Reasoning for Microservice Root-Cause Analysis

Source: arxiv.org
Author: Yuhao Wang, Zhen Qin, Xingliang Wang, Guochang Li, Weize Li, Shuiguang Deng
Date: Published Sep 18, 2026

Context

EviRCA separates collecting diagnostic evidence from asking an AI to explain an outage. Conventional code first converts operational measurements, request traces and logs into compact evidence cards. The then reasons over those observations through predefined read-only tools, without running code or exploring the raw data itself. Method description.

On OpenRCA, a using data from three enterprise systems, the authors report correct diagnoses in 40.6–43.9% of cases across two models, compared with prior baselines reaching 15.2%. Their failure analysis points to interpreting already-extracted evidence as the remaining bottleneck. The reported improvement supports preparing useful evidence before reasoning, while the absolute accuracy shows that diagnosis remained difficult. Benchmark findings.

Key takeaways · AI-distilled

EviRCA separates root-cause analysis into a deterministic stage that turns raw metrics, traces, and logs into compact 'evidence cards,' and an LLM stage that only reasons over those cards via read-only tools.
On the OpenRCA benchmark (three enterprise systems), EviRCA reached a 40.6-43.9% correct rate versus a prior best of 15.2% for agentic baselines that let the LLM explore raw telemetry directly.
The decoupled design also cut consumption 15-26x and execution time 3-20x compared to those baselines, per the paper.
EviRCA solves 'hard cases' requiring reasoning across time, components, and causes at once -- a setting where prior approaches scored near zero.
Failure analysis found the bottleneck is judging evidence already surfaced, not finding it -- meaning RCA accuracy hinges on the extraction stage's quality, not the LLM's reasoning power.

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.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

Read the source arxiv.org