cd /news/machine-learning/canonical-basis-realignment-for-tran… · home topics machine-learning article
[ARTICLE · art-115357] src=github.com ↗ pub= topic=machine-learning verified=true sentiment=· neutral

Canonical-basis realignment for Transformer LLMs: every hidden axis becomes independently measurable and controllable

A new technique called canonical-basis realignment rotates a Transformer's internal coordinate system into a lossless canonical basis, enabling every hidden axis to be independently measured and controlled without altering outputs or perplexity scores. Researchers applying it to models like Qwen and Pythia observed internal mechanisms including a bipolar oscillator, rhythmic layer-wise respiration, and a homeostatic defense that erases localized perturbations within a couple of layers, and found that a half-billion-parameter model's effective correlation rank may be as low as eleven independent patterns.

read1 min views1 publishedAug 29, 2026

The code essentially gives you a way to rotate a Transformer's internal coordinate system into a canonical basis that aligns with its own weight matrices in a lossless way. By absorbing the normalization gains directly into the adjacent weights and using orthogonal matrices built from the singular vectors of the model, you can transform architectures like Qwen or Pythia without altering their outputs or perplexity scores.

Applying this transform reveals the actual hidden geometric structures operating inside the network. Once the model is rotated into this new perspective, you can see its internal mechanisms that were previously opaque. The authors found things like a bipolar oscillator where specific axes form inhibitory pairs that fire against each other in perfect opposition. They also observed a kind of rhythmic respiration across layers where the model alternates between absorbing knowledge and filtering it. On top of that, it exposed a homeostatic defense mechanism that aggressively erases any localized perturbations within just a couple of layers.

Practically speaking, researchers now have a powerful lens for mapping out how models actually do reasoning. For example, it turns out that the effective rank of the correlation matrix in a half billion parameter model might be as low as eleven independent patterns. Reframing how we look at the internal activations of language models provides a standardized way to study their underlying architecture.

── more in #machine-learning 4 stories · sorted by recency
tinker-docs.thinkingmachines.ai · · #machine-learning
LoRA Primer
── more on @qwen 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/canonical-basis-real…] indexed:0 read:1min 2026-08-29 ·