Model-Level RSI: What It Means When a Model Improves Its Own Weights

2026-10-12 · VIDRAFT

aimachinelearningllmbenchmark

TL;DR

Most self-improvement in LLM systems happens around the model: better prompts, better agent scaffolds, better tool use. Model-level RSI is different. It changes the weights of the model itself, without a human handing it new labels. This post lays out the five stages we use to talk about RSI, where the hard line sits, and how to read public leaderboard results so you are measuring improvement rather than leakage.

The question that matters

Here is the honest question behind recursive self-improvement: can a model get better at a task when nobody gives it the answers?

It is easy to fake a yes. Tune on the evaluation set, pick the best of many runs, or let a fallback path quietly fabricate an answer, and the number goes up while nothing real changed. So the useful framing is not "did the score rise" but "did the weights improve under a rule that cannot see the test."

The five stages of self-improvement

We find it helps to separate five stages, because they are often lumped together.

1. Tool and prompt improvement. The scaffold around the model changes. The weights do not. This is where most agent frameworks live.
2. Human-labeled fine-tuning. People provide new labeled data and the model trains on it. Real improvement, but bounded by human labeling.
3. Model-generated data with human filtering. The model proposes, humans curate, the model trains on the survivors.
4. Model-level RSI. The model solves problems, a verifier that cannot see the answer key filters the good solutions, and the model trains on those. The weights improve with no human labels in the loop.
5. Open-ended self-direction. The model also decides what to work on next. This is the frontier and should be claimed carefully.

The jump from stage 3 to stage 4 is the real one. It is also where measurement discipline matters most.

How to not fool yourself

If you train on self-generated solutions, the filter is everything. Three rules keep the loop honest.

• Bind the filter to something outside the model. Code execution, an external reference, or a held-out answer key. A model grading itself with the same weights is not a filter.
• Keep training and test disjoint. If any evaluation item leaks into training, the gain is an illusion. Count unique items, not rows.
• Treat a positive margin as a hypothesis, not a win. A small rise inside run-to-run noise is not improvement. Use paired comparisons and report the interval.

These are not optional. They are the difference between a result and a press release.

Reading leaderboard results

Public leaderboards are useful precisely because an outside party runs them. A few numbers that are worth stating plainly, all from public evaluations:

• AIME 2026 and HMMT February 2026: full marks on the competition math sets.
• GPQA Diamond: 93.43% in the sub-500B class, a set of graduate-level science questions written to resist search.
• Breadth: top placements across eleven public benchmark categories, and sixteen categories on the Polaris drug-prediction leaderboard.

What makes these meaningful is not the size of the number. It is that the test was held by someone else, the questions resist memorization, and the same method holds across very different domains. A single high score can be luck. Breadth across independent boards is harder to fake.

A note on throughput

Self-improvement is cheap to talk about and expensive to run. The loop only closes in practice if serving is efficient, because every round means generating, filtering, and retraining. Pushing more useful tokens through the same hardware is what makes the loop affordable, and it is why inference efficiency and RSI belong in the same conversation.

FAQ

Is RSI the same as an agent that uses tools? No. Tool use is stage one. The weights do not change. RSI changes the weights.

Does RSI need human labels? Stage four is defined by not needing them. It needs a verifier that is bound to something external, not a human answer key.

How do you know a gain is real? Disjoint train and test, an external filter, and paired statistics with reported intervals. Without those three, treat the gain as unproven.

Why report leaderboard numbers at all? Because an outside party runs them on questions that resist memorization. Breadth across independent boards is the signal.

Further reading

The pattern that generalizes is simple to state and hard to execute: improve the weights under a rule the test cannot see, and measure with disjoint sets and paired statistics. Everything else is scaffolding.