IC-203Gemini 1.5 Pro's RCA-Agent accuracy drops 68.4% when code execution fails, far exceeding the drops for Claude 3.5 (17.9%) and GPT-4o (15.6%)

Junjielong Xu, Qinan Zhang, Zhiqing Zhong, Shilin He, Chaoyun Zhang, Qingwei Lin, Dan Pei, Pinjia He, Dongmei Zhang, Qi Zhang

SourceOpenRCA: Can Large Language Models Locate the Root Cause of Software Failures?

When the RCA-Agent encounters a failed code execution, the model must use the error feedback to revise its code or adjust its reasoning. Manual review showed that GPT-4o and Claude 3.5 effectively utilize execution feedback (empty output, exception tracebacks) to recover, while Gemini 1.5 Pro rarely does. Quantitatively, Gemini's accuracy on queries involving failed executions drops from 2.69% to 0.85% (a 68.4% reduction), compared to 17.9% for Claude and 15.6% for GPT-4o. This indicates that the agent framework imposes an additional error-tolerance requirement that Gemini does not meet.

Evidence
correlational
Key metric
Gemini 1.5 Pro: 0.85% correct on failed-execution queries, 68.4% drop; Claude 3.5: 9.31%, 17.9% drop; GPT-4o: 7.56%, 15.6% drop; Llama 3.1: 3.18%, 18.04% drop
Caveat
The analysis is based on manual review of intermediate agent steps and a subset of queries involving failed executions. The absolute accuracy numbers are very low, so percentage drops may be sensitive to small absolute differences.
Model
Gemini 1.5 / Gemini Pro 1.5 Gemini 1.5 Pro, Claude 3.5 Sonnet, GPT-4o, Llama 3.1 Instruct
Concepts
Failure mode
Related findings
IC-202, IC-204
Extraction
automatic-extraction