RSI-Anything: Predict Before Testing, the Self-Improving AI Loop With World Models

The RSI-Anything repository from the Gen-Verse organization on GitHub opened on October 8, 2026, reaching 122 stars and 10 forks in four days. The Apache-2.0 licensed work combines self-improving artificial intelligence research with a model that predicts the world. The topic is an agent model deciding in advance which of its proposed paths to really test in each round.

The idea in plain words is this: do not test everything that comes to mind; predict first, test the most promising one, look at the result, and grow smarter next round. It is like tapping the outside of four melons to hear which one is ripe instead of cutting all four open. Real testing is costly while guessing is cheap. An agent model thus makes more accurate decisions with the same budget every round.

Propose, predict, test, keep: a six-step loop

The loop has six steps: a policy or workflow proposes candidates, a world model predicts each candidate outcome, the top-scoring candidate enters real testing, prediction and observed result are recorded side by side, feedback updates the policy or the model, and a retention check decides which version stays. Policy and model can update on separate schedules; the policy can improve while a reliable model stands still, and vice versa. The kept policy and model pair shapes the next round candidates and learning data, so improvement never stays in a single round but piles up.

At the base sits JEPA-Anything: the team led by PhAI Labs built the Orthogonal Predictive Factorization technique, which splits a hidden target into complementary factors and predicts each factor through its own branch. A MarkTechPost analysis notes this design fixes the problem of strong signals crushing weak ones in single-piece prediction; the team tested the method as a world model across seven areas including vision, biology, clinical trajectories, control, molecular dynamics, physical fields and weather, beating the plain JEPA baseline on all ten dynamics tasks. An AIWeekly roundup reports that a biological intervention suggested by one learned factor was confirmed in cell and mouse experiments.

Program repair works like this: the 7B producer model named ReasonFlux-Coder-7B stays fixed and proposes four patches per round, while the 3B world model named W1 returns outcome odds in five classes for every patch and test pair. Only the top-scoring patch runs against real tests, for at most three rounds. 200 CodeContests tasks with three seeds make 600 evaluations; hidden-test success rises from 26.3 percent with a static checker to 29.2 percent with a fixed model setup and 31.2 percent with model adaptation. Test calls stay at 3.7 per task. CodeContests is the dataset the DeepMind team opened for competition-level program generation; the technical report on arXiv stresses the comparison used the same producer and the same candidate budget.

For generalization the same world model repeated the loop on two large collections: full-test success reaches 46.37 percent on the 1,000-task APPS Introductory measure and 33.05 percent on the 1,055-task LiveCodeBench v6 measure. The starting program stayed identical for every task and seed, public tests guided the work, and the full official test pack had to pass at the end. Random choice and the static checker stayed behind on both sets.

From screens to laboratories: one loop, four separate worlds

In screen work an agent tried the checkbox task from the MiniWoB++ pack on the same 100 cases across three versions: success climbed from 51 to 79 and then 99 percent, mean actions per episode fell from 5.30 to 3.21, and pointless repeated clicks dropped from 237 to zero. At every step a world model scores which click will work before it happens. The middle version acted less yet submitted wrongly 21 times; the final version acted slightly more and cut the fault to 1. On desktop work the OSWorld measure shows an 18.4 percent relative gain over the starting agent and a 5.5 percent gain over the RLAnything setup. The public OSWorld repository on GitHub measures such agents with 369 real computer tasks.

In laboratory planning the LabHorizon F200 measure tracked five consecutive versions of the same 4B policy model: the final score rose from 21.80 to 46.01 percent, passing the giant reference setup at 45.32 percent. A world model filters candidate sequences before execution. In chemistry Suzuki-Miyaura screens the best observed yield rose from 87.5 to 93.0 percent on the 20-run GC screen and from 82.16 to 96.15 percent on the 40-run flow LC-MS screen, with the run count unchanged.

The immune study built an end-to-end line: first agent workflows evolved, then a world model carrying cell, tissue and individual context was built, and an analysis agent with a fixed model performed prediction, self-correction and hypothesis ranking. The output is a candidate hypothesis suggesting SIRPα blocking together with IL-36γ axis tuning; wet-lab testing has not happened yet. The technical report, the GitHub repository and the models on HuggingFace are all public. The reproduction notes are concrete too: Linux with Python 3.10, one-command setup, four-stage training through the rsi plan and rsi run commands; the evaluation archive is 64 MB and the training add-on is 236 MB.

The honest limits are written in the report too: the laboratory score is planning success, not bench execution; the chemistry result is limited to two Suzuki screens with 20 and 40 runs; the immune hypothesis waits for testing. A world model learns to predict more accurately with every round. Still the picture is clear: when outside feedback is costly, predict first, test little, teach the result to both the decider and the predictor, and never keep a pair that does not fit. Any team with a growing test bill learns the same lesson.