The autonomous driving company said its new fact-based cognitive model cuts token costs by up to 98% and turns operational driving data into verified learning signals.
WeRide unveiled WITT, short for World Intelligence Toward Truth, as a Physical AI Cognitive Foundation Model built to convert real-world operational data into what it describes as trusted facts for AI training, evaluation and iteration. The model introduces Atomic Physical Facts, or APFs, as the smallest verifiable units of information about the physical world, and uses four functions—Fact Extraction, Fact Reasoning, Fact Verification and Fact Curation—to turn video, images and text into structured, validated learning signals. The company said WITT is designed to address a core problem in autonomous driving: large volumes of noisy real-world data, scarce long-tail scenarios (rare but important edge cases), and inconsistent scene interpretation by general-purpose AI models. Within WeRide’s architecture, WITT works alongside its GENESIS simulation model to form a feedback loop in which operational data are converted into facts and then into simulation and training inputs. WeRide said WITT can process up to 10,000 minutes of vehicle-operation video per day on a single GPU (graphics processing unit), reduce token costs by up to 98%, and deliver up to 200 times greater data-processing efficiency than larger general-purpose AI models in comparable workloads. It also said WITT’s average factual error rate in autonomous driving scenario understanding tasks is about one-third that of leading general-purpose AI models. The company framed the launch as part of a broader push to build cognitive foundation models grounded in real-world validation as Physical AI moves into larger-scale deployment.