Skip to main content
MICHIEL HORSTMAN
BOOK A CALL ↗
← All work

Independent project · model-behavior research

Study refusal as a behavior you can measure and intervene on.

Blasphemous is a research toolkit for examining refusal-related representations in transformer activations and testing interventions against an evaluation suite.

Conceptual artwork showing a language model activation space and a refusal direction
EDITORIAL PROJECT ARTWORK, NOT A PRODUCT SCREENSHOT

The project asks where refusal-related directions appear in model activations, and what changes when an intervention is tested against a fixed set of prompts.

Independent project

OPEN SOURCE REPOSITORY ↗

The system

  • Tools for analyzing refusal-related directions and activation geometry.
  • Intervention paths that include projection and LoRA-based methods, with Bayesian search for tuning experiments.
  • Evaluation that separates harmful-prompt refusal from harmless-prompt refusal and tracks distribution shift with KL divergence.

What shaped it

01

Measure both sides of refusal

The evaluation tracks refusal on harmful prompts alongside false refusal on harmless prompts, rather than optimizing one number in isolation.

02

Compare interventions under the same evaluation

Projection and LoRA approaches can be examined against shared prompt sets and scoring criteria.

03

Treat experimental methods as experimental

The repository labels optimal-transport intervention work as experimental and retains a projection fallback.

What this does not prove

  • Results depend on the model, prompts, data, and scoring setup. They do not establish general safety or behavior across models.
  • An improved score on a test suite is not proof that an intervention preserves every capability or generalizes to unseen prompts.

Surface. Thanks for diving.

Shark model by alextkd2003, CC BY, converted to ASCII.