Transformer models encode concepts in quiet spectral regions, syntax in high-variance ones

Researchers have identified a dual geometry within transformer representations, where concept directions anti-concentrate in the spectral tail while static unembedding-row contrasts concentrate in high-variance directions. This phenomenon was observed across 17 models and 4 language pairs, with further evidence from SAE features and linear probes on Gemma and Llama. The findings suggest that transformers may move semantic content to spectrally quiet regions during processing, allowing concepts to be manipulated with less grammatical interference. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Identifies a potential mechanism for how transformers process and store semantic information, which could inform future model architectures.

RANK_REASON This is a research paper published on arXiv detailing novel findings about transformer representations. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Gemma
• Llama
• Qwen 2.5

• paper
• other