Research map / Multi-agent systems and games / LLM agents
RECoRD: A Multi-Agent LLM Framework for Reverse Engineering Codebase to Relational Diagram
Yuan Xue, Xiaoyu Lu, Yunfei Bai, Yunan Liu, Hoiyi Ng
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RECoRD, a multi-agent LLM framework that reverse engineers codebases into relational diagrams to expose interpretable causal structure of algorithms.
Problem
Reverse engineering complex algorithms in industrial codebases into interpretable causal structures; the potential of LLMs for this task is unexplored.
Approach
A multi-agent LLM framework (RECoRD) that converts code into relational diagrams capturing the algorithm's logical and causal structure, leveraging LLM code understanding.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply DPO reinforcement fine-tuning with LoRA (ablation) | Phi-4 · MiniSCOT (full entity list) | relation F1: 0.690 -> 0.968 | Improved |
| apply DPO reinforcement fine-tuning with LoRA (ablation) | Qwen-7B-Instruct · Newsvendor (full entity list) | relation F1: 0.810 -> 0.895 | Improved |
| apply DPO reinforcement fine-tuning with LoRA (ablation) | Phi-4 · Newsvendor (LLM+H+AST entities) | relation F1: 0.647 -> 0.647 | No clear effect |
| apply DPO reinforcement fine-tuning with LoRA (ablation) | Phi-4 · Black-Scholes (full entity list) | relation F1: 0.556 -> 0.562 | No clear effect |
| apply DPO reinforcement fine-tuning with LoRA (ablation) | Qwen-14B-Instruct · Black-Scholes (LLM+H+AST entities) | relation F1: 0.483 -> 0.569 | Improved |
| fine-tune on News Vendor then transfer (ablation) | Qwen2.5-7B-Instruct · MiniSCOT (full entity list) | relation F1: 0.786 -> 0.857 | Improved |
| fine-tune on News Vendor then transfer (ablation) | Phi-4 · MiniSCOT (LLM+H entities) | relation F1: 0.121 -> 0.250 | Improved |
| use noisier LLM+H entity list instead of full ground-truth entity list (ablation) | RECoRD RFT · MiniSCOT | relation F1: 0.968 -> 0.467 | Hurt |
| add human filtering to AST+code entity extraction (ablation) | entity extraction agent · MiniSCOT | entity precision: 0.56 -> 0.83 | Improved |
| add human filtering to AST+code entity extraction (ablation) | entity extraction agent · Newsvendor | entity precision and recall: 0.89/0.73 -> 0.89/0.73 | No clear effect |
| add AST guidance to code-only entity extraction (ablation) | entity extraction agent · Black-Scholes | entity recall: 0.21 -> 0.58 | Improved |
| use AST guidance instead of code-only entity extraction (ablation) | entity extraction agent · Newsvendor | entity precision: 1.00 -> 0.89 | Hurt |
Closest papers by meaning
- Code Graph Model (CGM): A Graph-Integrated Large Language Model for Repository-Level Software Engineering Tasks (NeurIPS 2025 · similarity 0.86)
- RepoMirage: Probing Repository Context Reasoning in Code Agents with Perturbations (NeurIPS 2026 · similarity 0.85)
- RepoGraph: Enhancing AI Software Engineering with Repository-level Code Graph (ICLR 2025 · similarity 0.85)
- Nemotron-CORTEXA: Enhancing LLM Agents for Software Engineering Tasks via Improved Localization and Solution Diversity (ICML 2025 · similarity 0.85)
- DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems (ICLR 2026 · similarity 0.84)
- From Patches to Trajectories: Privileged Process Supervision for Software-Engineering Agents (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
