RSI-Anything
A general-purpose recursive self-improvement framework built on JEPA-Anything. It uses a predictive world model to guide action selection and training data filtering, introduces a retention mechanism to ensure sustainable improvement, and supports task domains including code repair, GUI operation, and scientific planning.
Timeline (last 90 days)
RSI-Anything is a general framework proposed to address two pain points of recursive self-improvement (expensive external feedback and world model misalignment caused by policy evolution), built on top of JEPA-Anything
RSI-Anything uses a predictive world model to guide the selection of actions and which data to use for learning
RSI-Anything introduces a retention mechanism with component checks and compatibility checks to decide whether an updated policy-model pair is worth retaining
In code repair tasks, the success rate using a static validator was 26.3%, rising to 29.2% after introducing predictive selection, and reaching 31.2% after world model adaptation
In GUI task evaluation, RSI-Anything reduced the average number of operation steps by 39.4%
Across five consecutive policy version iterations, the laboratory planning score improved from 21.80% to 46.01%
In reaction optimization tasks, RSI-Anything achieved an average best-observed yield of 96.15% under a budget of 40 experiments, outperforming Bayesian optimization's 82.16%
The RSI-Anything study presents an end-to-end immunological science application case that chains workflow evolution, predictive modeling, and therapeutic hypothesis generation
The retention mechanism implementation path of RSI-Anything is closer to system-level component validation than to parameter-level regularization constraints (such as EWC or PNN)
In the code repair experiments, the encoder (coder) was kept fixed, and the improvements were achieved under comparable test resource consumption
All Facts (10)
RSI-Anything is a general framework proposed to address two pain points of recursive self-improvement (expensive external feedback and world model misalignment caused by policy evolution), built on top of JEPA-Anything
50%UnverifiedRSI-Anything uses a predictive world model to guide the selection of actions and which data to use for learning
50%UnverifiedRSI-Anything introduces a retention mechanism with component checks and compatibility checks to decide whether an updated policy-model pair is worth retaining
50%UnverifiedIn code repair tasks, the success rate using a static validator was 26.3%, rising to 29.2% after introducing predictive selection, and reaching 31.2% after world model adaptation
50%UnverifiedIn GUI task evaluation, RSI-Anything reduced the average number of operation steps by 39.4%
50%UnverifiedAcross five consecutive policy version iterations, the laboratory planning score improved from 21.80% to 46.01%
50%UnverifiedIn reaction optimization tasks, RSI-Anything achieved an average best-observed yield of 96.15% under a budget of 40 experiments, outperforming Bayesian optimization's 82.16%
50%UnverifiedThe RSI-Anything study presents an end-to-end immunological science application case that chains workflow evolution, predictive modeling, and therapeutic hypothesis generation
50%UnverifiedThe retention mechanism implementation path of RSI-Anything is closer to system-level component validation than to parameter-level regularization constraints (such as EWC or PNN)
50%UnverifiedIn the code repair experiments, the encoder (coder) was kept fixed, and the improvements were achieved under comparable test resource consumption
50%