Research

Open research, including the experiments that didn't work.

We publish model releases, technical writeups, and negative results in one place. An Atom feed lives at /research/feed.xml.

All4Model releases2Technical2

Model release2026-05-27· 8 min

PredictLM v1: 0.751 cls / 0.609 reg on OpenML via test-time training

Two open-weight tabular models (Mini 13M, Base 26M, Apache-2.0) and a test-time training recipe that hits 0.751 classification accuracy on a 25-dataset OpenML benchmark. Plus 11 architecture experiments that didn't move the needle.

#predictlm #release #test-time-training #negative-results

Model release2026-05-26· 3 min

PredictLM-Mini: a 13M-parameter tabular foundation model with calibrated uncertainty

The smallest open-weight tabular foundation model with calibrated uncertainty out of the box. 13M parameters, 54 MB, Apache-2.0. Statistically tied with Base on classification accuracy and within ~4 pp R² on regression.

#predictlm #release #distillation #calibration

Technical2026-05-25· 4 min

Six architectural experiments. Five lost. Here's what we shipped.

Before settling on Mini's architecture, we ran six experiments at the 26–57M parameter scale. Five lost to the smaller model. Here's the full writeup, including the gradient-NaN debugging.

#predictlm #ablation #negative-results

Technical2026-05-24· 5 min

BarDistribution: why every regressor should return a distribution

A 1024-bin quantile head is the difference between a guess and a calibrated prediction. We explain how the head works, why softmax-over-bins beats Gaussian heads for in-context regression, and what the distribution unlocks.

#calibration #uncertainty #predictlm