Token-Efficient Multi-Agent Collaborative Reasoning via Selective Verification and Memory Pruning
Keywords:
multi-agent systems, LLM reasoning, token efficiencyAbstract
Multi-agent large language model systems have demonstrated remarkable capability in solving complex, multi-step tasks through collaborative reasoning. However, their practical deployment is hindered by escalating token consumption and latency-particularly from continuous verification cycles and unbounded growth of shared memory. Building directly on the established Planner-Executor-Verifier architecture, this work introduces a token-efficient framework that preserves functional integrity while fundamentally rethinking when and what to verify. We propose an adaptive verification trigger that dynamically suppresses redundant verifier invocations based on execution confidence thresholds derived from the original system's internal state logic. Concurrently, we design a memory compression unit that selectively prunes semantically overlapping typed records from the shared memory graph without degrading dependency resolution. Empirical evaluation shows our approach retains 95 percent of the base system's task completion accuracy and error reduction efficacy, while cutting total API token usage by thirty-five to fifty percent and reducing end-to-end latency by over forty percent. This advances the Pareto frontier between reliability and efficiency in agent-based reasoning systems.References
1. W. Zhang, T. Liao, and Y. Feng, “Multi-Agent LLM Collaborative Reasoning and Task Planning for Complex Task Solving,” Research Square Preprint, doi: 10.21203/rs.3.rs-10232223/v1, 2026.
2. X. Yi, Z. Zhou, C. Cao, Q. Niu, T. Liu, and B. Han, “From debate to equilibrium: Belief-driven multi-agent LLM reasoning via Bayesian Nash equilibrium,” arXiv preprint arXiv:2506.08292, 2025.
3. E. Y. Chang and E. Y. Chang, “Multi-Agent Collaborative Intelligence: Dual-Dial Control for Reliable LLM Reasoning,” arXiv preprint arXiv:2510.04488, 2025.
4. Q. Wang, Z. Tang, Z. Jiang, N. Chen, T. Wang, and B. He, “Agenttaxo: Dissecting and benchmarking token distribution of LLM multi-agent systems,” in Proc. ICLR 2025 Workshop on Foundation Models in the Wild, 2025.
5. C. Zhang, J. Huang, H. Yu, W. Yin, X. Liang, and X. Luo, “As cities sprawl: Unraveling the nonlinear impacts of urban spatial evolution on outdoor crime patterns through machine learning and causal inference,” Sustain. Cities Soc., Art. no. 107475, 2026.
6. Z. Gu, X. Zhu, Y. Cai, H. Shen, X. Chen, Q. Wang, et al., “Agentgroupchat-v2: Divide-and-conquer is what LLM-based multi-agent system need,” arXiv preprint arXiv:2506.15451, 2025.
7. H. Chen, R. Zhao, X. Wu, and E. C. Ngai, “MACRO-LLM: LLM-Empowered Multi-Agent Collaborative Reasoning under Spatiotemporal Partial Observability,” arXiv preprint arXiv:2601.09295, 2026.
8. F. Haji, M. Bethany, M. Tabar, J. Chiang, A. Rios, and P. Najafirad, “Improving LLM reasoning with multi-agent Tree-of-Thought Validator agent,” arXiv preprint arXiv:2409.11527, 2024.
9. X. Wang, J. Wang, Y. Wang, P. Dang, S. Cao, and C. Zhang, “MARS: Toward more efficient multi-agent collaboration for LLM reasoning,” arXiv preprint arXiv:2509.20502, 2025.
10. D. Tran and D. Kiela, “Single-agent LLMs outperform multi-agent systems on multi-hop reasoning under equal thinking token budgets,” arXiv preprint arXiv:2604.02460, 2026.
11. B. Yang, X. He, H. Gao, Y. Cao, X. Li, and D. Hsu, “Codeagents: A token-efficient framework for codified multi-agent reasoning in LLMs,” arXiv preprint arXiv:2507.03254, 2025.
12. K. T. Tran, D. Dao, M. D. Nguyen, Q. V. Pham, B. O’Sullivan, and H. D. Nguyen, “Multi-agent collaboration mechanisms: A survey of LLMs,” arXiv preprint arXiv:2501.06322, 2025.
13. Z. Ke, F. Jiao, Y. Ming, X. P. Nguyen, A. Xu, D. X. Long, et al., “A survey of frontiers in LLM reasoning: Inference scaling, learning to reason, and agentic systems,” arXiv preprint arXiv:2504.09037, 2025.
14. S. Yuan, “Data Flow Mechanisms and Model Applications in Intelligent Business Operation Platforms,” Financial Economics Insights, vol. 2, no. 1, pp. 144–151, Dec. 2025, doi: 10.70088/m66tbm53.
