Task Decomposition and Collaborative Reasoning in LLM-Based
Keywords:
Abstract
Autonomous decision-making systems based on large language models (LLMs) increasingly operate over tasks that require long-horizon planning, evidence integration, tool invocation, and explicit risk control. However, monolithic LLM agents often exhibit unstable task decomposition, cumulative reasoning errors, and weak calibration between local confidence and final decisions. This paper presents TDC-R, a task decomposition and collaborative reasoning framework for LLM-based autonomous decision-making systems. The framework represents a user objective as a dependency-aware task graph, routes subtasks to role-specialized reasoning agents, verifies intermediate states through a critique-and-consistency module, and fuses candidate decisions using reliability-weighted aggregation. The design is motivated by recent multi-agent LLM research, including the hierarchical collaboration method of Wang, Feng, and Fang, which emphasizes semantic communication, verifier-guided feedback, and dynamic collaboration control. We evaluate TDC-R on three decision-oriented benchmarks constructed from mathematical decision problems, multi-step planning tasks, and multi-hop evidence-based decision questions. Experiments compare TDC-R with direct prompting, chain-of-thought self-consistency, ReAct- style tool reasoning, least-to-most prompting, and debate-based multi-agent reasoning. TDC-R achieves an average decision accuracy of 82.8%, outperforming the strongest baseline by 5.9 percentage points, while reducing invalid intermediate-state errors by 22.4% relative to direct prompting. It also improves throughput-normalized utility under bounded interaction budgets. Ablation studies show that dependency-aware decomposition, semantic state routing, verifier feedback, and reliability-weighted fusion each contribute to final performance. The results suggest that effective autonomous LLM decision-making depends less on increasing the number of agents and more on structuring intermediate computation into verifiable, reusable, and cost-aware reasoning states.
References
[1] J. Wei et al., “Chain-of-thought prompting elicits reasoning in large language models,” in Proc. NeurIPS, 2022.
[2] Zhou, D. (2025, December). M-VP2: Microservice-Oriented Vulnerability Patch Planning-A Cost-Aware Approachusing Multi-Agent Reinforcement Learning. In 2025 5th International Conference on Computer, Internet of Things and Control Engineering (CITCE) (pp. 248-254). IEEE.
[3] X. Wang et al., “Self-consistency improves chain of thought reasoning in language models,” in Proc. ICLR, 2023.
[4] S. Yao et al., “ReAct: Synergizing reasoning and acting in language models,” in Proc. ICLR, 2023.
[5] S. Yao et al., “Tree of Thoughts: Deliberate problem solving with large language models,” in Proc. NeurIPS, 2023.
[6] Y. Du et al., “Improving factuality and reasoning in language models through multiagent debate,” arXiv preprint arXiv:2305.14325, 2023.
[7] G. Li et al., “CAMEL: Communicative agents for mind exploration of large language model society,” in Proc. NeurIPS, 2023.
[8] S. Hong et al., “MetaGPT: Meta programming for a multi-agent collaborative framework,” in Proc. ICLR, 2024.
[9] W. Chen et al., “AgentVerse: Facilitating multi-agent collaboration and exploring emergent behaviors,” in Proc. ICLR, 2024.
[10] Q. Wu et al., “AutoGen: Enabling next-gen LLM applications via multi-agent conversation,” in Proc. COLM, 2024.
[11] Hu, S. (2026). Research on Security Enhancement Methods for Adversarial Robust Large Language Model Intelligent Agents for Medical Decision-Making Tasks. arXiv preprint arXiv:2605.08257.
[12] Wang, S., Feng, Y., & Fang, X. (2026). A Large Language Model-Enabled Multi-Agent Collaboration Method for Complex Task Solving.
[13] K. Cobbe et al., “Training verifiers to solve math word problems,” arXiv preprint arXiv:2110.14168, 2021.
[14] K. Valmeekam et al., “PlanBench: An extensible benchmark for evaluating large language models on planning and reasoning about change,” in Proc. NeurIPS Datasets and Benchmarks, 2023.
[15] Chen, X. (2024, November). Cloud Storage User Behavior Analysis and Dynamic Replica Strategy Optimization Based on Improved RFM and Fuzzy Clustering. In International Conference on Cognitive based Information Processing and Applications (pp. 425-434). Singapore: Springer Nature Singapore.
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Journal of Engineering Systems and Digital Innovation
This work is licensed under a Creative Commons Attribution 4.0 International License.
