Interpretability·2025
On the Geometry of Concept Representation in Language Models
We show that high-level concepts in transformer models form predictable geometric structures, enabling targeted concept editing without retraining.
Olyxee LabExploring the frontiers
of intelligence
Olyxee Lab is our research division — dedicated to advancing the science of artificial intelligence. We work on fundamental problems in AI safety, interpretability, and capability, and publish our findings openly.
Research
areas
Active
Frontier Model Research
Pushing the boundaries of what AI systems can understand, reason about, and generate. We explore novel architectures and training paradigms.
Active
AI Safety & Alignment
Developing methods to ensure AI systems behave as intended. Mechanistic interpretability, reward modeling, and formal verification of AI behavior.
Active
Interpretability
Understanding how neural networks represent and process information internally. Making the black box transparent without sacrificing capability.
Active
Efficient Intelligence
Building AI systems that achieve more with less — smaller models, fewer parameters, lower energy. Intelligence shouldn't require a data center.
Exploring
Multimodal Reasoning
Systems that can perceive, reason across, and generate content spanning text, vision, audio, and structured data simultaneously.
Exploring
Emergent Capabilities
Studying how complex behaviors arise from simple training objectives. Understanding and predicting capability jumps in scaled systems.
Recent work
AI Safety·2025
Verification-First Deployment for Edge AI Systems
A framework for systematic pre-deployment verification of AI models targeting heterogeneous edge hardware.
Efficient Intelligence·2025
Sparse Attention Is All You Need: Efficient Transformers for Resource-Constrained Environments
We demonstrate that structured sparse attention patterns can match dense attention quality at 40% of the compute budget.
AI Safety·2024
Failure Mode Taxonomy for Deployed AI Systems
A comprehensive classification of how AI systems fail in production, with detection signatures and mitigation strategies for each mode.
Our approach
Open research
We publish our findings, share our methods, and contribute to the broader scientific community. Knowledge compounds when shared.
Safety by design
Every research direction is evaluated through the lens of safety. We don't build capabilities without understanding their implications.
From lab to product
Our best research becomes Olyxee products. The path from paper to production is short and deliberate.
Join us at
the frontier
We're looking for researchers and engineers who want to work on problems that matter.