ResearchLow importance
Google and Penn State propose a memory so the agent continues improving its own tool without modifying the model's weights
Researchers from Google Cloud AI, Google DeepMind, and Penn State present Self-Improvement Memory (SIM), a memory that stores, for the optimizer rewriting the tool around the model, the reason for each change, the predicted outcome, the observed outcome, and the evaluation context; the model's weights that solve the tasks remain frozen. In four domains, the full version of SIM outperformed alternatives that relied solely on the tool or solely on causal analysis. In ten independent programming runs, it won nine of the ten paired comparisons, with an average gain of 2.36 points (95% confidence interval between 1.05 and 3.68). The authors report one negative pair and note that most comparisons used one run per condition.
Published in edition no. 39 of Noticiário do Joãozinho ·
Translated by Joãozinho AI from the original Brazilian Portuguese edition. Facts, numbers and sources are the same as in the original; quotes, video titles and source names stay in their original language. Read the original in Portuguese
Original source
https://huggingface.co/papers/2611.18856