RobustBench-TC - Clean: leaderboard
RobustBench-TC (2026) tests tool-calling models under deployment noise grounded in real tool-calling failures, using 199 single-turn samples drawn from BFCL, API-Bank, RoTBench, ToolAlpaca and ToolEyes and scored by deterministic rule-based parsers. This board is tool-call accuracy in percent on the unperturbed reference samples; higher is better.
Metric: Tool-call accuracy (%) on the 199 unperturbed baseline samples of RobustBench-TC, the clean reference for its perturbations; RobustBench-TC samples drawn from BFCL V3 single-turn, API-Bank, RoTBench, ToolAlpaca and ToolEyes, each scored by its source benchmark's strict scorer after a format-tolerant tool-call parse, temperature 0 (DeepSeek-R1-Distill at 0.6), Qwen3-family models with thinking disabled; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 18 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | LoopTool-32B | 77.9 | ||
| 2 | Qwen 3 32B | Non-reasoning | 75.4 | #983 |
| 3 | Qwen 3 14B | Non-reasoning | 74.9 | #1284 |
| 4 | LoopTool-8B | 71.4 | #1497 | |
| 5 | O4 Mini | 70.9 | #372 | |
| 6 | Qwen 3.5 9B | Non-reasoning | 70.9 | #838 |
| 7 | ToolRL-Qwen2.5-1.5B | 70.4 | #2330 | |
| 8 | MUA-RL-32B | 68.3 | ||
| 9 | MUA-RL-8B | 67.8 | #1656 | |
| 10 | Qwen 3 8B | Non-reasoning | 67.3 | #1890 |
| 11 | TL-CodeLLaMA-2 | 65.3 | ||
| 12 | ToolRL-Qwen2.5-3B | 63.8 | ||
| 13 | MUA-RL-14B | 62.8 | #1674 | |
| 14 | ToolRL-Llama3.2-3B | 62.8 | ||
| 15 | DeepSeek R1 Distill Qwen 14B | 61.8 | #2739 |
No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Example tasks
How much traffic should I expect from Las Vegas to Los Angeles this weekend?
What is the area of a rectangle with length 12 meters and width 5 meters?
Interactive version: theaggregate.ai/benchmark?slug=robustbench-tc-clean · How It Works · Data refreshed daily, snapshot 2026-11-22.