Cross-Modal Syntax-NL Attention for Multi-Agent Reinforcement Learning in Collaborative Coding

ICLR 2026 Conference Submission25586 Authors

Keywords: Syntax-NL Attention

Abstract: We suggest a new communication protocol for multi-agent reinforcement learning (MARL) in collaborative coding where agents have to coordinate in coordinate (both structured code syntax and natural language (NL) messages). Conventional ways to treat these modalities separately, the result was suboptimal alignment between the communicational and the code semantic. The proposed method introduces a cross-modal attention framework that is able to dynamically bridge abstract syntax trees (ASTs) of code and NL messages in a jointly learned embedding space. A graph neural net encodes artistic of syntactic elements of code while a pretrained Transformer processes NL messages which are then aligned in the direction of weakly supervised contrastive learning making use of implicit training sign for execution outcome of code to guide the alignment without requirement of manual annotation. Also, the framework uses syntax-aware attention gates to select which message tokens are relevant to particular code nodes, which can result in more precise coordination during collaborative tasks.

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

Submission Number: 25586