Modularium

Modularium Research

AI research for model behavior we can understand.

We study how models act under pressure and turn the results into clear evidence.

Current study

What happens when the test pushes back?

We build evaluations where models use tools, make plans, and face constraints.

Evaluation track

Latest work

Small studies, clear claims.

Evaluation

Tests that show what a model does when the easy answer is not enough.

Status
Building
Area
Research
Read moreAlignment

Studies of incentives, feedback, and oversight.

Status
Scoping
Area
Research
Read moreInterpretability

Checks that explain why behavior changed.

Status
Planned
Area
Research
Read more

Notes

What we are thinking through.

  1. Evaluation should explain behavior, not just rank systemsResearch note · In preparation
  2. Agent audits need pressure, not theaterField memo · Drafting
  3. Mechanistic evidence as a publication standardEssay · Planned
  4. Small surface area, high scrutinyCompany note · Current