Root-Cause Attribution Is a Search Problem: Continual Search for Long-Horizon Agent Failures
The paper argues that long agent logs break one-shot failure diagnosis because the key evidence is sparse and buried.
The authors frame root-cause attribution as a search task, not a single LLM judgment over an execution trace. Their Continual Search method repeatedly pushes the judge to look for unresolved evidence before settling on a diagnosis. They test it across four existing RCA benchmarks and add MegaRCA-Mix, a 50-trial human-annotated set for longer, heavier traces. On MegaRCA-Mix, they report GPT-5.5 F1 rising from 0.349 to 0.498. HF Daily Papers' note
The authors frame root-cause attribution as a search task, not a single LLM judgment over an execution trace. Their Continual Search method repeatedly pushes the judge to look for unresolved evidence before settling on a diagnosis. They test it across four existing RCA benchmarks and add MegaRCA-Mix, a 50-trial human-annotated set for longer, heavier traces. On MegaRCA-Mix, they report GPT-5.5 F1 rising from 0.349 to 0.498. HF Daily Papers' note
score 4