PaperlayerView exact revision on arXiv

Source-linked AI summary

InfoAgent: Advancing Autonomous Information-Seeking Agents

Gongrui Zhang, Jialiang Zhu, Ruiqi Yang, Kai Qiu, Miaosen Zhang, Zhirong Wu, Qi Dai, Bei Liu, Chong Luo, Zhengyuan Yang, Linjie Li, Lijuan Wang, Weizhu Chen, Yuan Zhang, Xin Li, Zhaoyi Liu, Xin Geng, Baining Guo

arXiv:2509.25189v1cs.CLcs.AI

TL;DR

Deep research agents face challenges in training on sufficiently difficult data and operating in capable real-world tool environments. InfoAgent combines synthetic multi-entity questions with dedicated web search and browsing tools, and its Qwen3-14B-based agent achieves strong results across benchmarks. The paper also reports a scope boundary: its question-synthesis pipeline is evaluated on Wikipedia.

  • Problem

    Open-source deep research agents often search shallowly because training data are relatively simple, and simulated environments may be underpowered for challenging real-world problems.

    Source: §1 · p. 1

  • Method

    InfoAgent synthesizes harder questions from Wikipedia entity trees using subtree sampling and fuzzification, and provides a self-hosted search and browsing environment.

    Source: §1 · p. 2

  • Results

    InfoAgent establishes state-of-the-art performance among open-source models under 15B parameters across benchmarks, including 15.3% accuracy on BrowseComp.

    Source: §1 · p. 2, §4.2 · pp. 7–8

  • Takeaways & Limitations

    The dedicated tool supports effective and efficient training, and InfoAgent achieves higher accuracy with it than with a traditional tool.

    Source: §5 · p. 9

    The synthesis pipeline is evaluated on Wikipedia, whose biases and inaccuracies may be reflected in the resulting training data and models.

    Source: §5 · p. 10

Abstract

from arXiv · show

Building Large Language Model agents that expand their capabilities by interacting with external tools represents a new frontier in AI research and applications. In this paper, we introduce InfoAgent, a deep research agent powered by an innovative data synthesis pipeline and orchestrated web search tools. To construct challenging, hard-to-find queries,we build entity trees and apply sub-tree sampling with entity fuzzification to systematically increase question difficulty. Unlike prior work that relies heavily on commercial search tools, we develop a dedicated self-hosted search infrastructure, enhancing transparency of agent environments and facilitating further advancement of agent capacity. We evaluate the effectiveness of our data pipeline by measuring the average number of tool calls required to correctly answer a question, and also show that our agent yields better performance when equipped with our tools. Our \mbox{InfoAgent} is post-trained from Qwen3-14B using a two-stage recipe: cold-start supervised finetuning to instill long-horizon search behaviors, followed by reinforcement learning which significantly improves reasoning-driven tool use. With our methods, InfoAgent achieves 15.3\% accuracy on BrowseComp, 29.2\% on BrowseComp-ZH, and 40.4\% on Xbench-DS, outperforming prior open-source deep research agents such as WebSailor-72B and DeepDive-32B.

1 INTRODUCTION

InfoAgent addresses challenges in synthesizing sufficiently difficult training data and providing a capable interactive environment for deep research. Its data pipeline and self-hosted search infrastructure accompany strong benchmark results, including performance above larger open-source agents.

1 INTRODUCTION

  • Existing open-source agents often search shallowly because their training data are relatively simple, while simulated environments can be underpowered for challenging real-world problems.The paper calls for training data spanning varied, uncertain information and tools that support high-concurrency searches and consistent results during RL training.Source: §1 · p. 1
  • InfoAgent builds entity trees from Wikipedia and uses subtree sampling with entity fuzzification to create harder questions, alongside self-hosted search and browsing infrastructure.The paper says this approach targets long-horizon retrieval and conjunctive reasoning; its infrastructure provides fine-grained output control and a transparent experimental environment.Source: §1 · p. 2
  • 15.3% accuracy on BrowseComp; InfoAgent outperforms prior open-source deep research agents and also reports 29.2% on BrowseComp-ZH and 40.4% on Xbench-DeepResearch.The paper also reports first place on multiple benchmarks among models with fewer than 15B parameters and performance above some 32B and 72B models.Source: §1 · p. 2

2 RELATED WORKS

The related work situates deep research agents at the intersection of reasoning models, real-world tool use, and information retrieval. It contrasts iterative information-seeking agents with efficient but static and localized RAG systems.

2 RELATED WORKS

  • Reasoning models have extended test-time reasoning, while the paper argues that tools are needed to connect their capabilities with the real world.The cited examples include gold-level performance in mathematics and coding competitions.Source: §2 · pp. 2–3
  • RAG systems retrieve efficiently by treating passages as latent variables, whereas information-seeking agents iteratively use multiple sources and tools.The paper characterizes RAG retrieval as static and localized and agents as more flexible.Source: §2 · p. 3
  • Prior deep research agents use varied approaches, including knowledge-graph question synthesis, real-web search loops, memory-based continual learning, and multi-stage RL.Benchmarks such as BrowseComp, HLE, BrowseComp-ZH, xBench, and GAIA assess integration of more information sources than traditional RAG and knowledge-based QA benchmarks.Source: §2 · p. 3

3 INFOAGENT APPROACH

InfoAgent combines iterative reasoning and web-tool use with a pipeline for synthesizing multi-entity search questions and a dedicated search-and-browse environment. The approach targets sustained retrieval and reasoning through harder questions and higher-quality tool outputs.

3 INFOAGENT APPROACH

  • InfoAgent follows ReAct, iteratively adding tool observations and reasoning traces to the model context while using search for URLs and snippets and browse for deeper page investigation.The paper presents these tools as the agent’s web-based information-seeking interface.Source: §3 · p. 3

3.1 DATA SYNTHESIS PIPELINE

  • The data pipeline builds entity trees from Wikipedia facts, fuzzifies names, numbers, and wording, then samples subtrees to generate questions about root-entity attributes.Fuzzification makes shallow identification harder; questions use constraints consistent with a ground-truth entity space but difficult to resolve from shallow cues alone.Source: §3.1 · p. 4
  • The pipeline filters generated questions by testing whether o3 can solve them and synthesizes ReAct-style reasoning trajectories for supervised fine-tuning.The filtering keeps questions that o3 solves in at least one multi-round attempt; the generated trajectories include concise intermediate reasoning.Source: §3.1 · p. 5
  • On 800-example samples, InfoAgent’s dataset averages 20.3 tool calls with a broader distribution, versus 5.4 for ASearcher and 9.5 for DeepDive.The InfoAgent distribution has substantial density in the 20–50 call range and a non-negligible presence beyond 100 calls.Source: §3.1 · p. 5, Figure 3 · p. 5

3.2 CUSTOMIZED SEARCH TOOL DESIGN

  • The in-house search infrastructure is designed for scalable, high-throughput use and more transparent experiments than reliance on commercial search services.The paper identifies tool quality, availability, throughput, and consistency as important for agentic RL training.Source: §3.2 · p. 5
  • Search retrieves webpage content and selects relevant chunks using the query and search-engine snapshot, followed by embedding and reranking to generate a concise snippet.Snapshots are included among candidate chunks because nearly 15% of websites’ main text cannot be fetched; the two retrieval queries help identify relevant chunks.Source: §3.2 · pp. 5–6
  • Browse retrieves a webpage from a URL and returns a relevant long chunk selected with BM25, embeddings, and a reranker.The browse function uses the webpage title as its retrieval query and returns the URL with the selected chunk.Source: §3.2 · p. 6, Figure 4 · p. 6

4 EXPERIMENTS

InfoAgent leads the reported small open-source model comparisons across deep-research benchmarks, with results also competitive against larger models and strong on SimpleQA. The experiments compare benchmark performance across training setups and examine the roles of cold-start SFT, search-tool quality, and SFT trajectory length.

4.1 EXPERIMENTAL SETUPS

  • The experiments cover five deep-research benchmarks and compare proprietary and open-source baselines; Qwen3-14B receives cold-start SFT followed by GRPO-based RL on filtered synthesized problems.SFT uses 14k trajectories; RL selects 5.7k problems with SFT pass@4 between 0.25 and 0.75.Source: §4.1 · pp. 6–7

4.2 MAIN RESULTS

  • 15.3 on BrowseComp, 29.2 on BrowseComp-ZH, 40.4 on Xbench-DS, and 52.7 on WebWalkerQA are InfoAgent’s top scores in its sub-15B class; RL substantially improves over SFT, and English-only training generalizes to Chinese benchmarks.The evaluation table reports accuracy and computes InfoAgent scores as Avg@4.Source: Table 1 · p. 7, §4.2 · pp. 7–8

4.3 ABLATIONS

  • Using the designed search tool yields higher RL accuracy than Wiki Retriever, whose lower call counts do not produce higher accuracy; Table 2 also reports worse performance when Wiki Retriever is used.The authors describe their tool as providing clean, concise documents that match queries.Source: Table 2 · p. 8, §4.3 · p. 9

4.2 MAIN RESULTS

  • On BrowseComp, InfoAgent scores 15.3 versus WebSailor-72B’s 12.0 and MiroThinker-32B-DPO-v0.1’s 13.0; on WebWalkerQA, its 52.7 exceeds MiroThinker-32B’s 49.3 and WebThinker-32B’s 46.5.Source: §4.2 · p. 8
  • On SimpleQA, InfoAgent scores 90.4, near WebSailor-32B at 92.8 and Kimi-Researcher at 93.6, while o3 and Kimi-Researcher retain an edge on several benchmarks.Source: §4.2 · p. 8

4.3 ABLATIONS

  • Without SFT cold-start, the instruct model gains little reward and accuracy and makes fewer tool calls; the SFT model learns more varied queries, while repeated-query ratios decay during RL.The authors argue that high-quality SFT supports planning, retrieval, and backtracking needed for deep research and further RL.Source: §4.3 · pp. 8–9, Figure 5 · p. 8
  • SFT models trained on trajectories with at least 10 tool calls achieve higher accuracy across challenging benchmarks, but risk exhausting the 32k-token context; shorter trajectories under-generate calls and generalize poorly.The authors recommend mixing a small proportion of short trajectories with longer ones.Source: Table 3 · p. 9, §4.3 · p. 9

5 CONCLUSION

InfoAgent combines synthesized research queries with a dedicated search tool and evaluates a 14B model across multiple benchmarks. The conclusion also identifies limits in context length and source-corpus scope, ethical risks, and reproducibility challenges.

5 CONCLUSION

  • 20 tool calls are required on average for OpenAI o3 to answer the synthesized queries correctly, and InfoAgent matches or surpasses larger open-source models despite having 14B parameters.The conclusion compares InfoAgent with WebSailor-72B and DeepDive-32B.Source: §5 · p. 9
  • InfoAgent achieves higher accuracy with its dedicated search tool than with a traditional tool across the reported evaluation.The paper evaluates BrowseComp-en, BrowseComp-zh, WebWalkerQA, and Xbench-DS.Source: §5 · p. 9
  • A 16K-token context limit during reinforcement learning constrains the difficulty and range of problems the model can address.The authors say longer-context base models and more advanced RL infrastructure could mitigate this limit.Source: §5 · pp. 9–10
  • The synthesis pipeline is evaluated on Wikipedia, which may constrain question scope; broader-web data could increase difficulty and reduce the training–inference gap.Source: §5 · p. 10
  • Open Wikipedia content reduces privacy-violation risks but retains source biases and inaccuracies, while unsafeguarded deployment may exacerbate misinformation risks.The authors encourage human oversight, bias detection, and provenance tracking.Source: §5 · p. 10
  • The data pipeline can be reconstructed but may yield different data because of randomness; reproducing the custom search tool requires substantial iterative engineering.The engineering work includes reducing latency and optimizing distributed parallelism.Source: §5 · p. 10

A DATA SYNTHESIS DETAILS

The appendix describes subtree extraction, reasoning-data generation, and search-tool settings. It details how these components select connected entity nodes, prompt reasoning, retrieve results, and process page content.

A.1 SUBTREE EXTRACTION

  • Subtree extraction samples nodes and includes their paths to the root when the paths fit the remaining node budget.The selected nodes are used to build the output subtree.Source: §A.1 · p. 14

A.2 SFT REASONING CONTENT GENERATION

  • For SFT reasoning content, OpenAI o3 generates reasoning for the next tool call from prior reasoning, calls, and responses, with a lightweight prompt and a length restriction.The prompt asks for no more than five sentences.Source: §A.2 · p. 14

B.1 SEARCH ENGINES

  • The search tool primarily uses Google Search Web API and falls back to Brave and Bing when Google returns no results.The default region and language are us-en, changed to cn-zh for Chinese benchmarks.Source: §B.1 · p. 15

B.2 HYPER-PARAMETERS

  • Search returns up to five results, then ranks page chunks through BM25, embedding retrieval, reranking, and GPT-4o-mini snippet generation.Pages are divided into 128-token chunks; the final snippets use the original query with selected chunks.Source: §B.2 · p. 15
  • Browse uses 2048-token page chunks and successively retrieves the top 40, top 8, and top 1 chunks with BM25, embedding, and reranking.Source: §B.2 · p. 15

B.3 PERFORMANCE OPTIMIZATION

The search environment addresses latency and rollout bottlenecks with caching and a hybrid service strategy. Its snippet-generation prompt directs the model to prioritize query-relevant content and ignore irrelevant webpage material.

B.3 PERFORMANCE OPTIMIZATION

  • The environment faces a trade-off between high-quality retrieval and low latency, and scaling CPU clusters improves QPS without resolving per-request latency.Source: §B.3 · p. 16
  • Caching intermediate request results can dramatically reduce average latency in GRPO training, where trajectories for the same problem often share queries.Source: §B.3 · p. 16
  • A single unusually long trajectory that misses the cache can block rollout, so the system prioritizes low-latency paid search services for long-tail requests.These services handle about 15% of requests on average; other requests go to the self-hosted tool.Source: §B.3 · p. 16
  • Snippet generation focuses on query-relevant information when available and otherwise summarizes the content, while ignoring unrelated webpage material.The prompt also requests a relevance score from 0 to 1.Source: §B.3 · p. 16

C.1 EVALUATION DETAILS

The evaluation procedure uses an LLM judge to check extracted final answers against ground truth, while the RL description specifies grouped trajectory rewards and normalized advantages. The appendix also sets a context limit for evaluation.

C.1 EVALUATION DETAILS

  • The evaluator extracts the final answer from the model's last turn and uses o4-mini to check it against the ground truth.If the model keeps calling tools after reaching 64k context, it is forced to answer.Source: §C.1 · p. 17

C.2 REINFORCEMENT LEARNING ALGORITHM

  • GRPO generates groups of trajectories and estimates each trajectory's advantage from its normalized reward within the group.The reward is an LLM-as-judge correctness score; the displayed rule assigns correctness when the response equals the ground-truth answer.Source: §C.2 · pp. 17–18
  • The clipping threshold and KL-penalty weight are hyperparameters, and the experiments set the KL-penalty weight β to zero to encourage exploration.Source: §C.2 · p. 17

C.3 TRAINING DETAILS OF ABLATION EXPERIMENTS

The ablation models share the main experiment’s training settings, with evaluation adjustments and a controlled search-function change for comparison.

C.3 TRAINING DETAILS OF ABLATION EXPERIMENTS

  • Three ablation models use checkpoints at RL step 120 and a 32k maximum context length to speed evaluation.They otherwise share most main-experiment training settings.Source: §C.3 · p. 18
  • The comparison uses the Wiki Retriever from Search-R1, which encodes text with e5-base-v2 and retrieves from Wiki-18 by vector search.Both tools retrieve at most five documents.Source: §C.3 · p. 18
  • Only the search function changes to the Wiki Retriever; the browse function remains the same for fairness.Source: §C.3 · p. 18

D REINFORCEMENT LEARNING WITH PROCESS REWARD

The section tests whether intermediate target-entity recall can supplement correctness-based RL rewards. Increasing the recall-bonus weight raises training reward but does not improve final performance.

D REINFORCEMENT LEARNING WITH PROCESS REWARD

  • Intermediate artifacts from structured question construction provide process rewards based on whether target entities are recalled during a trajectory.The entity recall rate is added as a bonus when a trajectory fails to find the correct final answer.Source: D REINFORCEMENT LEARNING WITH PROCESS… · pp. 18–19
  • Target-entity recall can be measured with name exact match, source-URL exact match, or an LLM judge using retrieved snippets and pages.The LLM judge uses Qwen3-Reranker-0.6B; sampled trajectories showed it yields more non-zero recall bonuses, so the authors use it for the added reward.Source: D REINFORCEMENT LEARNING WITH PROCESS… · pp. 18–19
  • Increasing λ lifts the reward curve, but produces no significant accuracy-curve difference and does not improve final performance.At λ = 0.9, the reward curve stays about 0.2 above λ = 0; Table 4 reports no final-performance gain as λ increases.Source: Table 4 · p. 19, D REINFORCEMENT LEARNING WITH PROCESS… · p. 19, Figure 7 · p. 19
  • The authors suggest terminal recall bonuses may not teach the specific search step, while correctness rewards may suffice on their relatively simple training data.They leave the mechanism of process reward for agentic RL to future work.Source: D REINFORCEMENT LEARNING WITH PROCESS… · p. 19

E CASE STUDY

The case study traces an agent refining searches to identify a pitcher and answer a debut-team question. The authors state that LLM use was limited to editing the manuscript’s language.

E CASE STUDY

  • The example narrows searches using birthplace and birth-year clues, then browses Bill Dailey’s page and identifies the Cleveland Indians as his debut team.The page lists his debut on August 17, 1961, with the Cleveland Indians.Source: E CASE STUDY · pp. 22–23

F USE OF LLM

  • The authors say GPT-4o was used only for grammar, sentence structure, clarity, and readability; scientific content and conclusions were produced by the human authors.Source: F USE OF LLM · p. 24
Loading 2509.25189v1…