LLM-Driven Collaborative Model for Untangling Commits via Explicit and Implicit Dependency Reasoning

2025-07-14 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Software Development & Engineering, Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

ColaUntangle is a novel LLM-driven collaborative consultation framework designed to untangle software commits by explicitly modeling both explicit and implicit dependencies among code changes. It employs a multi-agent architecture where specialized agents, one for explicit dependencies (e.g., control/data flow) and another for implicit dependencies (e.g., semantic relationships), iteratively consult with a reviewer agent. The framework utilizes multi-version Program Dependency Graphs (δ-PDG) to capture structural and contextual information. Evaluated on 1,612 C# and 14k Java tangled commits, ColaUntangle achieved a 44% improvement on the C# dataset and a 100% improvement on the Java dataset over the best-performing baselines, demonstrating its effectiveness in automating commit untangling tasks. The DeepSeek-V3 model served as the reasoning engine.

Key takeaway

For software engineering teams struggling with tangled commits, adopting an LLM-driven multi-agent approach like ColaUntangle can significantly improve code quality and maintainability. You should consider integrating such a framework to automate the decomposition of complex commits, thereby streamlining code reviews and enhancing the accuracy of downstream tools like bug prediction models. This method reduces manual effort in identifying explicit and implicit code change relationships.

Key insights

ColaUntangle uses LLM-driven multi-agents to untangle commits by reasoning over explicit and implicit code dependencies.

Principles

• Atomic commits enhance code review and maintenance.
• Untangling requires distinguishing explicit and implicit dependencies.
• LLMs excel at semantic code understanding and reasoning.

Method

ColaUntangle extracts explicit and implicit contexts from δ-PDG, then uses a multi-agent system (explicit, implicit, reviewer agents) for iterative consultation to reach consensus on untangling.

In practice

• Construct multi-version Program Dependency Graphs (δ-PDG) for code change context.
• Implement a multi-agent system for complex code analysis tasks.

Topics

• Commit Untangling
• Large Language Models
• Multi-Agent Systems
• Program Dependency Graphs (δ-PDG)
• Software Engineering
• Code Review

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Software Engineer

Related on AIssential

• [2026.05 Week 1] Five Trending Repos of the Week — Robust AI agent systems require sophisticated error handling, persistent memory, and secure, efficient execution environments.
• Diagnosis Is Not Prescription: Linguistic Co-Adaptation Explains Patching Hazards in LLM Pipelines — Correcting LLM agent bottlenecks directly can degrade performance due to downstream modules' implicit adaptation to upstream errors.
• Harnessing Code Agents for Automatic Software Verification — A general LLM code agent with a verification harness can fully automate formal software proof generation.
• Long Context Isn’t Free — I Built a Safe Prompt-Pruning Layer That Makes LLM Systems Work — Deterministic, dependency-safe prompt pruning significantly reduces LLM context bloat without sacrificing critical information.
• LLM Consortium for Software Design Refinement: A Controlled Experiment on Multi-Agent Collaboration Topologies — Structural adversarial and cross-model review topologies significantly enhance multi-agent LLM software design.
• TraceDev: A Traceability-Driven Multi-agent Framework for Requirement-to-Code Development — TraceDev uses a multi-agent framework with a dynamic traceability graph to generate repository-level code from complex use cases.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.