Tencent's CAFE: Self-Improving Search Agents Need Co-Evolving Feedback

Tencent-Hunyuan · hf · 2026-08-26

Tencent's Hunyuan team proposes CAFE, coupling a search agent and a critic via shared parameters to learn in-trajectory corrective feedback. The method reduces hallucinations and improves search performance across benchmarks.

Original post →

More from coding & agent

coding & agent channel →