Collaborating Agents with Large Language Models for Question-Answering Tasks

Multi-agent debate has emerged as an effective strategy for improving the factual accuracy and reasoning quality of large language models, yet most prior work relies on heavyweight backbones or extensive deliberation budgets that hinder real-time use. Our paper introduces a lightweight inter-agent collab-orating workflow - Proponent, Opponent, and Arbitrator - that injects adversarial critique without inflating context length or model size. We instantiate each role with two compact reasoning-centric backbones, Qwen-2.5 (0.5 B) and DeepSeek-Rl (1.5 B), both of which provide strong mathematical skills. Evaluated on three rigorous multiple-choice math benchmarks - TAL-SCQ5K-EN, MMLU (MATH), and Mathqa - our framework lifts top accuracy and delivers absolute gains over a single-agent reasoning baseline and over a vanilla LLM baseline while incurring only a 1.8 x inference-time overhead. Ablation studies confirm that architectural collaboration, rather than parameter count, drives these improvements and that the fixed 1.2 k-token budget keeps the system practical for small models. The results suggest that targeted adversarial debate combined with lightweight arbitration is an effective path toward adaptive role specialisation and retrieval-augmented reasoning in future work.