Can LLMs beat classical HPO? A benchmark comparing classical, LLM-based, and hybrid methods on Karpathy's autoresearch.

[NeurIPS 2025] PARCO: Parallel AutoRegressive Combinatorial Optimization

Structured Primitives for Efficient Architecture Research

Inference and numerics for multi-hybrid AI model architectures

Pretraining infrastructure for multi-hybrid AI model architectures

A MAD laboratory to improve AI architecture designs 🧪

Hubble is a suite of fully open-source large language models (LLMs) for the scientific study of LLM memorization.

PyTorch-native post-training at scale

Python implementation of Bayesian online changepoint detection

Tora: Torchtune-LoRA for RL

AlgoTune is a NeurIPS 2025 benchmark made up of 154 math, physics, and computer science problems. The goal is write code that solves each problem, and is faster than existing implementations.

A lightweight library for beautiful game of life embeds.

A command-line tool for building static sites from Observable Notebooks

A comprehensive toolkit for streamlining data editing, search, and inspection for large-scale language model training and interpretability.

Official Implementation of wd1

Reproduce ICLR2025 Energy-Based Diffusion Language Models for Text Generation

Implementation of Sharpe-ratio-based active learning strategies for aligning large language models using Direct Preference Optimization (DPO).

Sample-Efficient Preference Alignment in LLMs via Active Exploration

Official implementation for the paper "Toward Scientific Reasoning in LLMs: Training from Expert Discussions via Reinforcement Learning"

A High-Efficiency System of Large Language Model Based Search Agents

Resa: Transparent Reasoning Models via SAEs

Uncertainty-guided Likelihood Tree Search

[ICLR 2026] Tina: Tiny Reasoning Models via LoRA

This package contains the original 2012 AlexNet code.

You like pytorch? You like micrograd? You love tinygrad! ❤️