T-Search: An Open Agentic Retriever and Playground for Hard Multi-Step Search

2026-10-05

AI Summary

This study addresses the limitations of retriever performance and its tight coupling with generators in multi-step search by proposing an open-source intelligent retrieval architecture. Built upon the Qwen3.6-35B-A3B model, the method integrates adversarially filtered synthetic task training, turn-sliced supervised fine-tuning, GSPO reinforcement learning, and recall reward optimization to achieve bounded multi-turn evidence retrieval decoupled from backend systems. As a key contribution, this work introduces TRuST, the first challenging Russian search benchmark. Extensive evaluations across seven benchmarks demonstrate that the proposed approach attains a Recall@10 of 56.0, surpassing the base model by 14.4 points. Furthermore, multi-turn fusion elevates this metric to 61.3, outperforming larger-scale open-source models.

Abstract

We present T-Search, an open-weight agentic retriever for hard multi-step search. Given a question and a search tool over a fixed corpus, it runs a bounded multi-round search and returns a ranked list of evidence chunks with short justifications, leaving answer generation to a downstream model, so backend and generator can be swapped without retraining. 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.

Problem

• multi-step search
• agentic retriever
• evidence retrieval
• hard search

Innovation

• Agentic Retriever
• Multi-Step Search
• Round-Sliced SFT
• GSPO
• Recall Reward