Deep Research agents are rapidly emerging as primary consumers of modern retrieval systems. Unlike human users who issue and refine queries without documenting their intermediate thought processes, Deep Research agents generate explicit natural language reasoning before each search call, revealing rich intent and contextual information that existing retrievers entirely ignore. To exploit this overlooked signal, we introduce: (1) Reasoning-Aware Retrieval, a retrieval paradigm that jointly embeds the agent's reasoning trace alongside its query; and (2) DR-Synth, a data synthesis method that generates Deep Research retriever training data from standard QA datasets. We demonstrate that both components are independently effective, and their combination yields a trained embedding model, AgentIR-4B, with substantial gains. On the challenging BrowseComp-Plus benchmark, AgentIR-4B achieves 68% accuracy with the open-weight agent Tongyi-DeepResearch, compared to 50% with conventional embedding models twice its size, and 37% with BM25. Code and data are available at: https://texttron.github.io/AgentIR/.
Figures & tables
Figure 1: Reasoning-Aware Retrieval ( AgentIR-4B ) vs. conventional retrieval (Qwen3-Embedding-4B) for a task from BrowseComp-Plus, paired with the Tongyi-DR agent. The task has been simplified for display.
Figure 2: Oracle reranking procedure used in DR-Synth (Section 3.3 )
LLM
Retriever
Accuracy
Recall
Search Calls
Tongyi-DR
BM25
33.98
46.83
32.92
Qwen3-Embed-4B
48.67
59.90
31.02
Qwen3-Embed-8B
50.72
61.78
30.43
ReasonIR-8B
51.03
63.62
31.15
Reason-Rewriter + Reason-Embed-8B
31.08
40.15
34.64
Agentic-R
44.70
47.67
31.05
Table 1: End-to-end evaluation on BrowseComp-Plus. For Tongyi-DR with the visit tool, Search Calls are reported as (search + visit). LLM-Rerank refers to listwise reranking the top-20 results with Qwen3-8B.
Agent
Method
Accuracy
Recall
Search Calls
Tongyi-DR
Qwen3-Embed-4B
48.67
59.90
31.02
AgentIR-4B (w/o Training)
55.54
66.13
29.21
AgentIR-4B (w/o Reasoning)
59.40
70.02
27.97
AgentIR-4B
66.27
78.86
25.91
oss-120b-high
Qwen3-Embed-4B
47.59
58.15
29.14
AgentIR-4B (w/o Training)
51.33
63.05
27.72
Table 2: Component ablation. All methods use Qwen3-Embed-4B as the backbone. “ AgentIR-4B (w/o Training)” prepends reasoning traces without additional fine-tuning. “ AgentIR-4B (w/o Reasoning)” trains on DR-Synth -generated WebShaper data (Section 3.3 ) but embeds only the query.
Agent
f(Ht)
Accuracy
Recall
Search Calls
Tongyi-DR
None
59.40
70.02
27.97
Current Reasoning ( AgentIR-4B )
66.27
78.86
25.91
Global Question
63.25
70.98
26.90
Prior Queries
63.13
71.55
27.12
Prior Queries & Reasonings
63.13
74.20
26.48
Prior Queries & Reasonings & Docs
60.00
70.73
27.18
Table 3: Ablation over alternative signals. All models are fine-tuned from Qwen3-Embedding-4B using DR-Synth (Section 3.3 ), varying only the trajectory transformation f(Ht) used as the retriever’s input. “None” embeds only the query, corresponding to the AgentIR-4B (w/o Reasoning) entry in Table 2 .
Figure 3: Effect of embedding k history turns. We fix the Agent to Tongyi-DR. Plot (a) shows the end-to-end accuracy of embedding past k turns, where “None” denotes the “ AgentIR-4B w/o Reasoning” entry in Table 2 . Plot (b) shows the ratio of unique clues covered by using k most recent reasonings, among all clues that have been covered. This is averaged across all trajectories for the k=all setting (Section 5.3 ).
Figure 4: (a) Reasoning for the query in Figure 1 after identifying the candidate artist, Otto Knows from Sweden. (b) Average number of correct vs. incorrect claims (hypotheses) using k most recent reasonings. This is averaged across trajectories for the k = all setting (Section 5.3 )
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: The prompt template used to embed for AgentIR-4B . At turn t , we fill in {reasoning} with τt and {query} with qt . Note that the duplicate “Query:” is intentional due to Qwen3-Embedding-4B Zhang et al. (2025) ’s instruction format.
Figure 6: Prompt template used for the oracle reranker (Section 3.3 ).
Figure 7: Prompt used for the “Prior Queries” ablation.
Figure 8: Prompt used for the “Prior Queries & Reasonings” ablation.
Figure 9: Prompt used for the “Prior Queries & Reasonings & Docs” ablation.
Figure 13: Prompt used to extract the ground truth multi-hop answer list (Step 1).
Figure 14: Prompt used to extract the number of correct vs. incorrect claims (Step 2).
The DeepResearch Agent System is a large language model system engineered for deep information retrieval, multi-step reasoning, and autonomous research tasks. Built upon a sparse activation architecture with 30 billion total parameters of which only 3 billion are activated per token, the system achieves state-of-the-art performance on multiple agent search benchmarks while delivering 3.2 times faster inference compared to dense counterparts of equivalent scale. The system supports a 128K-token context window with hierarchical attention mechanisms that yield 18.7% accuracy and 23.4% recall improvements over standard long-context approaches. A dual-mode reasoning engine provides both a ReAct paradigm for basic multi-step problem solving and an IterResearch mode for high-performance iterative research with up to 20 reasoning steps, collectively delivering a 31.2% accuracy improvement over single-pass baselines. Multi-tool coordination integrates retrieval, computation, web search, and file parsing modules to achieve 92.1% tool-use accuracy. A reinforcement learning optimization framework based on the GRPO algorithm provides token-level policy gradients that improve training stability by 35% and accelerate convergence by 42%. An automated data synthesis pipeline with seed-based expansion achieves a 92.5% usability rate. Benchmark results include 87.3% on Humanity's Last Exam, 85.3% on BrowserComp Chinese, and 91.2% on WebWalkerQA. The system is fully open-sourced, including data synthesis, training, and inference code, and supports applications in academic research, business analysis, R&D support, and education.
Yong Huang, Yulu Huang, for the team Collaboration
Deep research agents extend the role of search engines from retrieving keyword-matched pages to synthesizing knowledge, fundamentally changing how humans interact with information. However, frontier systems remain proprietary, while existing open agents often generalize poorly across different task types, leaving unclear how to train a broadly capable deep research agent. We release QUEST, a family of open models (ranging from 2B to 35B) that serve as general-purpose deep research agents designed to handle a wide range of long-horizon search tasks, with strong capabilities in fact seeking, citation grounding, and report synthesis. To build QUEST, we propose an effective training recipe combining mid-training, supervised fine-tuning, and reinforcement learning. Central to this recipe is a curated data synthesis pipeline based on unified rubric trees, which applies to different task types and enables synthesizing training data with verifiable rewards without human annotation. In addition, QUEST incorporates a built-in context management mechanism that enables effective long-horizon reasoning and knowledge synthesis. Using only 8K synthesized tasks, QUEST approaches or even surpasses frontier closed-source agents across eight deep research benchmarks spanning diverse task types, and achieves the best overall performance among recent open-weight agents. We released everything: models, data, and training scripts.
Deep research agents answer complex questions through iterative loops of searching, reading, and reasoning. Recent work on reasoning-intensive benchmarks such as BrowseComp-Plus shows that well-configured lexical retrieval can surface high-quality evidence, yet agents may still fail to connect documents carrying evidence to the gold documents. We identify a deep research agent's first retrieval move as an important design decision for this setting. We introduce Question's Gambit, a first-move retrieval module that decomposes the question into a set of clues, reformulates them into complementary searches, consolidates the retrieved results, and reranks the candidate pool before the agent begins its iterative search-and-reasoning process. This produces an opening context designed to support both clue aggregation and final-answer verification. We further evaluate on MultiHop-RAG to test whether these benefits transfer beyond BrowseComp-Plus to a more conventional multi-hop question structure. Experiments on BrowseComp-Plus show that Question's Gambit improves retrieval recall and downstream agent accuracy over strong baselines, improving answer accuracy from 83.1% to 90.5% with gpt-5.5 over Pi-Serini, the strongest reported agentic baseline. Our results confirm that effective agentic deep research depends not only on the tools available inside the loop, but also on the quality of the first move. We published our implementation publicly at https://github.com/radinhamidi/Question-s-Gambit.