NEURAL LEARNING RESEARCH
Logic.
Small, inspectable experiments in reasoning, numerical memory, and how a model learns from experience.
Research in progressLogic is a collection of research experiments that turn broad questions about learning into small systems that can be tested. The work spans neural proof-step verification, changes to network structure, learned memory, and online control.
Several experiments ask what a frozen network can do by writing and reading numerical memory. Others examine how a model assigns credit to earlier actions or responds when unfamiliar controls rotate, reverse, or change scale.
Each experiment has its own assumptions and limits. Reproduction instructions, comparison conditions, and recorded failures matter as much as the model itself; these are controlled investigations rather than a claim of general reasoning.
A closer look
Reasoning in small steps
Neural proof-step experiments examine individual rule applications using explicit training and evaluation labels.
Memory as a working space
Learned writers and readers explore how numerical records can preserve useful experience.
Questions with controls
Matched comparisons and interventions help distinguish learned behavior from supplied structure.