Latest Research
View allEndogenous subspace orthogonalization: targeted monosemanticity via LoRA
Ortho-LoRA uses a task-conditioned geometry loss to reduce overlap between adapter channels, with the aim of making model behavior easier to inspect and steer.
LoRA composition: capability retention without functional composition
A controlled study of when merged adapters retain skills, interfere with one another, and fail to execute an unseen ordered task.
DriftLab: measuring agent adaptation in non-stationary environments
A testbed for studying how agents notice, recover from, and retain changes in the worlds they work in.
Latest Notes
View allWhen Interpretability Outgrows Language
Why investigating frontier intelligence may require agent interrogators—and how machine-speed analysis can remain open to human challenge.
Agents Need Memory. But Where Should Memory Live?
Why learned procedures may need a different home from facts, documents, and past interactions.
Agents change, so does the world
Most agent evaluations assume the world will wait for the agent. This note asks what changes when the environment moves while an agent is still learning how it works.
Open for collaborations
Have a question worth unfolding?
We work with researchers and organizations on difficult questions in machine intelligence.
contact@unfoldinglabs.ai