Skip to content
Trending storyDeveloping

Meta says Muse Wonder Full generalizes better than Muse Spark 1.3 to unfamiliar tasks

2 articles1 sourcesince Oct 12Last article 1h ago ·

Overview

AISummary of 2 articles

Meta says Muse Wonder Full scores higher than the Muse Spark 1.3 trend predicts on tasks not covered by post-training.

In a test-time compute scaling comparison, both models improved similarly on tasks seen during post-training, but Muse Wonder Full did better on unfamiliar tasks. Meta attributes the gap to a stronger base model that carries post-training gains to tasks it did not anticipate.

A later Meta post from the same account repeats the claim, but frames the comparison as holding in-distribution scores equal before measuring out-of-distribution performance. The earlier and later posts describe the same result with different wording about the baseline.

Written by AI from the articles below · updated Oct 12, 1:51 PM ET

Check the sources:

Latest developmentOct 12, 1:29 PM ET
Meta says Muse Wonder Full generalizes better to unfamiliar tasks

Article timeline

The articles in this story. Times are ET.

Oct 12
  1. AI at MetaOfficial
    Meta says Muse Wonder Full generalizes better to unfamiliar tasks

    AIMeta compared test-time compute scaling on tasks seen during post-training with unfamiliar tasks, finding that Muse Wonder Full and Muse Spark 1.3 improve similarly out-of-distribution. At the same in-distribution score, Muse Wonder Full outperforms the Muse Spark 1.3 trend on unfamiliar tasks, which Meta attributes to a stronger base model.

    Image from @AIatMeta's post
  2. AI at MetaOfficial
    Meta says Muse Wonder Full handles unfamiliar tasks better than Muse Spark 1.3

    AIMeta compared test-time compute scaling for two models on tasks seen during post-training and on unfamiliar tasks. Both models improve similarly on familiar tasks, but Muse Wonder Full scores higher out of distribution than the Muse Spark 1.3 trend predicts. Meta attributes this to a stronger base model that transfers post-training gains to unanticipated tasks.

    Image from @AIatMeta's post

Heat trend

Not enough continuous observations to show a trend yet.