AI Trends

Trending story

Google and Other Research Teams Jointly Publish SIM Paper

1 reports1 story updatesUpdated Nov 19 09:33

Understand this story

AI overview

On November 19, 2026, research teams from Google Cloud AI, Google DeepMind, and Pennsylvania State University jointly published the paper “Self-Improving Memory: Enabling Continual Recursive Self-Improvement,” proposing Self-Improving Memory (SIM), which uses continuously updated memory to preserve the reasons for, expected outcomes of, and actual results from code modifications to an external runtime framework. SIM performed better in nine out of 10 independent programming experiments, leading by an average of 2.36 points; in cross-model office tasks, it scored 84.75 points, compared with 73.42 for the control approach. The experiments did not alter model weights, but the experience-inheriting approach scored 0.8 points lower on Gemini 4 Argon. There are currently no facts regarding subsequent developments.

AI-generated from public reports

Latest development · Report

Google and Penn State Introduce SIM, Enabling AI to Carry Optimization Experience Across Models and Tasks

Developments

1 confirmed developments

  1. Research teams from Google and other institutions jointly release a paper on SIM

    Source:36Kr · AI

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 19

  1. 36Kr · AI

    Google and Penn State Introduce SIM, Enabling AI to Carry Optimization Experience Across Models and Tasks

    Google Cloud AI, Google DeepMind, and Penn State jointly introduced Self-Improving Memory (SIM), which uses continuously updatable memory to retain the rationale, expected outcomes, and actual results of Harness code changes, enabling continuous recursive self-improvement across models and tasks. In 10 independent programming experiments, the SIM approach performed better in nine, with an average lead of 2.36 points. In cross-model office tasks, the SIM approach achieved a final score of 84.75, compared with 73.42 for the approach without SIM. The experiments modified the external runtime framework for large language models while leaving model weights unchanged, and the approach that inherited prior experience trailed by 0.8 points on Gemini 4 Argon.

Related stories

  1. Related storiesApple research team submits paper to arXiv disclosing the HALO system
  2. Related storiesNvidia releases Nemotron 3.5 Super VL and Bolt models
  3. Related storiesEnglish translation is being prepared.
  4. Related storiesGoogle DeepMind demonstrates using Antigravity to discover antimicrobial peptide candidates