HELM Robo-Reward-Bench - Core Scenarios: leaderboard
The "Core Scenarios" subtask of HELM Robo-Reward-Bench. Stanford HELM's Robo-Reward-Bench: vision-language models act as reward models over robot manipulation episodes, predicting task progress or success from video.
Metric: Mean Abs. Error (↓ better). Source: crfm.stanford.edu. Saturation forecast: Around June 2027. 25 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (xHigh) | 63.02 | #29 (GPT-5.5) |
| 2 | GPT-5.5 (2026-04-23) | 66.08 | #24 |
| 3 | GPT-5 Mini (2025-08-07) | 69.1 | #174 |
| 4 | GPT-5 | 81.07 | #104 |
| 5 | Gemini 3 Pro (Preview) | 85.08 | #66 |
| 6 | GPT-5.2 | 88.73 | #113 |
| 7 | GPT-5.1 (2025-11-13) | 90.09 | #105 |
| 8 | Gemini 2.5 Pro | 90.16 | #153 |
| 9 | Gemini 3 Flash (Preview) | 91.7 | #84 |
| 10 | Gemini 2.5 Flash | 94.27 | #242 |
| 11 | Gemini 2.5 Flash Lite | 98.98 | #421 |
| 12 | Llama 4 Maverick Instruct | 127.06 | #460 |
| 13 | GPT-5 Nano | 129.51 | #428 |
| 14 | Llama 4 Scout Instruct | 148.5 | #542 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-robo-reward-bench-core-scenarios · How It Works · Data refreshed daily, snapshot 2026-10-23.