Qwen3-1.7B LoRA-trained to zigzag its 'cat' representation across layers — but it just tracks the word 'even'

Sauers_ · x · 2026-10-20

Researcher @Sauers trained a rank-16 LoRA on Qwen3-1.7B to raise its internal 'cat' representation on even layers and lower it on odd layers (or the reverse) on request, while answering unrelated questions with KL to the original model minimized.

Surprising findings:

The behavior generalizes to new questions and wordings never trained on;
The model actually responds to the tokens "even"/"odd", not the semantics — merely mentioning "even" in a sentence (e.g. "explain why the number 4 is even") triggers the zigzag, as do negations and requests about a 'dog' representation;
An untrained side effect: after a cat-zigzag-on-even-layers request, the model answers "a cat" 46% of the time when asked to name an animal (10% for odd layers, 31% baseline) — the author doesn't know why.

Code and per-finding figures are open-sourced at SauersML/zigzag on GitHub.

Original post →

Qwen3-1.7B LoRA-trained to zigzag its 'cat' representation across layers — but it just tracks the word 'even'

More from Models