PhAI Labs Launches RSI-Anything: A World Model Drives AI Self-Evolution Across 5 Tasks

Researchers at PhAI Labs have published RSI-Anything, a recursive self-improvement (RSI) framework that they tested on five very different tasks: bug fixing, computer control, lab operation planning, chemical reaction optimization, and immune system research. The most telling experiment involves code. The team did not replace the model that writes the edits and did not increase the online testing budget. It only added a world model to predict the consequences of changes and select the best ones.

The idea: predict first, then change

Most self-improvement systems follow a "generate, run, check the result" loop. It is expensive. Every candidate check requires a real run: tests, a simulation, a lab experiment. The more complex the environment, the faster this brute-force search hits the budget ceiling.

RSI-Anything moves part of that work inside a world model. The model estimates what a proposed edit will lead to before the expensive check. Only candidates the model considers promising then reach the real environment. The executor's job stays the same: write code, click buttons, draft a protocol. A separate layer handles selection and the direction of improvement.

The choice of domains looks deliberately eclectic. A bug in a repository, an operating system interface, a lab protocol, a reaction in a flask, and the immune system have one thing in common: each is an environment with feedback, and checking in it is expensive or slow. That is where predicting an outcome can save the most.

What the experiments showed

The available description covers only coding in detail. The experimental conditions are tightly fixed:

• the code-generating model stayed the same;
• the online testing budget was not increased;
• the only change: a world model takes part in predicting results and selecting edits;
• the comparison was against the same executor without this layer.

This design rules out the most common explanations for a gain: a stronger base model and more attempts. If the improvement holds up, its source would be the quality of selection. The available materials give no numerical results, benchmark names, or size of the gain, so it is too early to judge the scale of the effect. Details on the other four domains have not been disclosed yet either.

Caution is warranted here for two reasons. A world model can make predictable errors, and selection would then lock in a systematic mistake instead of correcting it. Also, transferring the approach between domains has to be verified for each one separately: a chemical reaction and a code patch work differently, and a general framework does not guarantee the same benefit.

What teams should check in practice

For agent developers, the article offers a concrete takeaway. If checking candidates in your environment is expensive (CI that takes hours, a physical test bench, a lab run), it is worth assessing whether you can train an outcome predictor on accumulated logs. That is cheaper than increasing the number of attempts or swapping the base model.

When evaluating work like this, it helps to look at four things:

• whether the budget of real checks is fixed and whether it is the same across the compared variants;
• how the world model behaves on tasks outside the training distribution;
• what share of edits are selected wrongly, and what happens to them afterward;
• whether the result reproduces on public benchmarks, not only on internal sets.

For now, RSI-Anything remains a claim of a general path to self-improvement through prediction. It will become convincing when the authors publish full tables and code, and independent groups repeat the experiments in at least two of the five domains.