Skip to content
Trending storyDeveloping

Google DeepMind paper finds self-improving agents forget old tasks as they learn new ones

1 article1 sourcesince Oct 20Last article 2h ago ·

Overview

AISummary of 1 article

A Google DeepMind paper, "Are Self-Evolving Agents Continual Learners?", tests agents that edit their own prompts, memory, or skills and finds they lose earlier skills while learning new ones.

According to a report summarizing the paper, GEPA gained 6.2 points per task block on average but lost 4.3 points on other tasks, making it the fastest learner and also the method that forgot the most.

The same report says a later edit cut one rule's 15.5-point gain to 2.5 points. The paper's recommendations, as reported, are to keep task-specific rules separate and to retest old tasks after each update.

Written by AI from the articles below · updated Oct 20, 3:17 PM ET

Check the sources:

Latest developmentOct 20, 2:47 PM ET
Google DeepMind study finds self-evolving agents forget earlier tasks as they learn

Article timeline

The articles in this story. Times are ET.

Oct 20
  1. Rohan PaulXPick
    Google DeepMind study finds self-evolving agents forget earlier tasks as they learn

    AIA Google DeepMind paper, Are Self-Evolving Agents Continual Learners?, tests whether agents that edit their own prompts, memory, or skills keep old skills while learning new ones. The post says GEPA gained 6.2 points per block on average but lost 4.3 points on other tasks. It also reports that a later edit cut one rule's 15.5-point gain to 2.5, and it advises keeping task-specific rules separate and retesting old tasks after each update.

    Image from @rohanpaul_ai's post

Heat Trend

Not enough continuous observations to show a trend yet.