Huawei’s SearchArt trains a 27B long-horizon search agent with verified synthetic tasks
_reachsumit · x · 2026-07-29
Huawei presents SearchArt, a framework for training long-horizon search agents with verification-driven synthetic task generation and a multi-stage post-training pipeline.
- It synthesizes large-scale search-, research- and user-oriented QA pairs from web documents and automatically generated evidence graphs.
- A verification pipeline checks QA consistency, trajectory quality, and evidence relevance before training.
- The verified trajectories are then used for supervised fine-tuning and reinforcement learning.
- The paper reports a 27B model that matches larger frontier models on search benchmarks.
More from coding & agent
- DBHub shows how it upgraded to MCP 2026-07-28 with stateless core and header routing — rokk07 · 2026-07-29
- DBHub Adopts the New MCP Spec: Stateless Core and Caching Practices — db-master · 2026-07-29
- Kimi K3 reportedly works well in Kimi Code and Claude Code via the Responses API — zainhas · 2026-07-29
- Figma and Sentry MCP servers push design data and live errors into coding agents — heypearlai · 2026-07-29
- GitHub MCP Server and Context7 emerge as core tools for coding agents — heypearlai · 2026-07-29
- Cutting half your MCP servers may make your agent smarter overnight — heypearlai · 2026-07-29