Experiments
-
WP-003. Can a model explain its moves?
How well LLMs and computer vision models can reason through competitive games and explain their reasoning.
-
WP-002. Does a fixed relation vocabulary help agent memory?
Tests whether a small fixed set of relations (is_a, part_of, if_then, before and a few more) gives an LLM agent better recall per stored token than letting it write any relation it likes.
-
WP-001. Can an LLM dream?
Corrupting the chats an LLM agent rewrote into memory each night, the way Erik Hoel says dreams distort the day, lowered its LongMemEval recall from 76 to 60 percent and did not help it on new tasks.
What we do
-
Experiments
Each member owns one research project for the term.
-
Evals
We write evals that test one behavior and publish the code.
-
Reading group
We read one paper a week and talk it through.
-
Writeups
Results go on LessWrong and X, and the code goes on GitHub.