Proprioceptive AI, a startup founded by Logan Matthew Napolitano, has developed lightweight neural probes capable of detecting and suppressing hallucinations in large language models in real time without retraining the underlying model. The company claims an 85.8% reduction in confident-wrong outputs, those incorrect responses stated with full conviction. Each probe contains approximately 1 million parameters and adds just 0.003% overhead to the host model with sub-millisecond latency. Internal benchmarks show detection scores between 0.96 and 0.999 AUC and a 70 percentage point improvement on GSM8K tasks. The startup is developing a commercial product called “The Cradle” for models in the 3B to 32B parameter range and has filed over 100 provisional patents.
Source: Read the original article
