TraceBench (Low Noise, Direct): leaderboard
TraceBench (2026) generates controlled time-series root-cause attribution tasks from interpretable mechanical simulators, where an LLM agent decides whether and which system parameter changed; chance is about 19 percent. Episode accuracy under low observation noise when the agent submits its prediction directly, in percent; higher is better.
Metric: Accuracy (%; low observation noise; the agent submits its parameter prediction directly; controlled root-cause attribution tasks generated from three interpretable mechanical simulators (BallDrop, BounceBall, MassSlide): the agent reads time-series observations and decides whether a system parameter was altered and which one; chance 0.189, mean over three simulators). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (High) | 93.3 | #48 (GPT-5.5) |
| 2 | Claude Opus 4.6 (High) | 85.3 | #100 (Claude Opus 4.6) |
| 3 | Gemini 3.1 Pro (Preview) (High) | 76 | #91 (Gemini 3.1 Pro (Preview)) |
| 4 | MiniMax-M2.7 | 29.3 | #519 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=tracebench-low-noise-direct · How It Works · Data refreshed daily, snapshot 2026-11-12.