T-Search: An Open Agentic Retriever and Playground for Hard Multi-Step Search
arXiv cs.CL · 6 Oct 2026, 03:44 AEST
T-Search is an open-weight agentic retriever designed for challenging multi-step search tasks. It utilizes a bounded multi-round search to provide ranked evidence chunks with justifications, allowing for flexible integration with downstream models.
Key facts
- T-Search is built on Qwen3.6-35B-A3B and trained on adversarially filtered synthetic search tasks with round-sliced supervised fine-tuning followed by GSPO on a recall reward.
- Averaged over seven English and Russian benchmarks with gold evidence annotations, it reaches 56.0 Recall@10 with one rollout, 14.4 points above its base, and 61.3 with three fused rollouts, outperforming larger open models.
- We release the model, harness, live demo, and three benchmarks, including TRuST, the first native-Russian hard-search benchmark.
Read aloud
T-Search is an open-weight agentic retriever designed for challenging multi-step search tasks. It utilizes a bounded multi-round search to provide ranked evidence chunks with justifications, allowing for flexible integration with downstream models. T-Search is built on Qwen3.six-35B-A3B and trained on adversarially filtered synthetic search tasks with round-sliced supervised fine-tuning followed by GSPO on a recall reward. Averaged over seven English and Russian benchmarks with gold evidence annotations, it reaches fifty six point zero Recall@ten with one rollout, fourteen point four points above its base, and sixty one point three with three fused rollouts, outperforming larger open models. We release the model, harness, live demo, and three benchmarks, including TRuST, the first native-Russian hard-search benchmark.
Sources in this cluster
- T-Search: An Open Agentic Retriever and Playground for Hard Multi-Step Search
arXiv cs.CL · 6 Oct 2026, 03:44 AEST