JarvisBench - User Interaction: leaderboard
The user-interaction track of JarvisBench (NVIDIA, 2026): an LLM acting as an always-on intermediary answers progress questions from the user about background agent work, replayed causally at an early and a late checkpoint of 30 worker trajectories, plus a generated follow-up question each time. GPT-5.6-Luna grades each answer 0 to 2 for grounded, direct and useful content, rescaled to 0 to 100; higher is better.
Metric: User-interaction score (0-100), the mean grade of the answers to the fixed General question and one generated Follow-up question at an early and a late checkpoint of 30 causally replayed worker trajectories (20 single-agent, 10 multi-agent); each answer graded 0, 1 or 2 for grounded, direct and useful content by GPT-5.6-Luna and rescaled, failed answers counting as zero. Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Non-reasoning) | 96.3 | #30 (GPT-5.6 Sol) |
| 2 | Claude Opus 4.8 | 90.8 | #56 |
| 3 | DeepSeek V4 Pro (Non-reasoning) | 85.8 | #188 (DeepSeek V4 Pro) |
| 4 | GPT-OSS-120B | 84.6 | #700 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=jarvisbench-user-interaction · How It Works · Data refreshed daily, snapshot 2026-11-13.