TOBench: A Task-Oriented Omni-Modal Benchmark for Real-World Tool-Using Agents

Researchers introduced MM-ToolBench, a new benchmark for evaluating omni-modal AI agents on realistic end-to-end tasks.

It includes 100 executable tasks, 27 MCP servers, and 324 tools to test multimodal reasoning and tool use.

The harness checks whether agents can verify outputs, detect failures, and self-correct when tasks do not meet requirements.

Results show a large gap between current models and human performance in professional workflows such as customer service and content creation.