AdaPlanBench - Physical Plausibility: leaderboard
AdaPlanBench (2026) gives LLM agents 307 household planning tasks whose world and user constraints stay hidden until a proposed plan violates one, so the agent must re-plan from feedback; boards use the mid-level constraint profile. This board is the mean 1 to 5 rubric score of three LLM judges for whether the plans respect real-world physics and realistic tool effects; higher is better.
Metric: Physical plausibility rubric (1-5): whether the plan's actions obey real-world physics and its tool use would produce the claimed effects; mean score (1-5) from three rubric judges (GPT-4.1, DeepSeek-V3.2, Qwen3.6-Flash) over the agent's plans, 307 household planning tasks with progressively revealed world and user constraints at the mid-level constraint profile: a hidden constraint is disclosed only when a proposed plan violates it (GPT-5.4 constraint judges), and the agent re-plans within the turn budget; temperature 0 (GPT-5 family at its fixed default, mean of three runs); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 4.68 | #167 |
| 2 | GPT-5 Mini | 4.62 | #348 |
| 3 | GPT-5 Nano | 4.43 | #889 |
| 4 | Gemini 3 Flash | 4.28 | #164 |
| 5 | Gemini 3.1 Pro (Preview) | 4.26 | #94 |
| 6 | DeepSeek V4 Flash | 4.22 | #258 |
| 7 | Llama 3.3 70B Instruct | 3.82 | #1357 |
| 8 | Qwen 3 14B | 3.52 | #1233 |
| 9 | Qwen 3 32B | 3.5 | #960 |
| 10 | Qwen 3 8B | 3.48 | #1830 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=adaplanbench-physical-plausibility · How It Works · Data refreshed daily, snapshot 2026-11-13.