PaperlayerView exact revision on arXiv

Source-linked AI summary

Token-Efficient Multi-Agent Collaboration via System One-Guided Computational Division of Labor

Zihan Zhou, Xinzhe Hu, Hanxu Yang, Liangjian Wen, Zhao Kang

arXiv:2610.08155v1cs.MAcs.AI

TL;DR

As agent trajectories lengthen, coordination costs become a critical bottleneck to scalable agentic Web services. S1-MAS separates bounded coordination informed by state and evidence from generative task reasoning, reserving LLM computation for substantive reasoning. Across seven benchmarks, S1-MAS reports state-of-the-art accuracy among evaluated MAS baselines, with GPT-4o token reductions of 44.9%–97.2% and measured end-to-end latency reductions of 37.8%–93.0% versus AgentVerse, DyLAN, and SelfOrg.

  • Problem

    As agent trajectories lengthen, coordination costs become a critical bottleneck to scalable agentic Web services.

    Source: §1 · p. 1

  • Method

    S1-MAS separates bounded coordination informed by state and evidence from generative task reasoning, reserving LLM computation for substantive reasoning.

    Source: §3 · p. 3, §6 · p. 9

  • Results

    Across seven benchmarks, S1-MAS reports state-of-the-art accuracy among evaluated MAS baselines, with GPT-4o token reductions of 44.9%–97.2% and measured end-to-end latency reductions of 37.8%–93.0% versus AgentVerse, DyLAN, and SelfOrg.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    The results support separating coordination from reasoning as an effective pathway toward scalable and cost-efficient agentic Web services.

    Source: §1 · p. 2

Abstract

from arXiv · show

Large language model (LLM)-based multi-agent systems (MAS) have become a promising paradigm for complex information-seeking and reasoning tasks by enabling collaborative problem solving among specialized agents. However, existing MAS frameworks tightly couple task reasoning with coordination operations, including task selection, role assignment, message routing, and context management. As interactions grow, using powerful LLMs for these bounded control decisions introduces substantial token overhead and latency, limiting the scalability of agentic Web services. In this paper, we investigate whether coordination can be decoupled from expensive reasoning without compromising collaborative performance. We propose S1-MAS, a token-efficient multi-agent framework based on System One-guided computational division of labor. S1-MAS assigns bounded coordination decisions to lightweight System One models while reserving open-ended reasoning for capable LLM workers. Specifically, a lightweight controller selects inspection conditions, chooses subsequent tasks, and determines termination, while a compact reader retrieves condition-relevant evidence from authorized sources to support these decisions. Through a decision-evidence loop, selected tasks dynamically determine worker roles and source access, enabling adaptive collaboration without task-specific training. Extensive experiments on seven diverse benchmarks demonstrate that S1-MAS achieves superior accuracy while substantially reducing the inference cost. Across individual comparisons with AgentVerse, DyLAN, and SelfOrg on seven benchmarks, S1-MAS reduces GPT-4o token consumption by 44.9%-97.2% and measured end-to-end latency by 37.8%-93.0%. These results highlight its potential for scalable and cost-effective agentic Web applications.

1 Introduction

The introduction frames repeated LLM coordination as a growing cost in multi-agent collaboration and asks whether lightweight models can handle bounded control without compromising worker reasoning. It presents S1-MAS as a division of labor and reports accuracy and efficiency gains across seven benchmarks.

  • As agent trajectories lengthen, repeated LLM-generated control instructions, history interpretation, and state maintenance make coordination a computational bottleneck.The bounded choices include selecting tasks, deciding whether to inspect further, and determining whether evidence is sufficient.Source: §1 · pp. 1–2
  • The paper asks whether lightweight decision-making models can coordinate multi-agent systems while preserving the reasoning capability of powerful LLM workers.Source: §1 · p. 2
  • S1-MAS separates bounded coordination from open-ended reasoning, using a lightweight controller and reader while LLM workers handle substantive reasoning.The reader retrieves condition-relevant evidence to ground controller decisions, enabling adaptive collaboration without task-specific training.Source: §1 · pp. 2–3
  • Across seven benchmarks, S1-MAS achieves state-of-the-art accuracy among evaluated MAS baselines, reducing GPT-4o tokens by 44.9%–97.2% and measured end-to-end latency by 37.8%–93.0% versus AgentVerse, DyLAN, and SelfOrg.Source: §1 · pp. 2–3

2 Related Work

Prior work improves agent orchestration, communication efficiency, and the use of System One models with LLMs, but leaves costly coordination decisions as a distinct challenge. S1-MAS addresses that challenge through pretrained System One coordination and evidence-based communication.

Orchestration in Multi-Agent Systems

  • Orchestration methods adapt agent roles, participation, communication structures, or workflows, but generally retain expensive LLM-based coordination decisions.S1-MAS instead uses a pretrained System One controller for bounded coordination without task-specific training.Source: §2 · p. 3

Communication Optimization in Multi-Agent Systems

  • Communication optimization methods prune exchanges or compress messages, while S1-MAS targets the cost of deciding what information to access or exchange.The paper describes this as a complementary efficiency challenge.Source: §2 · p. 3

Combining System One Models with LLMs

  • Prior System One integrations address application-specific control, while S1-MAS combines pretrained System One coordination with evidence-based communication in an end-to-end LLM-based MAS.The paper reports reduced worker-token consumption and end-to-end latency while maintaining collaborative task performance.Source: §2 · p. 3

3 Problem Formulation

The paper formulates multi-agent coordination as sequential choices about worker tasks, information access, and stopping. It represents execution dependencies, maintains a state for coordination, and uses inspection evidence and a compiled menu of feasible tasks to guide the next decision.

  • Coordination is modeled as sequential decisions about worker tasks, the information workers receive, and when collaboration stops.The state after t completed worker executions is St; controller and reader calls are not counted in t.Source: §3 · p. 3
  • A directed acyclic communication graph records completed worker executions as vertices and direct response dependencies as edges.An edge connects an earlier execution to a later task when its complete response is provided as input.Source: §3 · p. 3
  • The coordination state combines the question and task–response archive with inspection history and the remaining source-transmission budget.The archive retains associated roles, sources, and observable check results; inspection records track attempts, not correctness.Source: §3 · pp. 3–4
  • The controller pairs a question sentence with a candidate answer for inspection, and the reader localizes relevant response evidence for the next decision.Source: §3 · p. 4
  • A deterministic compiler limits the task menu to candidates whose prerequisites are met and whose source inputs fit the remaining communication budget.If the controller selects a task, the worker receives its required source responses; selecting finish returns an archived response without another worker execution.Source: §3 · p. 4

4 Methodology

S1-MAS separates bounded coordination from open-ended worker reasoning through a controller, evidence reader, and LLM workers. Its decision–evidence loop supports adaptive collaboration without task-specific training.

4.1 Inspection Query and Evidence Localization

  • The controller and reader select inspection targets, localize relevant evidence, and use it to guide coordination while workers retain responsibility for substantive review.The reader's localized passages are attention cues, not correctness certifications.Source: §4.1 · p. 4

4.2 Task Construction and System One-Guided Selection

  • A deterministic compiler filters tasks by prerequisites and communication budget, then the controller chooses among feasible operations such as Solve, Review, Repair, or Contrast.The menu can include an eligible finish option and need not contain every operation.Source: §4.2 · pp. 4–5

4.3 Task Execution and State Update

  • The selected task determines the worker's role and authorized sources; for Review or Contrast, localized evidence highlights relevant parts while preserving the full source response.A cumulative communication budget counts transmitted source tokens, including repeated copies, to constrain context-processing overhead.Source: §4.3 · p. 5
  • Each worker response is archived with its task, role, sources, output identifier, and checks, while the communication graph links it to the source responses used.Attempted checks are recorded to avoid repetition but do not certify inspected content as verified.Source: §4.3 · p. 5

4.4 Adaptive Stopping and Answer Delivery Rule

  • When eligible, the controller can stop collaboration, and a deterministic delivery rule returns an archived response without an extra LLM call to generate or evaluate a final answer.Execution also stops at Tmax, an enforced limit rather than an additional controller decision.Source: §4.4 · pp. 5–6
  • For multiple-choice and numerical questions, an explicit two-source Contrast can make its latest parsed answer take precedence; otherwise delivery uses the most frequent nonempty answer key, breaking ties by earliest occurrence.For code, delivery favors a continuous usable implementation, and public interface checks do not establish hidden-test correctness.Source: §4.4 · pp. 5–6

5 Experiments

The experiments compare S1-MAS with single-model and multi-agent baselines across reasoning, code-generation, and Web question-answering tasks. S1-MAS achieves the highest mean accuracy on six main benchmarks while reducing token use and latency, and ablations and WebSRC results examine its coordination and evidence-localization components.

5.1 Experimental Setup

  • The evaluation covers seven benchmarks spanning knowledge and mathematical reasoning, code generation, and Web-based question answering, with comparisons against single-model, fixed-topology, and MAS baselines.Comparisons use shared task-solving models, questions, response formats, and scoring procedures; three-seed means and GPT token counts and latency are reported.Source: §5.1 · p. 6, §5 · p. 6

5.2 Main Results

  • 93.97% macroaverage accuracy is highest across six benchmarks, exceeding DyLAN by 0.85 and AgentVerse by 1.54 percentage points.S1-MAS also scores 91.94% on MMLU versus DyLAN’s 91.07%, 87.07% on AQuA versus MAD-M2(S)’s 86.60%, and 91.46% on HumanEval versus MACNET’s 90.36%.Source: §5.2 · pp. 6–7, Table 1 · p. 7
  • 78.3–94.0% fewer GPT tokens per question are used by S1-MAS than by the five MAS baselines across the six main benchmarks.It uses fewer input and generation tokens in each benchmark; the reported savings accompany accuracy gains.Source: §5.2 · p. 7, Table 2 · p. 8
  • 16.8 seconds per question is S1-MAS’s mean latency across the six main benchmarks, 58.2–91.7% below the five MAS baselines.The comparison includes coordination and evidence-localization time; adaptive stopping and delivery of an archived response avoid some sequential GPT generation and an additional judging or synthesis call.Source: §5.2 · p. 7, Figure 3 · p. 7

5.3 Ablation Study

  • 82.34% average accuracy under random routing is below S1-MAS’s 84.17%, despite similar GPT token consumption and decreases on all three ablation datasets.The comparison uses the same task-construction rules and execution framework; the controller selects inspection queries and then uses the resulting evidence to select a task or stop.Source: §5.3 · p. 8
  • 75.5% fewer GPT tokens are used by S1-MAS than by the GPT-routing variant, with 84.17% versus 83.14% average accuracy.The comparison reports 0.54 million GPT tokens for S1-MAS and 2.19 million for GPT routing.Source: §5.3 · p. 8
  • 1.76 percentage points higher average accuracy comes from retaining evidence localization and highlighting, with 11.1% more GPT tokens than removing the reader.The reader localizes evidence for subsequent task selection, and highlighting directs workers to relevant content in source responses.Source: §5.3 · p. 8

5.4 Effectiveness on Web-based Question Answering

  • On WebSRC, S1-MAS has the highest mean exact-match accuracy and lowest latency, using 51.2–85.0% fewer GPT-4o tokens than AgentVerse, DyLAN, and SelfOrg.The evaluation supplies each method with the same question and webpage HTML, without visual input.Source: §5.4 · p. 8, Table 4 · p. 8

6 Conclusion

S1-MAS decouples adaptive coordination from generative task reasoning by treating collaboration choices as bounded decisions informed by state and evidence. The conclusion reports reduced GPT token consumption with strong task performance and identifies multi-page Web retrieval and question answering as future work.

  • S1-MAS substantially reduces GPT token consumption while maintaining strong task performance, and routing ablations show System One achieves higher accuracy than random and GPT-based routing.Source: §6 · p. 9
  • The framework treats coordination and information exchange as bounded choices informed by current state and evidence, reserving powerful LLM computation for substantive reasoning.Source: §6 · pp. 8–9
  • The authors identify extending the approach to multi-page Web retrieval and question answering as future work.Source: §6 · p. 9
Loading 2610.08155v1…