PhAI Labs released RSI-Anything，AI Agent Code repair pass rate improves

Article

RSI-Anything lets AI improvements to be made while working, fixing more bugs without changing the programming model

According to a news brief from Beating AI news,AI research institution PhAI Labs released RSI-Anything, attempting to let AI Agent continually accumulate experience from real-world tasks and improve its own methods of working. This framework can be used for code repair, computer operation, and scientific research.

Its approach is to equip the Agent with a “world model” dedicated to predicting the possible results of different actions. The Agent first proposes several plans, the world model determines which is more likely to succeed, and then selects a plan for actual execution. The system subsequently compares the prediction with the actual result and improves the Agent’s action strategy. If the prediction is inaccurate, it also updates the world model. A new version can be retained for the next round of tasks only after passing tests.

For bug fixing, for example, the coding model generates four modification plans at once.RSI-Anything first predicts which one is most likely to pass the tests, runs only one in practice, and then uses the test results to improve subsequent judgments. This allows limited testing opportunities to be allocated to the most promising plans.

In 200 Code CodeContests problems, the coding ReasonFlux-Coder-7B remains unchanged throughout, and the number of tests is also broadly the same. With a traditional static verifier, the repair pass rate was 26.3%. After introducing world-model screening, it rose to 29.2%, and after updating the world model using actual feedback, reached 31.2%.

The team also applied this method to computer operation and scientific experiments. In laboratory planning tests, a 4B model was continuously updated through five versions, with the planning score rising from 21.80%to 46.01%.

The project code has been made public. At present, different tasks still use their own Agents and world models. The paper also acknowledges that whether execution strategies and world models can continue to improve jointly over the long term on the same task requires further experimental validation.

Original link https://m.theblockbeats.info/flash/371379

Key points

AI Research institution PhAI Labs released RSI-Anything。
The original text states that this framework can be used for code repair, computer operation, and scientific research.
The Agent first proposes multiple plans; the world model predicts the results and screens the plans for execution. The system then updates the action strategy based on the predictions and actual results, and also updates the world model when predictions are inaccurate.
A new version must pass tests before it can be retained and used for the next round of tasks.
When fixing bugs, the coding model generates four modification plans at once,RSI and -Anything predicts and selects one of them for actual execution.
In 200 Code CodeContests problems,ReasonFlux-Coder-7B remained unchanged and the number of tests was broadly the same; the repair pass rate with a traditional static verifier was 26.3%, after world-model screening it was 29.2%, and after updating the world model with actual feedback it was 31.2%。
In laboratory planning tests, a 4B model was continuously updated through five versions, with the planning score rising from 21.80%to 46.01%。
The project code has been made public.
At present, different tasks still use their own Agents and world models.
The paper acknowledges that whether execution strategies and world models can continue to improve jointly over the long term on the same task still requires further experimental validation.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This method is like assigning AI Agent a “results forecaster.” The Agent first prepares several approaches, the world model determines which is more likely to succeed, only one is executed, and the actual result is then used to adjust future selection. If the forecast itself is inaccurate, the world model is also updated.

Why it matters to readers

The original data show that, with the programming model unchanged and the number of tests broadly the same, adding plan screening and actual feedback resulted in a higher repair pass rate. This suggests that limited testing opportunities may be concentrated on more promising plans.

For newcomers, the key point is that it improves results without replacing the coding model; instead, it adds a process of prediction, screening, testing, and feedback-based updates outside the model. The project code has been made public, but the original text does not provide a code address.

Risks and unknowns

The provided material is a media news brief and does not include the original link to the paper or project code, so the relevant methods and data cannot be independently verified based on it alone.
It remains unclear whether the reported results can hold across more tasks, models, or independently reproduced experiments.
Whether the execution strategy and world model can continue to improve jointly over the long term within the same task remains to be validated.
Different tasks currently still require their own Agent and world model; cross-task applicability remains unclear.

PhAI Labs released RSI-Anything，AI Agent Code repair pass rate improves

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center