MapSatisfyBench - Tool Selection: leaderboard
Metric: Tool-selection score (%): set overlap between the valid tools the agent called and the annotated required tools, penalizing missing and redundant calls; underspecified map-service requests rebuilt from real anonymized user behavior chains, answered by a ReAct agent with 22 deterministic replayed map tools and a simulated user; GPT-5.3 simulates the user and judges (majority of three judgments), temperature 1; higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 15 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 | 49.42 |
| 2 | Claude Sonnet 4.6 | 49.09 |
| 3 | GPT-4.1 | 46.96 |
| 4 | Gemini 3.1 Pro (Preview) (Thinking) | 45.6 |
| 5 | Qwen 3.6 Plus (Thinking) | 45.36 |
| 6 | Qwen 3.6 Plus (Non-reasoning) | 44.93 |
| 7 | DeepSeek V4 Pro (Non-reasoning) | 43.82 |
| 8 | DeepSeek V3.2 (Non-reasoning) | 43.79 |
| 9 | Qwen 3 235B A22B (Non-reasoning) | 42.36 |
| 10 | Qwen 3 8B (Non-reasoning) | 33.49 |
| 11 | Qwen 3 30B A3B (Non-reasoning) | 28.59 |
Interactive version: theaggregate.ai/benchmark?slug=mapsatisfybench-tool-selection · How It Works · Data refreshed daily, snapshot 2026-09-29.