Latent Reasoning Works

LatentReasoningWorks

Is that structure in your data, or in your method?

Latent Reasoning Works builds open tools for that question. Run a dataset, or a model, through many methods, score the geometry of every result with one suite, and keep the record of how each was reached. A representation is good when it survives a change of method, and that test should be something anyone can run.

uv tool install --python 3.12 manyruns

The thesis

A UMAP, a PHATE, a foundation-model embedding, a transformer's hidden layer: each shows you a shape, and part of that shape belongs to the method. The picture is what gets published, so that part matters. We want to know which structure would be found again by a different method, a different instrument, someone meeting the data fresh. Our tools make that comparison routine: run the data through many methods, score every result with the same suite of geometric metrics, and keep a ledger of how each result was reached. The data can be a cell-by-gene matrix, a cohort of genomes, a protein run through a sequence model, or a language model's hidden states while it answers. A representation is good when it survives a change of method. These tools make that test something you run before you believe a figure.

The codebases

Released as they are ready. Each carries the flavor of a problem worked on here.

manyruns

load, assume, run, repeat

open betastars on GitHub

The app. Run manyruns, drop a file in the data folder, and it reads the data, tells you what it sees, offers the analyses it can honestly run, executes with a live trace, and explains the result in plain English. Every run lands in one ledger with its artifacts, so a result can be retraced and run again. One command does the same from a script, and a harness sweeps every recipe over every dataset to rank which metrics tell the workflows apart.

uv tool install --python 3.12 manyruns && manyruns

Open beta. The CLI and the record schema can still change.

manylatents

one geometry, learned through many latents

openstars on GitHub

The engine. Seventeen embedding methods, from PCA and UMAP to PHATE, diffusion maps, autoencoders and neural ODEs, behind one call, and a suite of geometric metrics that scores every result the same way: trustworthiness, continuity, spectral decay, persistent homology. A sweep runs the methods on one dataset and returns the embeddings and the scores side by side.

manylatents --multirun algorithms/latent=umap,phate,tsne data=swissroll metrics=trustworthiness

Extended by manylatents-omics for single-cell, population genetics, and DNA, RNA and protein encoders.

manyagents

ask many models, keep the trace

openstars on GitHub

Run one prompt through many language models, hosted or local, and score what comes back. For local models it keeps the hidden states that produced each token, cut into reasoning steps and stored next to the text, so manylatents can measure the trajectory's velocity and curvature the way it measures a dataset.

manyagents experiment=test_wandb
Latent Reasoning Works MIT licensed