RoboRMBench - Lexical Substitution: leaderboard
Vision-language models used as reward models for robot trajectories, re-scored after the goal description is paraphrased by synonym substitution. Score-crossing rate is how often the paraphrase reorders two trajectories. Lower is better. Source: arxiv.org/abs/2609.05401.
Metric: Score Crossing Rate (0-1). Source: arxiv.org. Saturation forecast: Not forecast. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 0.05 | #94 |
| 2 | Gemma 3 12B | 0.12 | #674 |
| 3 | GPT-5.1 | 0.15 | #145 |
| 4 | Gemma 3 27B | 0.18 | #606 |
| 5 | Gemini 3 Flash | 0.27 | #103 |
| 6 | Llama 4 Maverick | 0.27 | #486 |
| 7 | GPT-4o | 0.34 | #357 |
| 8 | Claude Haiku 4.5 | 0.48 | #287 |
| 9 | Gemini 2.5 Flash Lite | 0.48 | #432 |
| 10 | Llama 4 Scout | 0.48 | #648 |
| 11 | Gemini 2.5 Flash | 0.59 | #255 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=robormbench-lexical-substitution · How It Works · Data refreshed daily, snapshot 2026-11-05.