PersonalHomeBench - Counterfactual Reasoning (Agentic): leaderboard
PersonalHomeBench (2026) tests models as assistants in generated, persona-rich smart homes. Counterfactual reasoning asks paired questions whose hypothetical change does or does not alter the answer; accuracy in percent, acting through the benchmark household toolbox to gather information. Higher is better.
Metric: Accuracy (%) on 2,000 counterfactual questions, macro-averaged over the variants that change and that keep the ground-truth answer (exact match, else a Gemini 2.5 Pro semantic-equivalence judge), over the text-based split of PersonalHomeBench (1,100 generated, manually reviewed households with persona-based occupants and appliances), temperature 0, in the agentic setting (the model must gather household information and act through the benchmark's PersonalHomeTools toolbox, up to 15 turns); each model at its stated reasoning mode; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | GPT-OSS-20B | 63.65 | #1188 | |
| 2 | GPT-4 | 62.9 | #988 | |
| 3 | GPT-4o | 60.94 | #754 | |
| 4 | Qwen 3 30B | Thinking | 58.75 | |
| 5 | nemotron-3-nano-30B-a3B | 57.95 | #1902 | |
| 6 | Qwen 3 30B Instruct | 52.35 | #1371 | |
| 7 | Qwen 3 4B | Non-reasoning | 45.45 | #2379 |
No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=personalhomebench-counterfactual-reasoning-agentic · How It Works · Data refreshed daily, snapshot 2026-11-22.