JevGPT: Thinking Fast and Slow in Autonomous Research

1 Introduction

• A limited research budget constrains how many candidate sources can be inspected, leads pursued, and supporting evidence gathered.Time spent assessing sources affects subsequent research opportunities.Source: §1 · p. 1
• JevGPT achieves leading performance at lower cost and faster completion than comparably effective agents on fixed-corpus question answering and document ranking.Figure 2 presents quality and API cost; the NeuCLIRBench panel shows GPT-based methods, and JevGPT costs include both systems.Source: Figure 2 · p. 2, §1 · p. 2
• Retrieval coverage can exceed evidence use: relevant sources may be available without informing later research decisions.Increased retrieval depth can improve candidate recall even after preview recall saturates (Hsu et al., 2026).Source: §1 · p. 2
• The challenge is to make fast local judgments guide evidence acquisition as needs change, while inspecting many sources within budget.Partial findings may redirect search or reading.Source: §1 · p. 2
• Because planning uses the accumulated investigation while source judgments use explicit evidence objectives, independent judgments can run in parallel and inform the next objective.This links local findings back to the evolving research process.Source: §1 · p. 2
• JevGPT assigns research planning and integration to System Two, while System One guides source selection, passage localization, and continued reading.Selected original passages return to System Two to inform the next research action and final answer or ranking.Source: §1 · p. 2

Agentic Search and Autonomous Research

• Autonomous research systems interleave search, source inspection, and evidence integration, while fixed-corpus systems include bounded workspaces, browsing, and persistent corpus navigation.JevGPT connects local source judgments to subsequent research decisions.Source: §2 · p. 2

Dual-System Reasoning and Collaboration

• Dual-system accounts and model cascades allocate different roles or computation, while JevGPT assigns evidence objectives to System Two and source selection and reading to System One.The overview depicts the two systems, readers returning original passages, adaptive reading loops, and asynchronous discovery leads.Source: §2 · p. 3, Figure 3 · p. 3

Evidence Selection and Context Management

• Evidence-selection approaches include lexical retrieval, reranking, reflective retrieval, and corrective retrieval, whereas JevGPT uses explicit requests to guide passage selection, candidate inspection, and context reading.The cited prior methods include BM25 and RM3, RankGPT, Self-RAG, and CRAG.Source: §2 · p. 3

3 Method

• System Two issues searches or evidence requests and integrates returned findings, while System One assesses sources and guides further reading in the fixed-corpus research loop.Readers return selected original passages to inform later actions and the final output.Source: §3 · p. 3

3.1 Problem Formulation

• A task is a question or topic over a fixed corpus; System Two is a generative policy whose actions include searches, evidence requests, or a final answer or ranking.System One judgments take an objective and source region, with output types and objectives bounding each judgment.Source: §3.1 · pp. 3–4

3.2 System Two: Research Loop

• Evidence requests select retrieval units and an objective for local judgments; gathered original passages enter the research history, while post-search discovery returns findings asynchronously.Evidence memory tracks processing, and question answering and ranking use different objectives and gathering policies.Source: §3.2 · p. 4
• Hybrid BM25 ranks candidates for readers and evidence discovery, and independent source judgments run concurrently once objectives and candidates are fixed.Further inspection and context reading depend on findings already located.Source: §3.2 · p. 4

3.3 System One: Evidence Processing

• For question answering, Noul estimates whether a region addresses a local question and Choice selects original text blocks; gathered passages preserve source coordinates.Readers merge evidence around selected blocks, and questions are assessed independently so a passage can supply an intermediate fact.Source: §3.3 · p. 4
• Reading adapts among inspecting candidates, reading surrounding context, and returning current evidence; partial findings help System Two choose the next action.The reader returns when questions are supported or further inspection stops adding useful material.Source: §3.3 · p. 4
• For ranking, Score judgments assess regions on four relevance levels, and Noul judgments select useful bounded source pages; System Two uses returned estimates and passages to order results.The set returned may include candidates with permissive relevance estimates and incompletely assessed candidates.Source: §3.3 · pp. 4–5
• Asynchronous discovery assesses additional candidates against the original task, using located question clues or ranking relevance to prioritize leads for later inspection.Selected leads carry original excerpts into the research history.Source: §3.3 · p. 5

4.1 Experimental Setup

• Evaluation spans 830 BrowseComp-Plus questions and 105 NeuCLIRBench topics, with a 300-second budget for the main GPT-based comparisons.The tasks are factual question answering and document ranking, respectively.Source: §4.1 · p. 5

4.2 Quantitative Results

• JevGPT achieves the highest observed BrowseComp-Plus accuracy and NeuCLIRBench ranking quality on both metrics, with faster completion and lower cost than comparably effective agents.On NeuCLIRBench, it uses substantially fewer System Two tokens than Pi-Serini and has the lowest total cost among GPT-based methods.Source: §4.2 · p. 5, Table 1 · p. 6, Table 2 · p. 6

4.3 Evidence Analysis

• JevGPT and Pi-Serini have similar candidate coverage, but JevGPT has higher recall for inspected sources and original text delivered to System Two.Figure 4 measures gold-document recall separately for retrieved candidates, inspected sources, sources delivered beyond previews, and sources cited in the final response.Source: §4.3 · p. 6, Figure 4 · p. 6
• For Ancrum, the local evidence objective distinguishes identifying the village from establishing its by-name; System One returns “Little Lovely,” which System Two uses to answer correctly.The returned original sentence supplies a candidate for the next query and a basis for checking the answer.Source: §4.3 · p. 7, Figure 5 · p. 7

4.4 Ablation Study

• Hybrid BM25 and Evidence Discovery provide larger BrowseComp-Plus accuracy gains; Evidence Localization adds a smaller quality gain while reducing searches and increasing cost.The cumulative ablation comparisons are conditional on components already enabled, and the resource pattern is consistent with shifting effort from repeated search toward inspecting retrieved material.Source: §4.4 · p. 7, Table 3 · p. 8

4.5 Performance across System Two Models

• Planner preference depends on the trade-off: GPT-5.5 has highest accuracy, GPT-6.1 Sol retains strong accuracy at much lower cost, GPT-5.4 mini minimizes spending, and Terra finishes fastest.The figure compares cost and accuracy across System Two models; cost and elapsed time require separate consideration.Source: §4.5 · p. 8, Table 4 · p. 8

5 Conclusion

• JevGPT divides research planning and local evidence inspection between System Two and System One, which returns original passages for subsequent reasoning.The architecture supports broad parallel source inspection, proactive evidence discovery, and targeted evidence localization.Source: §5 · p. 8
• Across question answering and ranking, JevGPT achieves leading task performance with lower API cost and faster completion than comparably effective agents.The paper identifies aligning local computation with the planner’s current evidence needs as a route to more capable and efficient autonomous research.Source: §5 · p. 8

Limitations

• JevGPT coordinates fast local evidence judgments and research planning through separate components at the architecture level.A unified model that learns when to make rapid judgments and when to engage in extended reasoning is proposed as a next step.Source: Limitations · p. 9
• Joint training could let a unified model adapt the balance between rapid local judgments and extended reasoning to an investigation’s evolving needs.Source: Limitations · p. 9
• Evaluation covers question answering and document ranking over fixed corpora, with open-web research and long-form synthesis left for future exploration.Source: Limitations · p. 9

Ethical Considerations

• The study uses publicly released BrowseComp-Plus and NeuCLIRBench corpora, collects no additional source data, and redistributes neither corpus.Source: Ethical Considerations · p. 9
• The work involves no human participants or annotators, and retrieved sources may reproduce errors and biases that require independent reliability assessment.JevGPT returns original passages with source locations to support evidence inspection.Source: Ethical Considerations · p. 9
• Claude Code and OpenAI Codex assisted with code development and manuscript editing; GPT Image 2.5 generated illustrative artwork for Figures 1 and 3.The authors directed the research, reviewed outputs, and take responsibility for the methods, results, and manuscript.Source: Ethical Considerations · p. 9

A.1 Implementation

• The main configuration uses GPT-5.5 at medium effort for System Two and Jev-1.13.0 for System One, with a 300-second research budget per task.BrowseComp-Plus combines document and passage BM25 retrieval; NeuCLIRBench uses passage scores to refine document candidates.Source: §A.1 · p. 11
• System One scans retrieval units in batches of ten, up to 60 per request, and runs independent local judgments across QA regions or ranking documents.QA regions overlap and cover at most 60,000 UTF-8 bytes; ranking regions contain at most 48,000 bytes, while responses preserve source coordinates.Source: §A.1 · p. 11
• For QA, System Two supplies up to six evidence questions; regions qualify at τloc = 0.5, with original blocks selected by Choice probability and optional surrounding context.A second block may be selected under the stated probability thresholds, and adaptive reading can add up to two pages before reassessment.Source: §A.1 · p. 11
• For ranking, inspection continues when the estimated chance of finding more relevant documents reaches 0.5; fully assessed documents are retained at relevance probability θ = 0.25.Within retained documents, pages are selected using a 0.5 usefulness cutoff; discovery inspects eligible units among the first K = 60 after searches when at least 45 seconds remain.Source: §A.1 · p. 11
• QA discovery splits the question into up to six clue statements and prioritizes eligible sources by clue coverage and specificity, returning at most three sources.With τlead = 0.5 and m = 3, the system may return the highest-priority source containing a clue even if none meets the cutoff.Source: §A.1 · pp. 11–12
• Ranking discovery prioritizes fully assessed, not-yet-returned documents by relevance probability and returns up to ten at τlead = 0.5.Selected leads include original excerpts in subsequent responses, adding them to the research history for System Two to inspect or use in refining queries.Source: §A.1 · p. 12

A.2 Evaluation

• Evaluation uses all 830 BrowseComp-Plus questions and all 105 NeuCLIRBench topics; the two benchmarks assess fixed-corpus QA and document ranking.NeuCLIRBench queries combine English topic titles and descriptions with English machine translations of Persian, Russian, and Chinese documents.Source: §A.2 · p. 12
• BrowseComp-Plus answer quality is measured by accuracy, while NeuCLIRBench ranking quality uses nDCG@20 and F1@20 averaged across topics.Failed tasks score zero, and unjudged documents receive zero gain for ranking metrics.Source: §A.2 · p. 12
• A.2 studies use fixed research budgets: 300 seconds for cumulative ablations and System Two comparisons, and 3,600 seconds for System One and inspection-depth studies.The System Two comparisons use elapsed time; System One studies cap each local judgment at 600 seconds.Source: §A.2 · p. 12
• Efficiency measures report means per question for QA and per topic for ranking, with System One and System Two API costs reported separately.System Two token usage includes input, output, cached input, and reasoning tokens; API costs cover research requests, excluding grading and local computation.Source: §A.2 · p. 12
• Figure 4 tracks gold-document recall across retrieved candidates, inspected sources, sources with original text delivered beyond previews, and sources cited in answers.The analysis averages recall over all 830 BrowseComp-Plus questions and uses each method’s native retrieval depth.Source: §A.2 · p. 12

A.3 Baseline Settings

• BrowseComp-Plus baselines include Pi-Serini, Agentica, BCP/ReAct, and Reason-ColBERT; NeuCLIRBench baselines include BM25, RM3, Patapsco, RankGPT, and Qwen3-Reranker-8B.BCP/ReAct, Pi-Serini, and Agentica are also adapted to return ranked document lists under JevGPT’s 300-second research budget.Source: §A.3 · pp. 12–13

• The experiments test how adding components affects ranking quality and resource use on the 105 NeuCLIRBench topics.Source: B Supplementary Experiments · p. 13
• They compare System Two model choices for ranking quality and resource use on those topics.Source: B Supplementary Experiments · p. 13
• They also examine how candidate inspection depth affects ranking quality and resource use.Source: B Supplementary Experiments · p. 13

B.1 Ablation Study

• The ablation adds Hybrid BM25, Relevance Assessment, and Evidence Discovery sequentially to Pi-Serini, using GPT-5.6 Terra as medium-effort System Two.Source: §B.1 · p. 13
• Relevance Assessment provides the largest observed gain on both ranking metrics but raises API cost; adding Evidence Discovery slightly lowers nDCG@20 and raises F1@20 while reducing time, tokens, and estimated total API cost.Source: §B.1 · p. 13

B.2 Performance across System Two Models

• The System Two comparison uses five models on the same 105 NeuCLIRBench topics, holding Jev, Hybrid BM25, prompts, and evaluation budget fixed at medium reasoning effort.Source: §B.2 · p. 13
• GPT-5.5 achieves the highest ranking quality, GPT-5.6 Terra approaches its effectiveness and completes research fastest, while GPT-5.4 mini minimizes API cost with lower ranking quality.Source: §B.2 · p. 13

B.3 System One Models and Inspection Depth

• The study compares Jev and GPT-5.5 as System One at K ∈ {0, 20, 60, 120}, with GPT-5.5 fixed as System Two across 105 topics.For each search, K is the largest retrieved rank eligible for Evidence Discovery; System One performs requested Relevance Assessment and Evidence Discovery.Source: §B.3 · p. 13, Table 7 · p. 14
• Increasing K raises both models’ System One API cost, but neither nDCG@20 nor F1@20 improves monotonically.At K = 0, Evidence Discovery is disabled while requested Relevance Assessment remains enabled; K controls proactive inspection, not the total number of documents research can read.Source: §B.3 · p. 13
• Jev matches or exceeds GPT-5.5 in nDCG@20 at every matched K, exceeds its F1@20 at three depths, and runs in under one-quarter the time at under 1% of its System One cost.The comparison uses GPT-5.5 as System Two across 105 NeuCLIRBench topics.Source: §B.3 · p. 13, Table 7 · p. 14

JevGPT Query and Reading Trajectory

• For q556, System One localized the thesis acknowledgment naming Amos Makarau, enabling JevGPT to answer correctly, while Pi-Serini’s repeated reads missed the name and produced an incorrect answer.Both methods found the same thesis, but Pi-Serini’s recorded returns never exposed the acknowledgment paragraph.Source: Original Question: q556 · p. 15, JevGPT Query and Reading Trajectory · p. 16
• For q30, JevGPT inspected the answer-bearing box score beyond the initial five previews and verified Carl Herrera and Gabriel Estaba, while Pi-Serini never opened it and answered incorrectly.System Two checked the remaining playing-time condition after System One returned relevant game and player rows.Source: JevGPT Query and Reading Trajectory · p. 17, Original Question: q30 · p. 17, Pi-Serini Query and Reading Trajectory · p. 18
• For q234, System One selected the explicit Ancrum nickname “Little Lovely” rather than the ADHS promotional phrase, while Pi-Serini returned the phrase as the by-name.JevGPT answered correctly; Pi-Serini’s answer was incorrect.Source: Original Question: q234 · p. 19, Pi-Serini Query and Reading Trajectory · p. 20

NeuCLIRBench Topic

• In the Nenets topic case, System One exposed a passage about melting ice disrupting reindeer migration, which System Two placed near the top of its ranking.JevGPT used all 8 actions, compared with 8 of Pi-Serini’s 44 actions.Source: NeuCLIRBench Topic · p. 21

JevGPT Query and Reading Trajectory

• JevGPT ranked both Arctic-report versions and achieved nDCG@20 0.8208 in 77.9 seconds, while Pi-Serini omitted them and achieved nDCG@20 0.2801 in 232.7 seconds.The two versions count as separate documents in the official evaluation but represent one evidence source in this example.Source: JevGPT Query and Reading Trajectory · pp. 21–22
• After candidate exposure, Pi-Serini repeatedly saw the article’s preview, whereas JevGPT assessed its explicit account of disrupted migration, making the broad article useful for the ranking objective.The Arctic report states that the Nenets’ migration was becoming more complicated and that they failed to complete it in 2018.Source: JevGPT Query and Reading Trajectory · pp. 21–22

NeuCLIRBench Topic

• For the Japan-whaling topic, System Two requested an article’s omitted content to obtain its economic and political explanation, then placed the article second.JevGPT used all 8 actions, compared with 8 of Pi-Serini’s 50 actions.Source: NeuCLIRBench Topic · p. 23

• JevGPT initially scored the culture-and-interests article 3.0 but saw only its title; after correcting an invalid reference, it retrieved the full article.The article cites economic, institutional, and electoral interests.Source: JevGPT Query and Reading Trajectory · p. 23
• JevGPT reached nDCG@20 0.6621 in 59.6 seconds, versus Pi-Serini’s nDCG@20 0.3804 in 158.7 seconds.Source: JevGPT Query and Reading Trajectory · pp. 23–24
• Pi-Serini browsed and read relevant reports and expanded its queries, yet its submitted ranking omitted both the culture-and-interests article and the formal-withdrawal report.Source: JevGPT Query and Reading Trajectory · p. 24
• JevGPT used its initial relevance assessment to select further reading, then returned to the original article for substantive reasons.The case illustrates how explicit coverage information supports targeted follow-up within the research loop.Source: JevGPT Query and Reading Trajectory · p. 24