AgentIR: Reasoning-Aware Retrieval for Deep Research Agents

Abstract: 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/.

• The paper demonstrates that incorporating agent reasoning traces with query embedding significantly improves retrieval accuracy, achieving up to an 18% absolute gain.
• It introduces a novel data synthesis method that creates agentic sub-query document pairs from QA datasets to address under-specified multi-turn queries.
• Empirical results on the BrowseComp-Plus benchmark show robust improvements across diverse agents, reducing redundant search calls and computational overhead.

AgentIR: Reasoning-Aware Retrieval for Deep Research Agents

Motivation and Background

Deep Research agents, exemplified by recent advances in multi-turn RAG (Retrieval-Augmented Generation), are LLM-driven systems capable of autonomous, multi-step reasoning and information seeking. Unlike human users who express search objectives solely via queries, these agents interleave search calls with explicit reasoning traces, yielding a rich stream of context and intent signals. Conventional retrievers, whether sparse or dense, ignore these signals, leading to persistent ambiguity and retrieval failures when queries are under-specified.

The AgentIR framework directly addresses this omission by proposing a retrieval paradigm where the agent's explicit reasoning trace is concatenated with the issued query, and jointly embedded for document retrieval. The intuition is that reasonings encode the agent's evolving hypothesis space, reflections on prior search outcomes, and disambiguate the often cryptic queries produced in agentic search loops. A complementary challenge addressed is the lack of turn-level relevance data for retriever training in Deep Research, remedied by a data synthesis method that generates agentic sub-query document pairs from standard QA datasets.

Methodology

Reasoning-Aware Retrieval Paradigm

This paradigm utilizes the ReAct-style agent trajectory $(\tau_1, a_1, o_1, ..., \tau_t, a_t, o_t)$, extracting the current reasoning $\tau_t$ and query $q_t$ at each search turn. Rather than embedding $q_t$ only, the retriever $R$ is trained and invoked as $R(\tau_t, q_t)$, leveraging the joint context for improved retrieval relevance. The concatenation template is structured as shown in the paper's main prompt, ensuring compatibility with Qwen3-Embedding-4B instruction format.

Key advantages include:

• Disambiguation of Query: Reasonings clarify intent, transforming ambiguous queries into grounded search objectives (see (Figure 1)).
• Contextualization via Prior Results: Reasonings reflect on prior findings, narrowing the candidate space and preventing redundant or irrelevant searches.
• Hypothesis Injection: Agentic parametric knowledge allows generation of targeted subgenre/entity hypotheses, further refining retrieval scope.

  Figure 1: AgentIR retrieval compared to conventional Qwen3-Embedding-4B on a BrowseComp-Plus task, demonstrating context-grounded retrieval due to reasoning integration.

Data Synthesis for Training

Conventional IR datasets only offer global query-document relevance pairs, insufficient for agentic multi-turn retrieval. The proposed synthesis pipeline leverages agent rollouts on QA datasets: for each $(Q, A, P)$ triple, the agent with a query-only retriever produces a trajectory $\mathcal{H}_T$. Each turn yields $(\tau_t, q_t)$ pairs, and an oracle reranking procedure (based on LLMs) generates turn-level relevant documents and hard negatives (Figure 2).

Contrastive learning is applied over synthesized data using the chosen backbone (Qwen3-Embedding-4B), optimizing cosine similarity between $(\tau_t, q_t)$ and labeled positives, penalizing negatives with temperature $T=0.01$.

Figure 2: Oracle reranking workflow providing supervision for synthesized agentic retrieval training pairs.

Empirical Evaluation

AgentIR demonstrates strong performance on the BrowseComp-Plus benchmark, paired with Deep Research agents including Tongyi-DR, gpt-oss-120B, and GLM-4.7.

• Accuracy: 66–68% for Tongyi-DR, outperforming Qwen3-Embed-8B (51%) and BM25 (37%) by significant margins.
• Efficiency: Search calls reduced substantially relative to baselines (e.g., 25.9 vs. 32.9 for BM25).
• Generalization: Gains transfer to agents with distinct reasoning styles (gpt-oss-120B and GLM-4.7) without retraining.
• Computation: Outperforms LLM-based listwise reranking methods (~10% accuracy gain) without inference overhead.
• Independence of Components: Both reasoning-aware retrieval and fine-tuned training on synthesized data are independently effective; their combination yields maximal improvement.

Ablation and Analysis

Component ablations showed:

• Reasoning Inclusion: Prepending agent reasoning traces with frozen backbone boosts accuracy (+7 points).
• Agentic Training: Training on synthesized, agentic data using only queries adds further gains.
• Combined Approach: Embedding both reasoning and query with agentically trained retriever leads to largest jump (~18% absolute gain).

  Figure 3: Impact of embedding increasing numbers of prior reasoning turns; accuracy plateaus beyond current turn due to redundancy and filtering effects.

Ablations over trajectory variants revealed that alternatives such as prior queries, prior reasonings, global question, and full trajectory (including documents) are inferior to embedding only the current turn's reasoning trace and query. Notably, the reasoning trace acts as a curated summary, discarding outdated and incorrect hypotheses, while full history introduces retrieval noise and propagates errors.

Incorporating prior reasonings does not confer additional benefit: the current reasoning sufficiently summarizes relevant findings, and aggressive history expansion adds incorrect claims disproportionately (see (Figure 4)).

Figure 4: Agent reasoning after candidate identification and the distribution of correct vs. incorrect claims in prior history turns.

Implications and Future Directions

The AgentIR paradigm has practical implications for both agentic and conventional search. As information-seeking increasingly transitions to autonomous agents capable of deep multi-turn reasoning, retrievers must adapt to leverage the explicit intent and hypothesis signals now available "for free" in agent reasoning traces. Standard query expansion methods, like HyDE, hallucinate and misinterpret intent absent this context, leading to degradation in agentic settings.

Theoretically, AgentIR's findings motivate future research in context engineering for retrieval, not just for agents but for retrievers themselves. This includes principled selection and curation of trajectory signals, maximizing information density and minimizing irrelevant hypothesis noise. As Deep Research agents become mainstream, the paradigm signals a shift: agents, rather than humans, will dominate as consumers of retrieval systems, demanding new learning, supervision, and modeling strategies.

Conclusion

AgentIR introduces a reasoning-aware retrieval paradigm for Deep Research agents, leveraging rich contextual signals ignored by conventional retrievers. By embedding agent reasoning traces and synthesizing agentic training supervision from QA datasets, it achieves substantial accuracy and efficiency improvements, generalizing across agent types and settings. Analysis reveals that the agent's current reasoning acts as a context-optimal curator, summarizing relevant history and excising noise, and naive history accumulation underperforms. Future IR research should adapt to this agent-centered context, developing curation-optimized retrieval frameworks for multi-turn agentic workflows.