The Goal of Science Is Not to Win

Background

An RSI loop improves only if each round explores beyond what the system already knows. Agents can now propose candidates far faster than anyone can validate them, so the question becomes how the loop decides what to try next.

The pain

The common answer is generate-and-rank: sample many candidates, score them at the end, keep the best. Feedback arrives late and rarely. The agent learns whether it was heading the right way only after the most expensive evaluation.

In between, it flies blind. It keeps refining answers in the region it already knows and keeps optimizing the one number it can see. The result looks strong under the test and fragile everywhere else.

The challenge

Real discovery is as sparse as a crash in fuzzing: most tries reveal nothing. The loop needs a signal that is cheap, arrives after every try, and measures progress before the final check. That signal must guide the search without becoming a new target to overfit. The final check has its own problem. A wrong answer can fit the agent's own data very well, so the gate has to stand apart from the data the agent steers by.

Our approach

We treat research the way a greybox fuzzer treats a program. A fuzzer rarely finds a bug on any one run, yet every run returns coverage, a cheap signal of which parts of the input space have been touched. The fuzzer aims its next input at what remains untouched.

Our research agents follow three rules. Every experiment exposes a cheap, dense signal of progress. That signal chooses the next experiment, so each step builds on what the last one revealed. The final validation is protected from adaptive reuse.

In practice the agent keeps two maps: one of the world, and one of its own ignorance. After each round it asks whether its theory explains everything it has seen, and which regions it has never touched. Then it runs the test most likely to break its current theory.

The outcome

We ran two agents on the same hidden physical system, differing only in that habit. The agent that tracked its own ignorance found the true law. The one that kept refining in a familiar corner never did, even though its wrong law fit its own data closely.

What held the losing agent back was its feedback. It had plenty of ideas, and nothing told it where to look next.

On silicon

Design-space exploration is this problem in its purest form. Pipeline depth, cache policy, arbitration, and floorplan variants span a huge space. Full synthesis and sign-off take hours per candidate, so a loop cannot wait for sign-off on every try.

It needs fast proxies: lint, simulation coverage, quick synthesis, early timing and congestion estimates. On advanced nodes these proxies can drift far from sign-off, so the loop should keep measuring how well each proxy predicts the final result and lean on it only as far as it holds.

With that in place, the loop tracks which regions of the design space and which corner cases it has not touched, and it prefers the candidate most likely to break the current best design. Sign-off and the held-out regression stay independent. The loop never tunes against them.