P-Bench
Harbor: pendingP-Bench P-Bench is a benchmark for evaluating statistical analysis agents on realistic, open-ended hypothesis-testing problems.
- Samples
- 425
- Released
- 2026
- Scores
- 0
P^3IBench
Proactive Perception-Prediction-Intervention Benchmark
A benchmark with 533 video-query samples evaluating proactive video assistants on perception, risk prediction, and intervention capabilities in real-life streaming scenarios.
How to run
ltb · p-3ibench
$ npm i -g lettokenburn-cli$ ltb bench info p-3ibench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
12 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.6 Luna | 46.6 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen3-VL-8B-Instruct | 41.7 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| JoyVL | 31.6 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen3-VL-4B | 28.6 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| VideoChat3 | 28.1 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Em-Garde | 27.9 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| InternVL3.5-8B-Instruct | 26.8 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen3.5 4B | 26.2 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| qwen3.5-9B | 26.1 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen2.5-VL-3B | 16.7 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| GLM-4.1V-9B-Thinking | 10.2 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| MMDuet2 | 8.9 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
P^3IBench contains 533 samples from 429 unique videos across daily perception, risk alert, and event intervention tasks. It uses a hybrid multi-LLM and human annotation pipeline to ensure precise temporal grounding of response targets.
Scoring uses PRS-B and PRS-U metrics, rewarding correct, timely responses while penalizing missed or unnecessary outputs. Evaluation of 12 models reveals significant struggles with calibrated response decisions, with the best model achieving only 46.59 PRS-B.