Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore Salesforce Research's groundbreaking Agentic Uncertainty Quantification (AUQ) framework in this 15-minute video breakdown that demonstrates how to operationalize Kahneman's "System 1 vs. System 2" cognitive framework directly into large language model architectures. Discover how this training-free approach addresses the critical "Spiral of Hallucination" problem plaguing AI agents in long-horizon tasks by treating verbalized confidence as a dynamic control signal rather than merely a performance metric. Learn about the implementation of a non-differentiable switching function that intelligently bifurcates inference between a fast, memory-augmented processing path and a slower, reflective "inverse calibration" loop utilizing Best-of-N sampling techniques. Examine how this architectural innovation moves beyond fragile ReAct loops toward AI agents capable of mathematically detecting epistemic uncertainty and actively budgeting computational resources to resolve ambiguities through internal self-reflection and memory systems. Analyze whether Salesforce's approach successfully creates more reliable AI agents that can doubt themselves constructively and make more informed decisions in complex, multi-step reasoning scenarios.
