jrzhang's picture
Add model predictions and per-task leaderboard
37e7c8f verified
|
Raw History Blame Contribute Delete
3.2 kB
metadata
pretty_name: ActiveVision Results
license: cc-by-4.0
language:
  - en
task_categories:
  - visual-question-answering
tags:
  - benchmark
  - leaderboard
  - model-predictions
  - visual-reasoning
  - active-vision
size_categories:
  - 1K<n<10K
configs:
  - config_name: predictions
    default: true
    data_files:
      - split: train
        path: data/predictions.jsonl
  - config_name: leaderboard
    data_files:
      - split: train
        path: data/leaderboard.jsonl

ActiveVision Results

Model predictions and per-task scores for the ActiveVision benchmark: every final answer behind the leaderboard. 58 chain-of-thought configurations and 9 tool-using coding agents, each on all 85 items (17 tasks × 5).

🌐 Per-task results page · 🖼️ Benchmark dataset · 📄 Paper · 💻 Evaluation code

from datasets import load_dataset
preds = load_dataset("activevisionai/ActiveVision-Results", "predictions", split="train")  # 5,695 rows
board = load_dataset("activevisionai/ActiveVision-Results", "leaderboard", split="train")  # 68 rows

# join with the benchmark images and questions on `id`
items = load_dataset("activevisionai/ActiveVision", split="train")

predictions

One row per (configuration, item).

Field Description
id, task, category the benchmark item (same values as in activevisionai/ActiveVision)
track cot (pure chain-of-thought, no tools) or agent (tool-using coding agent)
model model id, e.g. gpt-6-astra, claude-opus-5-5
harness none for chain-of-thought; codex or claude-code for agents
reasoning_effort effort / thinking level the model was run at
input image; text-only marks a no-image control run
prediction the extracted final answer (empty if the model gave no parsable answer)
answer ground truth
correct graded with the rule below

leaderboard

One row per configuration plus the human reference (track = human): overall correct out of total, accuracy, cost_per_item_usd (API list price; empty for humans), and one column per task with the number of items correct out of 5. The human row is the mean of 3 unaided participants.

Grading

The prediction is the contents of the model's last <answer>...</answer> block, compared by exact match after normalizing case, whitespace, and separators ("C, A, G" ≡ "CAG"; integer answers also match numerically, "07" ≡ 7). See scoring.py.

Only final answers are released — no reasoning text and no agent tool-call traces.

Citation

@article{zhang2026exam,
  title={An exam for active observers},
  author={Zhang, Jiarui and Tao, Muzi and Wang, Shangshang and Liu, Ollie and Ma, Xuezhe and Neiswanger, Willie},
  journal={arXiv preprint arXiv:2607.16165},
  year={2026}
}