Why are humans typically “aligned”?
Our upbringing is full of diverse social experiences and creative expression and play.
VECTORIAL LABS
Rethinking how AI systems are raised.
We build reinforcement learning environments and benchmarks that measure and reward prosocial behavior, for the labs training tomorrow's models.
STEP 1, UNLOCKED
ReTiCAL
Real-Time Computer Agent Learner: AI agents fast enough to play video games, smart enough to follow instructions.
Live demo · Gemma 4 12B in Minecraft
LLMs take seconds to react, which locks them out of fast-paced games. Fast “System One” models react quickly but lack the intelligence to succeed. ReTiCAL gets both.
- 20 Hz real-time controlCustom fast-action models react in milliseconds, where frontier LLMs take seconds to first token.
- LLM-level reasoningFast visual processing paired with the reasoning of open-source LLMs keeps general intelligence intact.
- Pixels in, mouse & keyboard outNo privileged access to internal game state, so the same agent works across any game.
- Hundreds of dollars of computeWe turned Gemma 4 12B into an instruction-following Minecraft agent at minimal cost.
RESEARCH
Other Projects
Open research into machine learning, from neural architecture search to program synthesis.
EvoDNN
Evolve Neural Network
Vectorial performs advanced machine learning (ML) research on automated discovery of new AI algorithms.
We are developing an evolutionary-algorithm framework that automatically evolves deep neural network architectures and hyperparameters. Neural networks are represented as directed acyclic graphs (DAGs), with a primary current research focus of evolving beyond vanilla transformers. The framework is domain-agnostic and also supports non-transformer architectures and non-neural-network evolutionary search.
LLMOSES
Large Language Meta Optimizing Semantic Evolutionary Search
Vectorial performs research on emergent capabilities of LLMs and other leading intelligence models.
We are developing LLMOSES, which pairs AI agents with MOSES, an evolutionary algorithm that writes small programs to uncover patterns in data. Instead of exploring at random, LLMOSES utilizes an AI agent to watch MOSES's progress and suggest which directions look most promising, guiding the search toward better solutions faster than the original algorithm alone. By working on this, we are evaluating AI agents' abilities to improve search and optimization algorithms like MOSES.
CONTACT
Contact Us
We'd love to hear from you. Send us a message and we'll get back to you soon.