WORLD MODELS
Request draft Right for the Right Reasons
Investigates whether a world model’s confidence tracks the reasons its predictions succeed. The draft considers predictability and surprise as signals that may reveal a mismatch.
Working papers.
Questions currently being developed into technical drafts. Each entry describes an investigation in progress.
COUNTERFACTUAL LEARNING
Request draft Twin Worlds
Uses paired worlds to ask what a model learns from their differences. By changing one condition at a time, the work examines whether a representation captures a causal distinction.
GENERALIZATION
Request draft Competence, Not Structure
Tests whether apparent structural understanding reflects reusable competence. The focus is on behavior under changed tasks, rather than performance on familiar examples alone.
ENERGY-BASED MODELS
Request draft When Energy Decomposes
Studies when an energy function can be separated into meaningful constraints. The question is whether that structure helps explain how a model rules out inconsistent possibilities.
LANGUAGE MODELS
Request draft How Deep It Goes
Examines how far language models can track nested beliefs about other agents. The draft looks for ways to distinguish a shallow answer pattern from deeper recursive reasoning.
These are working descriptions, not published papers.