What it is
A deep-research agent that recursively audits its own provisional answers, preserves verified evidence, and launches targeted follow-up search to improve long-horizon research. Instead of only extending one search trajectory, it turns verification into control: audit a provisional answer constraint-wise, preserve verified parts, and refine or restart. It also learns its own context-update tool to compress history into a reusable research state.
Key results
- WideSearch: 82.0, beating Kimi-K2.6 at 80.8
- HLE: 52.4, above GLM-5 at 50.4 and Qwen-397B at 48.3
- xbench-2510: 71.0, above Miro-1.7 at 62.0
Shows that recursively verifying and refining partial answers is more effective than simply searching longer for deep research tasks.