SMH-Bench (Direct Reasoning) - Compositional Control: leaderboard

Metric: Success rate (%) on Compositional Control (TC2, 250 multi-device, state-dependent, room-dependent and selection-based operations); Direct Reasoning: the full serialized home state (rooms, devices, current values, service schemas) and the request in one prompt, one structured JSON answer replayed in the HomeEnv simulator; final home state checked against the goal with untouched devices preserved; clarification and query answers judged by GPT-5; higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 13 models tracked.

Top models

#ModelScore
1Gemini 3.1 Pro (Preview)81.2
2GPT-5.478.8
3Claude Sonnet 4.673.2
4Claude Haiku 4.573.2
5Qwen 3.5 397B A17B72
6GPT-5.4 Mini66
7DeepSeek V3.2 (Thinking)60.8
8MiniMax-M2.755.6
9Qwen 3.5 Plus51.6
10GLM-550
11Qwen 3.5 9B38
12Qwen 3.5 4B35.6

Interactive version: theaggregate.ai/benchmark?slug=smh-bench-direct-reasoning-compositional-control · How It Works · Data refreshed daily, snapshot 2026-09-29.