Qwen3's internal states collapse thousands of conversations into a thin sheet

What thousands of conversations look like inside a language model

Qwen3's internal states collapse thousands of conversations into a thin sheet

The author maps Qwen3-1.7B's residual-stream states for 900 English, Chinese, German, and Croatian Q&A examples plus 90 Python coding tasks. In a fixed 3D PCA view, conversations form a thin sheet holding 99.3% of English variation, yet that view discards 82.4% of the full 2,048-dimensional variation. Languages overlap with distinct centers; code sits near English. Refitting the axes on conversation averages alone raises captured variation from 42.4% to 80.1%, showing the map's shape depends on what it is optimized to preserve.

The visible sheet contained almost all the variation in the 3D view, but that view left out 82.4% of the original variation.
  1. netruk44

    I’ve been playing around with mapping steam games + their reviews based on Qwen embeddings. I revisit this project every year around this time to see what improvements I can make to make a “better” map than last year (According to my own preferences). [0/1]

    One thing I learned this year is that naive text mapping (embed a corpus of reviews with a Qwen embedding model, use umap to plot, then compare where things fall) primarily compares embedded text length more than anything else. What a review says affects the position less than how long the review is.

    This was a little surprising to me. For whatever reason, I didn’t think the embedding would so strongly describe the length of the embedded passage. So most of this year’s time was spent fixing that artifact.

    [0]: 2025’s map: https://storage.danieltperry.me/share/steam-map/

    [1]: 2026’s map (still a work in progress, both the ui and the review processing are incomplete): https://storage.danieltperry.me/share/steam-map-refresh/

  2. Scene_Cast2

    The methodology feels either heavily undercooked to the point of being misleading, or heavily under explained. Why are we in 3D? At which point did dimensionality reduction hit? And if we're doing dimensionality reduction, does the "plane" really mean anything?

  3. supermatt

    Isn’t this last layer analysis basically saying that after a bunch of Chinese tokens you get a Chinese token, and after some Latin tokens you get a Latin token? That seems expected given the setup. Would be interested in seeing what translations look like in the middle layers.

  4. harttrav

    umap (https://umap-learn.readthedocs.io/en/latest/) or tsne (https://scikit-learn.org/stable/modules/generated/sklearn.ma...) might be an interesting alternative to pca for dimensionality reduction

  5. rustyhancock

    I wish there had been a third broad language style such as arabic although perhaps Qwen doesnt use arabic.

    It made me think what if totally different knowledge bases could exist in a model where as my mental model was similar to OPs translation from pure knowledge source.

    It also explains why rewriring prompts to another language sometimes lets you bypass restrictions.

More from this day

2026-11-15