RobustBench-TC - Schema Drift: 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 This board is tool-call accuracy in percent under a first tool call reporting that a parameter is no longer valid; higher is better.

Metric: Tool-call accuracy (%) under the SchemaD perturbation (transition component: first call returns “parameter X no longer valid”); RobustBench-TC samples drawn from BFCL V3 single-turn, API-Bank, RoTBench, ToolAlpaca and ToolEyes, each scored by its source benchmark's strict scorer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 2031. 18 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1O4 Mini52.3#368
2MUA-RL-8B48.2#1640
3MUA-RL-32B44.2
4Qwen 3 8BNon-reasoning40.7#1873
5DeepSeek R1 Distill Qwen 14B40.7#2771
6Qwen 3 14BNon-reasoning40.7#1269
7ToolRL-Qwen2.5-3B38.2
8MUA-RL-14B37.7#1658
9LoopTool-32B37.2
10ToolRL-Qwen2.5-1.5B37.2#2299
11LoopTool-8B36.7#1473
12Qwen 2.5 1.5B Instruct35.7#7470
13Qwen 2.5 3B Instruct35.7#5392
14Qwen 3 32BNon-reasoning33.7#972
15Llama 3.2 3B Instruct28.6#7241

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

Interactive version: theaggregate.ai/benchmark?slug=robustbench-tc-schema-drift · How It Works · Data refreshed daily, snapshot 2026-11-21.