SwarmSys: Decentralized Swarm-Inspired Agents for Scalable and Adaptive Reasoning

1 Introduction

• Fixed roles and static communication limit adaptability, contributing to redundant exploration and inefficiency in long-horizon or dynamic tasks.Recent adaptive approaches are also described as limited by centralized orchestration or manually designed topologies, restricting scalability and long-term stability.Source: §1 · p. 1
• SwarmSys coordinates agents through lightweight pheromone-like traces that encode contextual utility, balancing exploration and convergence without centralized control.Its Explorers, Workers, and Validators iteratively explore hypotheses, execute subtasks, and check consistency; adaptive profiles and reinforcement support reallocation and convergence.Source: §1 · p. 2
• Up to 10.7% higher accuracy and 9.9% better sub-task correctness are reported for SwarmSys against baselines such as GPTSwarm.The paper also reports that GPT-4o-based agents approach GPT-5 performance.Source: §1 · p. 2
• Qualitative analyses report emergent behaviors, knowledge diffusion, specialization balance, and self-regularization as hallmarks of collective intelligence.The paper presents its framework, adaptive coordination mechanism, and evaluation across diverse reasoning tasks as three contributions.Source: §1 · p. 2

2.1 Overview of SwarmSys

• Tasks proceed through iterative matching, collaboration, and profile updates, with new event profiles used to retrieve candidate agents before collaborative rounds.Explorers propose decompositions, Workers execute subtasks after consensus, and Validators verify intermediate results; updated profiles feed subsequent iterations.Source: §2.1 · pp. 2–3, Figure 3 · p. 3

2.2 Profiles as Adaptive Memory Units

• Agent profiles track roles, abilities, workload, and performance history, while event profiles record task descriptions, dependencies, progress, and participating agents.Embeddings and workload indicators refresh after each round, and historical success or failure informs future matching; together, profiles form distributed memory.Source: §2.2 · p. 3

2.3 Embedding-Based Matching with Exploration–Exploitation Dynamics

• Embedding-based matching balances expertise, workload, and exploration by comparing agent and event representations and adapting each agent’s exploration rate to recent performance.Agent representations combine competence and availability; events encode descriptions, dependencies, progress, and milestones. Normalized cosine similarity estimates compatibility, while an ε-greedy policy explores or exploits matches.Source: §2.3 · p. 3

2.4 Pheromone-Inspired Optimization

• Validated contributions reinforce agent–event compatibility, while idle or invalid matches receive no reinforcement and gradually lose competitiveness as profiles evolve.Exploration preserves diversity and helps prevent deadlock, while exploitation prioritizes promising matches; bounded probabilities support stability.Source: §2.4 · p. 4

3.1 Experiment Setting

• Evaluation spans symbolic, research-synthesis, and scientific-programming tasks, using dataset-specific metrics and category-specific baseline selections.The four benchmarks differ in domain focus, reasoning type, and data format.Source: §3.1 · p. 4, Table 1 · p. 4

3.2 Overall Results

• +12.5% Accuracy and +10.8% Coverage over GPTSwarm on exam-style reasoning; SwarmSys-8 narrows the GPT-5 gap by over 70%.Table 2 covers single-subject, multi-subject, and Olympic-level exams.Source: §3.2 · p. 4, Table 2 · p. 5
• +2.3% overall RACE score and +3.7% instruction-following over Grok Deeper Search on research-level reasoning.The reported gains are on DeepResearch Bench.Source: §3.2 · p. 5, Table 3 · p. 5
• +2.5% Pass@Main and +11.9% Pass@Sub over the best GPT-4o baselines on SciCode.The authors report a larger Pass@Sub gain and connect it to modular code generation and localized validation.Source: §3.2 · p. 5, Table 4 · p. 5

3.3 Ablation Study

• Removing roles drops accuracy from 56.3% to 43.2% and coverage from 52.4% to 41.4% at A=4; adaptive matching raises coverage by up to +27.3%.The ablations also report gains plateauing beyond 14 Workers.Source: §3.3 · p. 5

3.4 Swarm Effect: Emergent Collective Intelligence

• On the Exam task, Accuracy and Knowledge Coverage rise from 74.3/79.7 at A=4 to 77.3/81.0 at A=14, then gains stay within < 1% at A=20 or A=32.The reported plateau follows sufficient coverage of subtask diversity and knowledge space.Source: §3.4 · p. 6, Table 5 · p. 6
• Role-based debate and validation yield higher reasoning-path diversity and smoother convergence, as the communication topology evolves from hub–spoke to a small-world mesh.Explorers propose subgoals, Workers attempt partial solutions, and Validators consolidate outcomes.Source: §3.4 · p. 6, Figure 4 · p. 6, Figure 5 · p. 7
• Agents reuse, revise, or correct content discovered by others, while consensus dilutes weaker agents’ errors and supports a distributed memory and decision space.The paper describes these as emergent knowledge diffusion and self-regularization.Source: §3.4 · p. 6

4.1 Agent Behavior

• Agent competence embeddings shift by 0.14 ± 0.03 cosine per round; Explorers vary most, while Validators remain stable, supporting specialization through distributed memory.The analysis uses experiment event logs to examine profile adaptation, interaction topology, and contribution balance.Source: §4.1 · p. 7
• Communication evolves from a validator-centered hub–spoke pattern to a distributed-consensus small-world structure, with local clustering rising from 0.28 to 0.47.The authors attribute denser cross-links to pheromone reinforcement and describe the transition as emerging without explicit central control.Source: §4.1 · p. 7
• Accepted contributions have normalized entropy Hc=0.72, versus GPTSwarm’s 0.41, indicating more balanced participation under the exploration–exploitation policy.The measure uses each agent’s share of accepted contributions, computed from its contribution acceptance rate.Source: §4.1 · p. 7

4.2 Error Analysis and Case Study

• SwarmSys sometimes fails on strict temporal or symbolic alignment, with five identified failure types including premature convergence, reinforcement bias, mode collapse, constraint omission, and communication deadlock.In one case, early validator acceptance reinforced one solution branch and suppressed valid alternatives, producing overfitting on partial evidence.Source: §4.2 · p. 7

LLM-based Multi-Agent Systems

• CAMEL (Li et al., 2023) and AutoGen (Wu et al., 2023) use explicit roles and dialogue, while SwarmSys coordinates through role-guided interaction and adaptive matching without a global controller.The passage characterizes centralized orchestration and fixed pipelines in prior systems as limiting scalability and adaptability.Source: §5 · p. 8

Collaborative Reasoning and Debate

• CoT and ToT (Yao et al., 2023; Besta et al., 2024), MAD (Liang et al., 2024), Longagent (Zhao et al., 2024), and ROBIN (Ghareeb et al., 2025) use structured deliberation or debate, whereas SwarmSys uses debate locally within a self-organizing swarm.The passage says these prior approaches generally use fixed communication topologies and lack dynamic adaptation or memory persistence across reasoning episodes.Source: §5 · p. 8

Adaptive Coordination and Swarm-inspired Reasoning

• GPTSwarm (Zhuge et al., 2024) uses graph optimization and stigmergic feedback, while SwarmSys unifies dynamic profiling, embedding-based matching, and pheromone-inspired reinforcement for multi-event reasoning.The passage also contrasts SwarmSys with profile-like adaptation and iterative self-improvement systems described as task-specific.Source: §5 · p. 8

6 Conclusion

• SwarmSys combines role-specialized collaboration, dynamic profiling, and pheromone-inspired reinforcement to coordinate without centralized control.The conclusion describes the framework as decentralized and swarm-intelligence-inspired.Source: §6 · p. 8
• Across diverse reasoning and research domains, SwarmSys consistently outperforms strong baselines and exhibits emergent collective behaviors.Source: §6 · p. 8
• The authors conclude that scaling coordination can rival scaling model size.Source: §6 · p. 8

Limitations

• Decentralized coordination increases communication overhead, which may reduce efficiency in latency-sensitive settings.Source: Limitations · pp. 8–9
• Agent profiling relies on text-based embeddings and heuristic updates; learnable or gradient-based mechanisms are proposed for more precise skill modeling.Source: Limitations · p. 9
• Experiments focus primarily on reasoning and research-oriented tasks, leaving embodied or real-time interactive environments as open directions.Source: Limitations · p. 9

• Experiments use public or synthetic datasets and involve no human subjects, personal data, or sensitive content.Source: Ethics Statement · p. 9
• The authors identify bias propagation and unreliable autonomous coordination as potential ethical issues in LLM-based multi-agent systems.Source: Ethics Statement · p. 9
• The authors advocate human oversight, fairness, and accountability in future applications.Source: Ethics Statement · p. 9

• Workers solve assigned subproblems with transparent justification, share solutions, and defend or revise them during structured debate.Source: §A.1 · p. 11
• Validators check workers’ solutions for logic, accuracy, and completeness, reconcile discrepancies, and confirm group consensus.Source: §A.1 · p. 11
• Explorers analyze and decompose problems, identify solution strategies, and may shift toward higher-pressure roles.Source: §A.1 · p. 11
• The general prompt classifies tasks by reasoning type and requires ordered explorer–worker–validator rounds that end with a validator-announced final answer.It calls for multiple solution paths, comparison, and resolution of disagreements through structured logical debate.Source: §A.1 · p. 12

A.2 Dataset Settings

• Exam-style benchmarks cover single- and multi-subject settings, while DeepResearch Bench evaluates research synthesis and SciCode evaluates scientific programming.Exam tasks comprise interrelated and unrelated subtasks; SciCode problems are decomposed into subproblems.Source: §A.2 · pp. 12–13

A.4.2 Embedding and Matching

• Agent representations combine competence information with availability reflecting workload and readiness.Competence draws on declared abilities and historical performance; the final representation is their sum.Source: §A.4.2 · p. 13
• Event embeddings integrate task descriptions, dependencies, progress, and milestone metadata in a shared space with agent embeddings.Source: §A.4.2 · p. 14
• Agent–event compatibility uses normalized cosine similarity, while a dynamic ε-greedy policy balances exploration and exploitation based on recent success.Exploration samples potential matches in proportion to similarity; exploitation emphasizes high-compatibility matches.Source: §A.4.2 · p. 14
• The matching mechanism is described as supporting adaptability through evolving embeddings, stability through probabilistic sampling, and robustness through exploration–exploitation balance.Source: §A.4.2 · p. 14

A.5 Cost

• Table 8 compares average per-question model costs for baselines and systems used in the main experiments.Source: Table 8 · p. 14