Proprioceptive AI, founded by Logan Matthew Napolitano, is developing neural probes to monitor and correct large language models in real time. The technology identifies and suppresses hallucinations without requiring model retraining or weight adjustments, keeping the base model frozen during deployment.
The probes consist of approximately 1 million parameters, adding 0.003% overhead to the host model. They operate with sub-millisecond latency. Internal benchmarks show an 85.8% reduction in confident-wrong outputs and AUC scores between 0.96 and 0.999 for detection efficacy. On GSM8K tasks, the system demonstrated a 70 percentage point improvement.
The company is developing a commercial product named The Cradle, intended for models with 3B to 32B parameters. The technology utilizes Koopman operators and has over 100 provisional patents filed. The probes have shown separation ratios between 125x and 1,376x across model families like LLaMA and Mistral.
While the company reports cross-model transfer capabilities, all current benchmarks are internal. Independent validation and testing on models exceeding the 32B parameter threshold remain necessary to confirm the effectiveness of the behavioral channels at a larger scale.
