SMH-Bench (Interactive Agent) - Atomic Control: leaderboard
SMH-Bench (2026) category on single-intent device operations phrased clearly, colloquially or noisily, scored in the executable HomeEnv smart-home simulator. In the Environment-Interactive Agent setting the model starts with partial observability and must query and control the home through tools in a ReAct loop. The board is the success rate in percent; higher is better.
Metric: Success rate (%) on Atomic Control (TC1, 150 single-intent device operations under clear, colloquial or noisy wording); Environment-Interactive Agent: a ReAct loop that starts from partial observability and queries and controls the HomeEnv simulator through tools under a call budget; 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: Estimated already saturated. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 94 | #94 |
| 2 | GPT-5.4 | 94 | #123 |
| 3 | Claude Sonnet 4.6 | 93.3 | #161 |
| 4 | Qwen 3.5 Plus | 91.3 | #232 |
| 5 | GPT-5.4 Mini | 89.3 | #389 |
| 6 | GLM-5 | 88.7 | #260 |
| 7 | Claude Haiku 4.5 | 88 | #568 |
| 8 | Qwen 3.5 397B A17B | 88 | #246 |
| 9 | DeepSeek V3.2 (Thinking) | 87.3 | #406 (DeepSeek V3.2) |
| 10 | Qwen 3.5 9B | 76.7 | #803 |
| 11 | Qwen 3.5 4B | 69.3 | #1175 |
| 12 | MiniMax-M2.7 | 57.3 | #518 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=smh-bench-interactive-agent-atomic-control · How It Works · Data refreshed daily, snapshot 2026-11-13.