WeAct - GAIA: leaderboard
WeAct (2026): eight deep-research benchmarks using a MiroFlow-derived harness and Qwen3-32B; task-macro accuracy. Prompt methods, full SFT, individual or merged LoRAs and dynamically selected hard/soft adapters are distinct systems, not the original checkpoint. Accuracy (%); higher is better.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | WeAct Study Qwen3-32B - WeAct -soft | 46.6 | |
| 2 | WeAct Study Qwen3-32B - ManuSearch | 41.75 | |
| 3 | WeAct Study Qwen3-32B - Full SFT | 41.75 | |
| 4 | WeAct Study Qwen3-32B - WeAct -hard | 40.78 | |
| 5 | WeAct Study Qwen3-32B - Least-to-Most | 37.86 | |
| 6 | WeAct Study Qwen3-32B - TSPA | 36.89 | |
| 7 | WeAct Study Qwen3-32B - ReAct | 35.92 | |
| 8 | WeAct Study Qwen3-32B - Single LoRA | 35.92 | |
| 9 | WeAct Study Qwen3-32B - DARE-TIES | 35.92 | |
| 10 | WeAct Study Qwen3-32B - TIES-Merging | 34.95 |
Interactive version: theaggregate.ai/benchmark?slug=weact-gaia · How It Works · Data refreshed daily, snapshot 2026-10-21.