Post-training recipe, not model family, drives multi-LLM diversity
Fine-tuning method matters more than base model for diverse multi-agent behavior.
Multi-LLM systems rely on diverse conversational behaviors from different models to deliberate, judge, or coordinate. Prior work suggested drawing one model per family for behavioral diversity, as LLMs prefer outputs from their own family in offline tests. But this hasn't been tested in interactive multi-LLM settings—the real-world use case.
A new paper from Luyang Zhang and colleagues (arXiv:2606.20632) studied this with a massive 940,000-chain corpus and a 1.6M-chain factorial using same-base Llama models. Their validated headline metric? Hedging. A reasoning-distilled Llama checkpoint shifted by 18% depending on which same-base partner it replied to—larger than any cross-family gap. Qwen, closed-API, and runtime checks confirmed the pattern. The results identify post-training recipe as a first-class axis for multi-LLM panel composition, proving model family alone is an incomplete proxy for conversational diversity.
- Study uses a 940,000-chain corpus and 1.6M-chain same-base Llama factorial for analysis
- A reasoning-distilled Llama checkpoint showed an 18% hedging shift based on partner, exceeding cross-family differences
- Post-training recipe (fine-tuning method) is the primary driver of conversational diversity, not model family
Why It Matters
For building multi-agent systems, choose models by fine-tuning method, not just base family, for true behavioral diversity.