Cosmos-Reason1-7B: benchmark results
Provider: Other.
Aggregate Rating 1541 ± 26, rank #1830 of 10916 rated models, from 63 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| SpatiaLQA - Content F1 | 56.1 | Step-content F1 (%, 0-100): harmonic mean of the pooled reca | 57.5 |
| SpatiaLQA - Precondition F1 | 18.3 | Precondition F1 (%, 0-100): harmonic mean of the pooled reca | 56.4 |
| FlagEval EmbodiedVerse - RealWorldQA | 66.8 | Score | 47.8 |
| OVO-S-Bench - Egocentric Perception | 44.79 | Level 1 (instantaneous egocentric perception) multiple-choic | 46.7 |
| Causal-Plan-Bench - Physical Feasibility | 40 | Accuracy (%) on the physical feasibility (why the action is | 46.2 |
| EPIC-Bench - Affordance Region | 5.37 | Manipulation score (0-100) on affordance region localization | 45.5 |
| EPIC-Bench - Placement Region | 32.75 | Manipulation score (0-100) on placement region (where an obj | 44.3 |
| FlagEval EmbodiedVerse - All-Angles Bench | 49.48 | Score | 39.1 |
| EPIC-Bench - Ground Detection | 24.68 | Navigation score (0-100) on ground detection (localizing tra | 35.2 |
| OVO-S-Bench - Generative Spatial Reasoning | 45.52 | Level 3 (generative spatial reasoning: simulation, consisten | 33.3 |
| FlagEval EmbodiedVerse - SAT | 60 | Score | 32.6 |
| Causal-Plan-Bench - Plan Diagnosis and Repair | 28 | Open-ended bad plan diagnosis and repair (find the wrong ste | 30.8 |
Interactive version: theaggregate.ai/model?slug=cosmos-reason1-7b · How It Works · Data refreshed daily, snapshot 2026-11-08.