AI Should Build Its Own Research World Model

Background

RSI means turning experience into lasting improvement. Experience can live in two places: in model weights, through training, and in an explicit record the system keeps. During a run the weights are frozen, and when the context window ends, whatever the agent learned by trial and error disappears.

The pain

Every new session starts from zero and repeats experiments that already failed. Training folds some experience back into weights, but it happens between rounds, and the knowledge it stores is hard to read, extract, or debug.

Saving the final answers does not fix this. A list of solutions to past problems carries none of the reasoning that produced them, so it breaks the moment the problem changes.

The challenge

The record has to hold understanding. It must stay honest, so refuted beliefs never quietly return. It must say where each lesson holds. It must answer questions from an agent that is stuck. And it has to be written during the work without derailing it.

Our approach

We give the agent a research world model: a continuously evolving record of what it knows. It names the objects and concepts it meets. It writes numbered claims, and each claim carries its evidence, its conditions, its confidence, and a test that would prove it wrong. Refuted claims stay on file, marked refuted. Dead ends become constraints on whoever comes next.

Current beliefs can be rewritten. The history of how they were reached is append-only. A research manager writes to the record only when a question reaches closure. A read-only interface lets an agent query the record and get a falsifiable prediction back. A new agent reads the record before it acts and asks it when it gets stuck.

The outcome

We dropped an agent with frozen weights into an unknown puzzle world with no rules. It built this record as it went, derived solutions on paper before spending moves, and corrected its own wrong laws on file. When one session ran out, the next picked up from the record and finished the job. A copy of the same model, given the same compute and only the correct answers to earlier stages, made no progress.

In that run, the capability lived in the agent together with its world model.

On silicon

Picture a timing-closure campaign on one block. Each run tries a fix: resize cells, add a pipeline stage, change the floorplan, relax a constraint. The world model should hold the working state of that campaign: which paths fail and the current theory of why, which fixes were tried, what the slack did, and which theories the reports refuted.

The next run, or the next agent, should start from that state and ask it before trying something. Engineers can read it and correct a wrong theory directly.

This gives a self-improving design loop two speeds. The world model improves after every run. The model improves when audited traces train it.