These experiments explore networks whose nodes predict the signals they receive and update their own states from local error information.
The NEON work studies energy functions and iterative state updates. A related JAX implementation sends probabilistic messages between nodes, including an estimate of uncertainty. Other language-model experiments add an error-optimization refinement stage and compare variants through ablations.
The common question is how much coordinated learning can emerge from local prediction and update rules. The implementations make different choices about messages, inference, and training, so they are experiments within a shared subject rather than one settled architecture.
The work has not established a general replacement for ordinary gradient-based training.