Contrastive-Aligned Knowledge Distillation for Collaborative Code Completion via Multi-Agent Reinforcement Learning

ICLR 2026 Conference Submission25606 Authors

Keywords: Contrastive-Aligned Knowledge

Abstract: We introduce a novel multi-agent reinforcement learning (MARL) framework for code completion in a collaborative manner, and address the important issue for successful collaboration in code completion: balancing semantic alignment and specialized expertise among the agents. The proposed method incorporates Contrastive Alignment Module (CAM) and Distilled Knowledge Transfer (DKT) mechanism, which allows agents to share coherent representations without losing domain-specific knowledge. CAM embeddings between agents might be aligned through a contrastive learning goal and would create a coordinate measurement of the space in which all embeddings agree (without homogenizing individual capabilities), but DKT would dynamically distil some knowledge from a high-performing teacher agent to others using a regularized KL-divergence goal.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 25606