Quanta

This repository contains the code for the paper "Neural Dynamics as the Composition of Quantized Units". We experiment with a discrete approximation of learning dynamics in which reusable computations are acquired in an ordered, input-selective way. It provides the models, controlled Hierarchical Sparse Parity experiments, NumberNaming Q-discovery pipeline, and Q-alignment experiments used to test loss decomposition, resource scaling, executable recovery, and steering while keeping experiment logic and saved-artifact contracts explicit.

Repository structure

• quanta/ contains reusable datasets, models, Q-programs, experiment code, metrics, and plotting utilities.
• configs/ contains one maintained configuration for each config-driven replication.
• scripts/ contains portable training and analysis entry points for the main experiments.
• tests/ checks configuration, training, discovery, compilation, and analysis behavior.
• .experiments/ and .artifacts/ are ignored output directories created at runtime.

Main replications

Install the environment and run the tests:

uv sync
uv run python -m pytest

Run the compact loss-decomposition control through the config CLI, or reproduce the composed-task panel:

uv run main.py scaling_laws/cnand/loss_decomposition
uv run python -m scripts.reproduce_loss_decomposition

Run every point in the parameter- and data-scaling panels. These are long GPU campaigns; each script also accepts one --alpha and, for parameter scaling, one --width.

uv run python -m scripts.reproduce_parameter_scaling --all
uv run python -m scripts.reproduce_data_scaling --all

Run Q-discovery and its executable Q-model, then compile a program-derived Q-model and run the matched Q-alignment/control pair:

uv run python -m scripts.reproduce_q_discovery --device cuda
uv run python -m scripts.reproduce_q_steering --device cuda

Outputs, including resolved configurations, status files, summaries, trajectories, and checkpoints, are saved under .experiments/.

Paper

Jacopo Minniti, Aravinth Kulanthaivelu, Richard Sproat. 2026. "Neural Dynamics as the Composition of Quantized Units". https://arxiv.org/abs/2609.32487.