ParamBench (3-shot): leaderboard
ParamBench (2026) tests how accurately LLMs fill the parameters of tool calls to real cloud-network APIs, with instances graded into five difficulty levels by parameter nesting, cross-parameter dependencies and values derived from earlier calls. Exact match in percent on the 293 held-out test calls from APIs unseen in training, where the whole parameter object must equal the gold object; three in-context examples from other difficulty levels, higher is better.
Metric: Exact match (%; share of the 293 held-out ParamBench test calls, 24 cloud-network APIs unseen in training, whose generated parameter object equals the gold object exactly; per-call protocol with the target API schema and the outputs of earlier calls, 3-shot prompting). Source: arxiv.org. Saturation forecast: Around November 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 41.3 | #123 |
| 2 | Qwen 3.6 Plus | 39.2 | #222 |
| 3 | DeepSeek V4 Pro | 38.9 | #188 |
| 4 | Claude Opus 4.7 | 34.5 | #79 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=parambench-3-shot · How It Works · Data refreshed daily, snapshot 2026-11-13.