[R] Anyone experimenting with heterogeneous (different base LLMs) multi-agent systems for open-ended scientific reasoning or hypothesis generation?

A viral discussion among AI researchers is probing whether multi-agent systems built with fundamentally different large language models (LLMs) outperform homogeneous setups for complex scientific tasks. The core question, raised on platforms like Reddit, asks if combining agents powered by distinct base models—such as OpenAI's GPT-4, Anthropic's Claude 3.5, and Meta's Llama 3—can enhance open-ended reasoning, hypothesis generation, and reduce systemic biases inherent in any single model's training data. Most current multi-agent frameworks use a single underlying model with different roles or prompts, but this new heterogeneous approach aims to leverage diverse "priors" or worldviews to simulate a more robust, committee-like deliberation process.

The technical premise is that each LLM has unique strengths, blind spots, and reasoning patterns. A system could, for instance, use Claude for cautious, constitutional reasoning, GPT-4 for creative leaps, and a specialized model like Gemini for data analysis, having them debate to reach conclusions. While anecdotal experiments exist, the field lacks rigorous papers or benchmarks comparing heterogeneous vs. homogeneous agent performance on tasks like literature review or experimental design. The community's call for evidence highlights a frontier in AI research where combining diverse AI 'minds' could accelerate scientific discovery and improve the reliability of AI-assisted research, pushing beyond the limitations of any single model's architecture.