MiniWeb Training Study - Heldout62: leaderboard
MiniWeb v1 contains 369 human-authored tasks, 37 primitives, and 65 sites. The main 10-agent study uses Gemini 3.5 Flash with a majority of 3 judge votes. Source interfaces comprise 4 general VLMs using browser-use and 6 native agent loops. The training study uses a Qwen3.5-4B planner and rank-64 LoRA executor, with 27,646 skill steps or 27,644 task-centric steps. Decoding is greedy. MiniWeb and WebArena-Lite use 3 runs; live WebVoyager and OnlineMind2Web use 1 run on normalized task sets of 536 and 280 tasks. WA-Lite has 154 tasks and uses a Gemini 3.5 Flash-Lite fuzzy judge for 43 cases. The planner uses browser-use 0.13.3 flash mode, v6, with at most 25 turns. Exact harnesses and trained weights are retained. The exact source No Skills, training, and family-specialist arms retain their configurations. The full 369-task main cohort remains separate. Metric: Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MiniWeb Qwen3.5-4B MiniWebAgent (planner + trained executor) | 58.1 | |
| 2 | MiniWeb Qwen3.5-4B Planner + family specialists | 52.7 | |
| 3 | MiniWeb Fara1.5-4B native loop | 48.4 | |
| 4 | MiniWeb Qwen3.5-4B Planner + task-centric executor (27.6K steps) | 47.8 | |
| 5 | MiniWeb Qwen3.5-4B Planner, untrained executor | 45.2 | |
| 6 | MiniWeb Qwen3.5-4B Single agent, trained | 39.2 | |
| 7 | MiniWeb Qwen3.5-4B Single agent, untrained | 38.7 | |
| 8 | MiniWeb Fara1.5-4B in study harness | 34.7 | |
| 9 | MiniWeb Qwen3.5-4B Single agent, trained without the final done steps | 34.4 |
Interactive version: theaggregate.ai/benchmark?slug=miniweb-training-study-heldout62 · How It Works · Data refreshed daily, snapshot 2026-10-28.