T-Search: An Open Agentic Retriever and Playground for Hard Multi-Step Search
T-Search separates multi-step evidence retrieval from answer generation, so the retriever and downstream model can be swapped independently.
The paper describes an open-weight agentic retriever built on Qwen3.6-35B-A3B. It runs bounded multi-round searches over a fixed corpus and returns ranked evidence chunks with short justifications. Across seven English and Russian benchmarks with gold evidence, it reports 56.0 Recall@10 with one rollout and 61.3 with three fused rollouts. The release includes the model, harness, live demo, and three benchmarks, including TRuST for native-Russian hard search. ArXiv · AI/CL/LG's note
The paper describes an open-weight agentic retriever built on Qwen3.6-35B-A3B. It runs bounded multi-round searches over a fixed corpus and returns ranked evidence chunks with short justifications. Across seven English and Russian benchmarks with gold evidence, it reports 56.0 Recall@10 with one rollout and 61.3 with three fused rollouts. The release includes the model, harness, live demo, and three benchmarks, including TRuST for native-Russian hard search. ArXiv · AI/CL/LG's note
score 7