LLAMIA-Bench Evaluates LLM and Non-Language Agent Collaboration

LLAMIA-Bench introduces six chess-based tasks to study the communication bottleneck caused by verbalizing continuous representations from non-language agents into text for LLM orchestrators.