acceptodds
Under review as a conference paper at ICLR 2027

Argus: Evidence Assembly for Scalable Deep Research Agents

Abstract

Deep research agents have achieved remarkable progress on complex information seeking tasks. Even long ReAct style rollouts explore only a single trajectory, while recent state of the art systems scale inference time compute via parallel search and aggregation. Yet deep research answers are composed of complementary pieces of evidence, which parallel rollouts often duplicate rather than complete, yielding diminishing returns while pushing the aggregation context toward the model's limit. We propose Argus, an agentic system in which a Searcher and a Navigator cooperate to treat deep research as assembling a jigsaw from complementary evidence pieces, rather than brute forcing the whole answer in parallel. The Searcher collects evidence traces for a given sub-query through ReAct-style interaction. The Navigator maintains a shared evidence graph, verifying which pieces are still missing, dispatching Searchers to gather them, and reasoning over the completed graph to produce a source-traced final answer. We train the Navigator with reinforcement learning to verify, dispatch, and synthesize, while independently training the Searcher to remain a standard ReAct agent. The resulting Navigator supports rollouts with a single Searcher or many in parallel without retraining. With both Searcher and Navigator built on a 35B-A3B MoE backbone, Argus gains 5.5 points with a single Searcher and 12.7 points with 8 parallel Searchers, averaged over eight benchmarks. With 64 Searchers it reaches 86.2 on BrowseComp, surpassing every proprietary agent we benchmark, while the Navigator's reasoning context stays under 21.5K tokens.

Primary area:
foundation or frontier models, including LLMs
Keywords:
DeepResearch, Test-time scaling, RL, Verification

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aeTTOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Argus, a Searcher–Navigator framework for deep research agents. Independent Searchers collect information through ReAct trajectories, while a Navigator organizes the retrieved findings into a shared evidence graph, identifies missing or conflicting information, and requests targeted follow-up searches. The Navigator then synthesizes a final answer using the assembled graph instead of the complete search histories. The Searcher and Navigator are trained separately, with the Navigator optimized through SFT and RL. Experiments across eight benchmarks show consistent improvements over a standalone Searcher. Additional analyses demonstrate the benefits of graph-based synthesis, targeted verification, and transfer across different Searcher backbones.

Critical strengths
  1. Strong empirical performance. Argus consistently improves over the standalone Searcher across eight benchmarks, achieving average gains of 5.5 points in Solo mode and 12.7 points in Parallel mode.
  2. Well-motivated evidence assembly framework. The shared evidence graph integrates findings from independent Searchers and guides targeted verification, providing a structured approach to multi-agent coordination.
  3. Meaningful component ablations. The paper evaluates Navigator training, targeted follow-up searches, and graph representations. Notably, the full graph improves BrowseComp accuracy by 5.2 points over text-only synthesis using identical Searcher rollouts.
  4. Strong transferability across backbones. The same trained Navigator improves performance with different Searcher models, including DeepSeek and Seed-2.0-Pro, without additional training.
Critical weaknesses
  1. Unclear source of performance gains. Argus achieves substantial improvements, but the total inference budgets are not matched with the baselines. It remains unclear whether the gains primarily come from the proposed Searcher–Navigator architecture or simply from additional test-time computation.
  2. Missing evaluation of broad information seeking. The evaluation primarily focuses on deep, answer-oriented search tasks such as BrowseComp and GAIA. However, Argus is designed to coordinate parallel Searchers and assemble evidence from multiple sources. A benchmark such as WideSearch would be particularly relevant for evaluating information coverage, completeness, and aggregation across many independent findings.
  3. Redundant parallel exploration. Independent Searchers may retrieve overlapping evidence, wasting computation. The paper lacks analysis of evidence redundancy and marginal information gain.
Questions for authors
  1. Inference budget fairness. Could the authors clarify how much of the performance improvement comes from the proposed architecture rather than additional search computation? How would Argus compare with existing approaches under matched inference budgets?
  2. Broad information seeking. Have the authors considered evaluating Argus on WideSearch? This benchmark would better assess whether the proposed evidence assembly framework can improve information coverage and completeness across parallel searches.
  3. Parallel search efficiency. How much evidence overlap exists among Searchers, and how does unique evidence coverage scale with the number of Searchers?
2 comments
AuthorsOpenReview ↗

Response to Reviewer aeTT

Thank you for identifying the broad-search evaluation as a useful test of the framework. We now report WideSearch results, then address compute attribution and evidence overlap.

A1. New evaluation on WideSearch

We evaluated the full WideSearch evaluation set with the same model backbone, tools, and scorer for the standalone Searcher and Argus:

SystemItem F1
Standalone Searcher69.0
Argus, K=8 initial Searchers75.9
Absolute difference+6.9

Evaluation uses the official WideSearch implementation and corresponding LLM judge, with the same evaluation configuration for both methods. The reported metric is Item F1 (f1_by_item). This extends the evaluation beyond deep, single-answer search to broad information collection, with a 6.9-point gain.

A2. Evidence organization versus additional computation

As you noted, Table 4 holds Searcher rollouts fixed. All three variants also use the same synthesis model, input-length limit, and truncation settings. The full graph obtains 74.5%, versus 72.0% for the bare graph and 69.3% for text-only input. The 5.2-point gain therefore requires neither additional retrieval nor a stronger synthesis model. It supports the structured representation, including relations and claim annotations, rather than topology alone.

Table 2 adds a search-allocation control: replacing targeted follow-ups with independent exploration reduces accuracy from 74.50% to 70.33% over three 200-question inference runs. Table 3 tests transfer of the same Navigator without retraining: Argus-Parallel exceeds K=8 LLM aggregation by 18.0, 9.5, and 8.6 points across the three Searcher backbones.

These controls support design benefits, but neither they nor the standalone-versus-K=8 WideSearch comparison establish matched-total-compute superiority. Figure 4 is within-system scaling evidence: black labels count cumulative Searcher tokens, and green labels count final graph-based synthesis input plus output only, excluding earlier Navigator processing and graph maintenance. The current “final graph context” wording needs correction; the green values are not total Navigator cost or graph-input length alone.

A3. Parallel overlap and marginal evidence gain

Initial Searchers explore independently and can retrieve overlapping material. The Navigator consolidates their returns and uses unresolved requirements or conflicts to guide subsequent searches. Consolidating records for synthesis cannot recover computation already spent on overlapping exploration.

The new WideSearch result supports broad-search output quality, while overlap and marginal evidence gain remain unmeasured; neither that result nor Figure 4 is presented as a redundancy analysis.

We hope the requested broad-search evaluation and the complementary controls strengthen the empirical case for the framework within these stated limits.

Thank you for your reply. I have no other questions. I maintain my score.

Reviewer_CyxCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Argus addresses duplicate parallel searches, missed evidence, and overly long histories by separating the Searcher from the Navigator. The latter organizes trajectories into an evidence graph with provenance and support/refutation relations, initiates targeted follow-up searches, and then clears the working context and synthesizes an answer from the graph. The Navigator is trained through SFT and RL; across eight benchmarks, Solo and eight-way initial parallelism improve performance over a standalone Searcher by an average of 5.5 and 12.7 percentage points, respectively.

Critical strengths
  1. The graph serves both follow-up searches and final synthesis, with clear responsibilities for each module; the same Navigator can be transferred to different Searchers without retraining, which has practical value

  2. Improvements are observed on all eight benchmarks with the same Searcher.

Critical weaknesses
  1. The core comparisons remain confounded by compute budgets. Tables 2, 3 do not match actual total compute, and a budget-matched verifier baseline with tool access is missing.

  2. The reliability of the evidence graph has not yet been empirically established. Support states and edges are predicted by the model, and URL deduplication does not guarantee source independence; edge accuracy and the rates at which follow-up searches correct or introduce errors are not measured

Questions for authors
  1. With total tokens or latency matched, how much of an advantage do targeted follow-up searches retain over independent exploration and tool-enabled verifiers? Please report Searcher and Navigator costs separately.

  2. Could you annotate a sample of edges and states, report the proportions of answers that follow-up searches correct or impair, and describe how reprinted evidence or evidence from the same source is handled?

  3. Please provide the SFT scale, evaluation prompts and scorer, and the budget for each configuration. How is graph acyclicity guaranteed in practice?

3 comments
AuthorsOpenReview ↗

Response to Reviewer CyxC

Thank you for the specific questions. We clarify what the submitted controls establish, add a descriptive paired answer-outcome analysis, and provide the implementation details we can substantiate.

A1. Higher accuracy with lower measured token use

Table 3 addresses initial Searcher count: with Searcher-35B-A3B, all three parallel configurations start K=8 Searchers. Majority-Vote scores 56.2%, LLM-Aggregation 56.5%, and Argus-Parallel 74.5%. To examine total resource use, we reuse the two K=8 baseline evaluations and add Argus at K=6, using the same BrowseComp evaluation set, scorer, and model configurations from that Table 3 row:

MethodInitial Searchers KAccuracy (%)Mean total tokens (M/question)
Independent Searcher + Majority-Vote856.22.10
Independent Searcher + LLM-Aggregation856.52.20
Argus668.01.98

Argus exceeds Majority-Vote by 11.8 percentage points and LLM-Aggregation by 11.5 points while using 5.7% and 10.0% fewer total tokens, respectively. Thus, the observed accuracy advantage at this operating point does not depend on a greater aggregate token count. The totals include input and output across all roles: initial and follow-up searches, Navigator trajectory processing, graph construction and updates, final synthesis, and the LLM-Aggregation call where applicable. Table 3 controls initial parallelism, while this supplementary result demonstrates higher accuracy at lower measured total-token use. K counts initial Searchers, not later follow-up dispatches.

The submitted ablations identify complementary design contributions. Table 4 fixes Searcher rollouts, synthesis model, input-length limit, and truncation settings: text-only input scores 69.3%, the bare graph 72.0%, and the full graph 74.5%. The full representation gains 5.2 points without additional retrieval or a stronger synthesis model. Table 2 retains further search but replaces targeted follow-ups with independent exploration: Full obtains 74.50 ± 2.00%, versus 70.33 ± 2.08% over three 200-question inference runs, a 4.17-point difference. Together, these results connect evidence organization and search allocation to the advantage over independent-search aggregation. The aggregation baselines are distinct from a tool-using verifier.

Figure 4 reports cumulative Searcher tokens in black and final graph-based synthesis input plus output in green. Its green values exclude earlier Navigator trajectory processing and graph updates; they are not a substitute for total-token accounting.

A2. Answer-level evidence and graph reliability

We analyzed paired Searcher and Argus-Solo outcomes from three recorded evaluation snapshots using the same checkpoint. The evaluations targeted a fixed 200-question subset. Not all questions had completed outputs from both systems, so the paired counts below include only jointly completed questions. WC denotes Searcher wrong / Argus-Solo correct; CW denotes Searcher correct / Argus-Solo wrong.

SnapshotWCCWNet additional correct answers for Argus-Solo
13418+16
23115+16
34220+22

Argus-Solo produces 16, 16, and 22 more correct answers on the respective paired sets. Each comparison includes both directions of disagreement and yields a positive net difference. These recorded snapshots are separate from the supplementary aggregation comparison above.

For graph reliability, Section 2.2 treats edge and support labels as model predictions. URL deduplication and clustering consolidate records; support has no fixed URL-count threshold, and distinct URLs are not assumed to be independent sources. After consolidation, reasoning uses the maintained state and new returns rather than revisiting old raw records. The paired counts above provide the answer-level reliability measurement; a separate manual edge/state audit has not been performed.

AuthorsOpenReview ↗

Part 2: Response to Reviewer CyxC

A3. Reproducibility and the graph implementation

Navigator SFT uses 2,562 training and 500 held-out validation examples for checkpoint selection; the lowest-held-out-loss checkpoint initializes RL, which uses 5,298 questions (Appendix A). These are separate from the Searcher's approximately 10K SFT trajectories. Appendix A reports graph observation in windows of 15 trajectory rounds, with at most eight windows, for Navigator training rollouts; this is not a per-Searcher search-step cap.

We use the official evaluation implementations and corresponding LLM judges for BrowseComp and WideSearch, with the same evaluation configuration across compared methods within each benchmark. WideSearch is reported as Item F1. These benchmark judges are separate from the DeepSeek-V3.1-Chat training reward judge in Appendix A.

K counts initial Searcher invocations; total dispatches also include follow-ups: D = K + Σᵣmᵣ, subject to a dispatch cap. This is not a total-token or latency measurement. The implementation does not enforce acyclicity: “directed evidence graph” is the accurate term. Appendix A's format validation is not a cycle check, and Table 4's “Full DAG” label should be corrected.

We would appreciate your assessment of the 5.2-point fixed-rollout gain, the 4.17-point search-allocation difference, the explicit paired-output counts, and the 11.8- and 11.5-point advantages over independent-search aggregation with lower measured total-token use. These complementary comparisons connect the component design to answer quality and resource allocation.

Thanks for the clarification. I would like to maintain my score.

Reviewer_hDDwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

A system that uses multiple search agents to investigate a question, then a trained Navigator gathers their findings into a shared record of claims and sources. The Navigator identifies gaps or disagreements, sends out targeted searches, and uses the collected evidence to write the final answerIt reports answer-accuracy gains over its standalone Searcher across eight benchmarks, with larger gains from more initial searchers.

Critical strengths

A clear architecture for coordinating independent searches around shared evidence; testing across eight benchmarks plus transfer to different searcher backbones; and ablations of Navigator training, targeted follow-up, and graph representation.

Critical weaknesses
  • Argus’s targeted follow-up searches are conditioned on the Navigator’s current evidence graph and claim assessments. Although the system can seek corroboration and resolve contradictions, the paper does not evaluate whether query framing reinforces an initial hypothesis or whether the Navigator reliably seeks disconfirming evidence. This may bias the assembled evidence toward what the system is already looking for.
  • Search-result aggregation does not provide a representative estimate of how prevalent a view is. In particular, finding a political opinion in retrieved pages establishes that the view exists, not how common it is. The paper does not evaluate whether Argus can distinguish evidence supporting a claim from evidence indicating its prevalence or representativeness.
Questions for authors

The paper does not measure end-to-end cost or latency or compare systems at matched total inference budgets. Thus, the accuracy gains’ practical efficiency remains unclear. Could you clarify?

1 comment
AuthorsOpenReview ↗

Response to Reviewer hDDw

Thank you for recognizing the architecture and component experiments. We supplement the submitted case with a paired answer-outcome analysis, distinguish evidence support from population prevalence, and clarify the scope of the compute results.

A1. Argus can actively challenge and overturn an initial hypothesis

We analyzed paired Searcher and Argus-Solo outcomes from three recorded evaluation snapshots using the same checkpoint. The evaluations targeted a fixed 200-question subset. Not all questions had completed outputs from both systems, so the paired counts below include only jointly completed questions. WC denotes Searcher wrong / Argus-Solo correct; CW denotes Searcher correct / Argus-Solo wrong.

SnapshotWCCWNet additional correct answers for Argus-Solo
13418+16
23115+16
34220+22

Argus-Solo delivers net gains of 16, 16, and 22 correct answers on the respective paired sets, with both directions of disagreement included in each comparison.

The submitted example in Appendix D (pp. 16–17) shows a mechanism for such revision. The Searcher proposes Jesse Duroha, but the Navigator keeps the claim unverified because constraints remain unchecked. Query v3 explicitly seeks alternative candidates. The subsequent record identifies a thesis-year conflict, marks the original claim as contradicted, and supports Nicholas Constant instead. Thus, the workflow includes an explicit search beyond the initial hypothesis and a decision to reject it.

These results establish a concrete answer-level benefit on the examined pairs. Appendix D complements them with an explicit mechanism: the Navigator searches for alternatives and rejects an initial candidate when conflicting evidence emerges. The workflow can therefore use counterevidence to revise its initial hypothesis.

A2. Claim support and population representativeness are different tasks

A claim that a view exists and a claim that it is prevalent require different evidence. Argus's support label expresses the Navigator's assessment of evidence for the specific claim. Section 2.2 recognizes that different URLs can share a source and sets no fixed URL-count threshold for support. This defines the scope of the method precisely: evidence assembly for information-seeking tasks, with prevalence claims requiring population-level evidence rather than repeated citations.

A3. Higher accuracy with lower total-token use

Table 3 already compares independent-search aggregation with Argus at the same initial Searcher count. With Searcher-35B-A3B and K=8, Majority-Vote scores 56.2%, LLM-Aggregation 56.5%, and Argus-Parallel 74.5%. Thus, the advantage in that table does not come from starting more Searchers.

Furthermore, to compare performance under comparable total-token budgets, we now report a resource comparison that reuses the K=8 Majority-Vote and LLM-Aggregation evaluations from Table 3 and adds a new Argus evaluation at K=6. All three use the same BrowseComp evaluation set and scorer, with the model configurations from Table 3's Searcher-35B-A3B row. Argus achieves 68.0% accuracy at 1.98M total tokens per question, compared with 56.2% at 2.10M for Majority-Vote and 56.5% at 2.20M for LLM-Aggregation. Argus therefore improves accuracy by 11.8 and 11.5 percentage points while using 5.7% and 10.0% fewer tokens, respectively. At this measured operating point, the accuracy advantage is obtained with lower total-token use, not a larger token budget.

The accounting sums input and output tokens across all roles, including follow-up searches, Navigator trajectory processing, graph construction and updates, final synthesis, and the LLM-Aggregation call where applicable. These are mean observed totals per question. K denotes initial Searchers, not all follow-up calls.

The submitted component controls explain why the organization matters. Table 4 holds the retrieved rollouts, synthesis model, input limit, and truncation settings fixed: the full representation scores 74.5%, versus 69.3% for text-only input, a 5.2-point gain without additional retrieval. Table 2 retains further search but replaces targeted follow-ups with independent exploration: 74.50% versus 70.33%, a 4.17-point difference. These controls complement the aggregation comparison by examining evidence representation and search allocation separately.

The paired answer gains, explicit counterevidence case, and representation control establish concrete benefits of the method. Table 3 shows the advantage at the same initial Searcher count; the supplementary K=6 result shows that Argus also outperforms the two K=8 aggregation baselines while using fewer total tokens.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.