Table 1.

  LLM Inference and Decoding Hyperparameters.

Hyper parameter	Description	Default Value	LLM Role / Effect
	Number of LLM agents in the debate	3	Total agents (mixture of cooperative + adversarial)
	Number of adversarial LLM agents	1	Controls adversarial presence inside the group
	Number of debate rounds per question	3	Depth of iterative LLM interaction and persuasion
	Number of repetitions of each configuration	1	Re-runs to estimate robustness and variance
	Number of adversarial candidate arguments generated for scoring	10	Number of alternative arguments sampled from the adversarial model for selection
	GPU device index for HuggingFace model inference	1	Hardware assignment for non-OpenAI LLMs
	Number of parallel completions for argument generation/selection	1	Used to generate multiple argument candidates via OpenAI models
	Device allocation strategy for HF models	“auto”	Automatically maps LLM weights across available GPUs/CPU
	Maximum number of new tokens generated by HF models	1000	Controls length of agent responses and debate arguments
	Whether HuggingFace model uses stochastic sampling	True	Enables non-deterministic sampling for diverse arguments and responses[42]
	Sampling temperature for HF decoding	0.6	Controls randomness—higher = more diverse generations
	Nucleus sampling probability cutoff	0.9	Restricts sampling to top-p portion of probability mass