daisylin/primbench-results-v2

PrimBench v2 — Full Trajectory Results

Browser-agent benchmark trajectories from the PrimBench v2 sweep (stock browser-use harness, post-hardening).

• Repo: https://github.com/i2x-ai-lab/PrimBench
• Task picks: 1049 per model (519 clean + 530 intervention)
• Harness: stock_browseruse_eval.py with concurrency=4, max_steps=40, timeout=600s

Directory layout per model

<model_dir>/
├── summary.json          top-level aggregation (score/pass/avg)
└── tasks/
    └── <task_id>__<cond>/
        ├── trajectory.json
        └── screenshots/*.png

Models in this release

Model	Pass %	Avg Score
Gemini 3 Flash	52.9%	0.701
GPT 5.4	48.1%	0.657
GPT 5.4 mini	24.2%	0.387
Qwen3-VL-235B	7.9%	0.143

(Opus 4.7 and Gemini 3.1 Pro pending — will be added once runs complete.)

See the top-level summary.json in each model dir for the full per-task pass/fail/score + evaluator details.