UnMaskFork
Test-time scaling where multiple masked diffusion LMs collaboratively unmask an answer, guided by MCTS.
-
Interactive Research Blogs
The AI Picbreeder Experiment: In Search of Automatic Open-Endedness
We replace Picbreeder's human users with AI (large Vision-Language Models), and ask which ingredients let AI search open-endedly the way people do.
CoffeeBench
A benchmark for evaluating how much net income an LLM agent can generate as a coffee roaster over 90 days in a multi-agent economy with two farmers, two roasters, and two retailers.
Digital Ecosystems
Watch digital species compete, cooperate, and evolve in real time.
Doc-to-LoRA and Text-to-LoRA
Hypernetworks that generate LoRA adapters from documents or text descriptions.
Digital Red Queen
LLMs evolve assembly programs that battle for control of a virtual machine.
DroPE
Extending LLM context windows by dropping positional embeddings after pretraining.
RePo: Context Re-Positioning
A lightweight module that assigns semantically-informed positions to tokens, enabling better attention on noisy contexts and long sequences.
Petri Dish NCA
Multiple neural cellular automata species compete and adapt in a shared environment.
Sudoku vs GPT-5
Why Sudoku variants remain a grand challenge in AI reasoning.
Continuous Thought Machines
Neural networks that use timing and synchronisation as a representation for thought.
Sudoku-Bench
A benchmark testing creative reasoning through Sudoku variants with invented rules.
TinySwallow
A small Japanese language model that runs entirely in the browser.
ASAL
Foundation models automatically discover diverse artificial life simulations.