Multi-Agent Reinforcement Learning for Interactive Code Debugging with Human Feedback and Memory

Anjana Krishnamoorthy, Kartik Ivatury, Benyamin Ahmadnia


Abstract
This paper introduces an interactive Python debugging framework that combines multi-agent reinforcement learning, Natural Language Processing (NLP), and long-term memory. Two Proximal Policy Optimization (PPO) agents specialize in syntax and logic errors, generating candidate fixes that developers can accept, reject, or refine. A BERT-based module encodes natural language feedback into dense embeddings and quality scores, which shape reward signals for Reinforcement Learning from Human Feedback (RLHF). To support personalization, the system uses dual FAISS indices to retrieve past fixes based on code-error pairs and developer explanations. Evaluated on a synthetic dataset of 200 Python programs, our approach achieves an 88% syntax-fix rate and 45% logic-fix rate within five suggestions—outperforming one-shot Large Language Model (LLM) baselines. In addition, the system improves the quality of the explanation, as measured by BLEU, ROUGE, and CodeBLEU. By integrating multi-agent specialization, linguistic feedback, and memory-driven retrieval, our framework delivers a more efficient, adaptive, and developer-aligned debugging experience.
Anthology ID:
2025.ranlp-1.70
Volume:
Proceedings of the 15th International Conference on Recent Advances in Natural Language Processing - Natural Language Processing in the Generative AI Era
Month:
September
Year:
2025
Address:
Varna, Bulgaria
Editors:
Galia Angelova, Maria Kunilovskaya, Marie Escribe, Ruslan Mitkov
Venue:
RANLP
SIG:
Publisher:
INCOMA Ltd., Shoumen, Bulgaria
Note:
Pages:
595–603
Language:
URL:
https://aclanthology.org/2025.ranlp-1.70/
DOI:
Bibkey:
Cite (ACL):
Anjana Krishnamoorthy, Kartik Ivatury, and Benyamin Ahmadnia. 2025. Multi-Agent Reinforcement Learning for Interactive Code Debugging with Human Feedback and Memory. In Proceedings of the 15th International Conference on Recent Advances in Natural Language Processing - Natural Language Processing in the Generative AI Era, pages 595–603, Varna, Bulgaria. INCOMA Ltd., Shoumen, Bulgaria.
Cite (Informal):
Multi-Agent Reinforcement Learning for Interactive Code Debugging with Human Feedback and Memory (Krishnamoorthy et al., RANLP 2025)
Copy Citation:
PDF:
https://aclanthology.org/2025.ranlp-1.70.pdf