We present TACO, a family of multi-modal large action models designed to improve performance on complex questions that require multiple capabilities and demand multi-step solutions. 

To address the challenges in generating multimodal instruction data, we developed ProVision, a scalable, programmatic framework that employs scene graphs and human-written programs to systematically synthesize vision-centric instruction data. 

From new open source models to evaluation frameworks, our AI Research team has been moving the needle in AI. Take a look at some of our 2024 highlights.

Discover how implementing effective data governance for AI can ensure security, compliance, and smarter decision-making.

We trust generative AI algorithms with important decisions, but know little about how they work.

Salesforce’s trusted AI architecture for red teaming leverages automation to scale ethical AI testing, utilizing a tool called fuzzai to simulate diverse adversarial scenarios and enhance model robustness. By automating adversarial prompt generation and response validation, fuzzai helps secure AI interactions while reducing human exposure to harmful content.