ATP-Bench: Towards Agentic Tool Planning for MLLM Interleaved Generation
The episode discusses 'ATP-Bench,' a paper testing MLLMs' ability to plan tool usage for complex, real-world queries. Hosts note that current models struggle with coherent planning, proposing Agentic Tool Planning as a fix. They introduce Multi-Agent MLLM-as-a-Judge (MAM) for evaluating the model's intent and planning ability.