CUJBench (Retrieval Baseline): leaderboard

CUJBench (2026) tests LLM agents on diagnosing the root cause of injected failures in web applications from snapshots of browser evidence and backend observability. This board is root-cause accuracy in percent when retrieved text evidence is given in a single call without tools; higher is better.

Metric: Root-cause accuracy: single-prediction root-cause accuracy (A@1, %: component and fault type both correct; a run that never submits counts as wrong) on the 25-scenario CUJBench evaluation subset of browser-to-backend failure diagnoses (OpenTelemetry Demo and Tractor Store applications), one run per scenario; no tool loop: BM25-retrieved chunks of the text-serialized evidence are given in a single call, without screenshots; higher is better. Source: arxiv.org. Saturation forecast: Around 2035. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Sonnet 4.620#173
2Llama 4 Scout12#1818
3GLM-4.6V12#710
4Qwen 3 VL 235B A22B Instruct8#532
5Gemini 3.1 Pro (Preview)4#99

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=cujbench-retrieval-baseline · How It Works · Data refreshed daily, snapshot 2026-11-21.