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.

• The paper introduces Argus, which employs cooperative Searcher and Navigator agents to assemble evidence through a directed acyclic graph.
• It achieves state-of-the-art performance by leveraging parallel Searcher rollouts and compositional verification, thereby minimizing evidence redundancy.
• The system is trained via contrastive reinforcement learning, ensuring auditable, source-traced reasoning and robust generalization across diverse backbones.

Argus: Cooperative Evidence Assembly for Scalable Research Agents

Motivation and Limitations of Parallel Rollouts

Recent advancements in large agentic LLMs for complex web-based information-seeking tasks have relied heavily on two rollouts paradigms: sequential ReAct-style trajectories and test-time parallelism with shallow aggregation. While parallel rollouts—with aggregation via voting, best-of-N, or LLM-based synthesis—help to scale inference compute, they fundamentally lack compositional evidence management. Each trajectory explores only a linear, isolated path, leading to high evidence redundancy, saturation on returns, and context inflation in aggregators. The fundamental bottleneck is a lack of shared, structured state—preventing effective composition of complementary evidence drawn from diverse sources and exploration angles.

Argus System: Agentic Evidence Assembly via DAG Composition

Argus addresses this shortfall with a two-agent cooperative model built on a structured, source-traced evidence graph. The core system components are:

• Searcher: Stateless, ReAct-based agent that processes a single query to return a trajectory of thoughts, tool calls, and observations. The Searcher backbone is fine-tuned via supervised learning, allowing arbitrary parallelization.
• Navigator: The orchestrator. It maintains a shared directed acyclic graph (DAG) encoding evidence (nodes, with URLs) and claim nodes, connected by support and contradiction edges. The Navigator parses Searchers' rollouts into this shared graph, continually labels claims (support/contradiction/unverified), identifies gaps or contradictions, and dispatches new targeted Searcher queries at those gaps. Upon graph completion, it synthesizes the final answer purely from the graph and question, offering auditable, source-traced reasoning for every claim.

This pipeline—search, verify, synthesize—decouples evidence collection (distributed across any number of Searcher rollouts) from the synthesis and verification of reasoning, enabling non-redundant evidence gathering and tractable compositional synthesis.

Agent Training via Reinforcement Learning

The Navigator is trained end-to-end with a contrastive RL scheme. The reward is computed by contrasting synthesis over the post-verification versus pre-verification graph, using LLM-as-judge scoring. The GRPO objective incorporates group-relative advantages akin to PPO but regularized against a reference policy. This approach credits the Navigator for verification steps that causally improve answer correctness, directly optimizing iterative evidence assembly and downstream synthesis. Notably, the policy is invariant to the number of parallel Searchers at inference, enabling the Navigator to generalize to arbitrary compute configurations.

Empirical Evaluation: SOTA Results and Scaling Analysis

Argus was evaluated on eight benchmarks spanning adversarial browsing (BrowseComp, BrowseComp-ZH), tool-use multistep reasoning (GAIA, SEAL-0, xbench DeepSearch-2510), and expert-level scientific QA (HLE, FrontierScience Olympiad/Research). The system was benchmarked in multiple modes: plain Searcher, Argus-Solo (single Searcher + Navigator), and Argus-Parallel (Navigator with 8 Searchers).

Key empirical findings include:

• Argus-Parallel achieves state-of-the-art (SOTA) accuracy on five of eight benchmarks. Gains in Pass@1 compared to the raw Searcher average 5.5 points (solo) and 12.7 points (8-parallel). At 64 Searchers, Argus reaches 86.2% accuracy on BrowseComp, exceeding all baselines, including proprietary large-scale agents.
• Parallel scaling remains log-linear: Evidence duplication is minimized, so gains do not saturate, and the Navigator's context grows sub-linearly with aggregate Searcher output (25.6M tokens from Searchers compressed to a 21.5K token graph, ~1200:1 compression).
• Compositional verification dominates naive aggregation: Argus-Solo outperforms majority voting/best-of-N even at higher K. Structured graph composition at the Navigator unlocks compounding benefit from parallel evidence collection.
• Zero-shot backbone transfer: The Navigator policy transfers without retraining across heterogeneous Searcher backbones (open-source or proprietary), with Argus-Parallel achieving the highest accuracy on each configuration.
• Ablation: Structured DAG input alone provides significant uplift over flat concatenation, and discrimination of support/contradiction arcs further amplifies synthesis accuracy.

The key result is that parallel compute in Argus is spent on closing distinct evidence gaps—not duplicating effort—yielding practical, scalable improvements unconstrained by aggregator context windows.

Implications and Theoretical Significance

This work demonstrates that scaling deep research agents is fundamentally a problem of joint evidence assembly, not just increased computation. Argus's compositional, source-traced state is directly aligned with information-theoretic principles of evidence corroboration, making downstream reasoning fully auditable and increasing reliability in information-seeking tasks. Architecturally, Argus decouples synthesis and verification from evidence collection, bypassing the context window wall faced by flat aggregation methods and opening up new avenues for modular agent design.

Practical implications include:

• True scalability: Compute investments scale accuracy efficiently, not just model size.
• Auditability and traceability: Every claim in a synthesized answer is source-justified, providing transparency for high-stakes applications.
• Generalization: The Navigator's policy is backbone-agnostic, facilitating system integration and transfer learning.

Limitations and Directions for Future Work

Argus currently functions as a heavy-duty solver; inference costs, both in compute and wall-clock time (up to 25.6M tokens/64 Searchers per question), are non-trivial—rendering it unsuitable for latency-sensitive consumer tasks. The ultimate ceiling for answer recall is set by Searcher backbone strength and web corpus coverage; Argus cannot recover correct answers from unavailable or inaccessible sources.

Open research questions include: optimal tradeoff between compute/budget allocation and practical deployment constraints; further abstraction of the evidence graph for richer downstream reasoning; and extending compositional verification to domains beyond web-based QA, such as complex planning or legal/cyber-physical reasoning.

Conclusion

Argus presents a principled and effective approach to scaling deep research agents, shifting from parallel trajectory aggregation to compositional evidence assembly in a graph-structured state. The agent achieves leading performance on challenging benchmarks with persistent accuracy scaling, robust auditability, and transferability across Searcher architectures. This compositional allocation of compute is poised to become a foundational mechanism for building scalable, reliable, and interpretable agentic systems in information-seeking tasks.

Reference: "Argus: Evidence Assembly for Scalable Deep Research Agents" (2605.16217)