UAC-WM: An Uncertainty-Aware Multi-Agent Coordination Framework Based on World Models

UAC-WM (Uncertainty-Aware Coordination with World Models) is an innovative framework that treats multi-agent coordination as a dynamic control problem. Its core lies in using an online uncertainty estimator and a world model-driven controller to adaptively select coordination strategies based on changes in task uncertainty, enabling a paradigm shift from reasoning to interaction in code reasoning tasks.

Traditional multi-agent coordination often uses fixed strategies (fully distributed or centralized), but real-world task uncertainty changes dynamically over time, making fixed strategies difficult to adapt. UAC-WM evolved from the predecessor project MARS: MARS v1 implemented a multi-agent pipeline and calculated the Coordination Uncertainty Index (CUI) post-hoc but lacked risk components; UAC-WM v2 transforms static post-hoc diagnosis into a dynamic online controller, responding to CUI changes in real time in interactive environments such as code repair.

UAC-WM consists of four core components:

1. Explicit State Abstraction: Uses structured WorldState and Candidate representations to replace free-text states, providing a reliable foundation for coordination decisions;
2. Online Uncertainty Estimator: Calculates CUI (a scalar value ∈ [0,1]) from four dimensions: belief entropy, confidence variance, answer entropy, and validator risk;
3. Adaptive Coordination Controller: Executes actions (TERMINATE/ROLLBACK/BRANCH/MERGE/CENTRALIZE) based on a threshold strategy, with the merge threshold of 0.30 inherited from MARS empirical results;
4. World Model-Guided Validation: Integrates real test execution, static checking, rollout risk assessment, and online learning to improve result credibility.

UAC-WM uses a three-agent pipeline:

• Locator: Identifies target files that need editing;
• Patch: Generates code repair solutions (uses full file rewriting to improve the reliability of local small models);
• Validator: Applies patches, runs tests, assesses risks, and learns from feedback. Baseline comparison methods include: single (single agent), fixed_centralized (fixed centralized), fixed_peer (fixed distributed), self_consistency (self-consistency).

UAC-WM provides a complete experimental framework:

• Local Quick Validation: Includes self-contained test tasks that can run end-to-end in the Ollama environment (e.g., qwen2.5:7b);
• SWE-bench Lite Extension: Supports standard benchmarks in the code generation field, checking out repository benchmark commits and running tests;
• Trajectory Analysis: Records each round's state, uncertainty signals, coordination actions, success status, and token costs to support subsequent analysis.

Application Value:

• Automatic Code Generation: Improves the success rate of automated code repair;
• Complex Task Solving: Adaptively adjusts the balance between exploration and integration;
• Multi-Agent Research: Provides an extensible framework that supports module replacement and ablation experiments. Technical Highlights:
• Interpretability: Rule-based threshold strategy with traceable decisions;
• Modularity: Clear component division (world_model/uncertainty, etc.) for easy extension;
• Local Model Friendly: The Patch agent uses full file rewriting to adapt to resource-constrained scenarios.

UAC-WM represents an important direction in multi-agent coordination research: from fixed strategies to adaptive strategies, from post-hoc analysis to online decision-making, from pure reasoning to interactive execution. Its uncertainty-aware coordination mechanism provides new ideas for building more intelligent and reliable multi-agent systems. As the capabilities of large language models improve, this mechanism may become one of the standard components of future intelligent systems.