TaskEvolving Deep Research (Agent, Without Hints): leaderboard
TaskEvolving (Wang, Chen et al., 2026) evolves 500 NQ-Open questions into verifiable deep research tasks, each a graph of atomic steps with fact-grounded rubric checkpoints scored by a Qwen3.7-Max judge. In this setting the model searches the web with Google search and page scraping tools and receives a short everyday question that hides those constraints. Score is the percentage of rubric points met; higher is better.
Metric: Rubric score (%; fact-grounded pointwise checkpoints of 500 evolved deep research tasks, evidential plus analytical, judged by Qwen3.7-Max; agent mode with Google search and web scraping tools; query without hints: a short everyday question). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Pro | 68 | #181 |
| 2 | GLM-5.2 | 68 | #137 |
| 3 | Claude Sonnet 5 | 67 | #108 |
| 4 | Qwen 3.7 Max | 67 | #142 |
| 5 | GPT-5.6 Terra | 67 | #60 |
| 6 | Gemini 3.5 Flash | 57 | #98 |
| 7 | MiniMax-M2.5 | 42 | #590 |
| 8 | Qwen 3.6 Flash | 40 | #422 |
| 9 | Qwen 3 32B | 23 | #960 |
| 10 | Qwen 3 14B | 21 | #1235 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=taskevolving-deep-research-agent-without-hints · How It Works · Data refreshed daily, snapshot 2026-11-12.