Researchers present T-Search, an open-weight agentic retriever designed for hard multi-step search tasks. Built on Qwen3.6-35B-A3B and trained with adversarially filtered synthetic data, it runs bounded multi-round searches to return ranked evidence chunks with justifications.
- Achieves 56.0 Recall@10 with one rollout and 61.3 with three fused rollouts across seven English and Russian benchmarks.
- Outperforms larger open models while leaving answer generation to downstream models for flexible swapping.
- Includes the release of the model, harness, live demo, and TRuST, the first native-Russian hard-search benchmark.
The authors consider this important as it enables backend and generator components to be swapped without retraining.