Keywords
LLMs, Sustainable AI, Multi-Agent Systems, NLP

Technologies
Python, PyTorch, HuggingFace, Streamlit, CUDA

Modern AI has become increasingly dependent on large language models that demand substantial computational resources, limiting accessibility and raising environmental concerns. This project investigates whether orchestrating multiple small language models (SLMs) can close the performance gap with larger systems, without the associated resource cost.

Two collaborative strategies were developed. The first, a multi-model voting framework, had three SLMs independently generate solutions before collectively selecting the best. The second, a cross-reflection pipeline, assigned each model a distinct role: planning, refinement, and implementation.

Voting consistently underperformed the strongest individual model, DeepSeek Coder, which achieved 56% accuracy. A key finding was a profound positional bias: models systematically avoided middle-positioned candidates regardless of code quality, receiving just 12% of votes compared to 41% and 47% for first and last positions respectively. The cross-reflection pipeline achieved 48% overall, but outperformed all individual models on harder tasks, improving from 30% to 40% on both medium and hard difficulty levels.