Adaptive Confidence Gating in Multi-Agent Collaboration for Efficient and Optimized Code Generation

TL;DR

This work addresses the challenge of high-quality code generation with small language models by introducing DebateCoder, a three-agent MAC framework that uses structured debate, adaptive confidence gating, and reviewer-guided debugging to overcome reasoning bottlenecks. The method achieves substantial accuracy gains (e.g., 70.12% Pass@1 on HumanEval) and reduces API overhead by about 35% compared to prior multi-agent approaches, demonstrating that resource-efficient models can reach competitive performance via optimized collaboration. Key contributions include the four-stage DebateCoder pipeline, the 95% gating threshold for efficiency, and the orthogonal pre-generation debate plus post-generation debugging loop that improves robustness. The findings suggest practical implications for scalable, cost-effective automated software engineering using small LMs in conjunction with carefully designed multi-agent protocols.

Abstract

While Large Language Models (LLMs) have catalyzed breakthroughs in automated code generation, Small Language Models (SLMs) often encounter reasoning bottlenecks and failure loops when addressing complex logical requirements. To overcome these challenges, we propose DebateCoder, a multi-agent collaborative framework designed to improve the reasoning ability of SLMs (e.g., Pangu-1B) in resource-constrained environments. DebateCoder uses a structured role-playing protocol with three agents: User Agent (A_UA), Technical Agent (A_TA), and Quality Assurance Agent (A_QA). It also includes an Adaptive Confidence Gating mechanism with a 95% threshold to balance accuracy and inference efficiency. In addition, we introduce a multi-turn deliberation module and a reviewer-guided analytical debugging loop for orthogonal pre-generation debate and post-generation refinement. Experiments on HumanEval and MBPP show that DebateCoder achieves 70.12% Pass@1 on HumanEval, outperforming MapCoder while reducing API overhead by about 35%. These results indicate that collaborative protocols can mitigate limitations of small-parameter models and provide a scalable, efficient approach to high-quality automated software engineering.

Figures (4)

• Figure 1: overview of debate coder
• Figure 2: Prompt for Initial Planning and Confidence Gating.
• Figure 3: Prompt for Code Reviewer.
• Figure 4: Prompt for Debugging Agent.