ToolRobustBench - Tool Interface: leaderboard
ToolRobustBench (Jilin University, 2026) is a stage-wise diagnostic benchmark for tool-calling agents, with perturbation families aligned to tool selection, argument binding, output handling and runtime feedback. The success rate is the share of instances with a fully correct tool call on a native 0 to 1 scale, higher is better, under interface-level perturbations to tool names, descriptions and boundaries.
Metric: Success rate (0-1, native; interface-level perturbations to tool names, descriptions and boundaries; over ToolRobustBench single-family perturbation instances (clean, light, medium, heavy); binary per-instance tool-call success). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 0.97 | #158 |
| 2 | DeepSeek V4 Flash | 0.96 | #254 |
| 3 | Qwen Plus | 0.93 | #788 |
| 4 | DeepSeek V4 Pro | 0.93 | #181 |
| 5 | Gemini 2.5 Pro | 0.9 | #262 |
| 6 | Gemini 2.5 Flash | 0.89 | #452 |
| 7 | GPT-5.4 Mini | 0.86 | #389 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=toolrobustbench-tool-interface · How It Works · Data refreshed daily, snapshot 2026-11-12.