Disentangling Direction and Magnitude in Transformer Representations: A Double Dissociation Through L2-Matched Perturbation Analysis

2026-02-13 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, extended

Summary

A study on Pythia-family models (410M-1.4B parameters) investigates whether the direction (orientation) and magnitude (L2 norm) of Transformer hidden states serve distinct functional roles. Using L2-matched perturbation analysis, a method ensuring identical Euclidean displacements for both perturbation types, the research reveals a "cross-over dissociation." Angular perturbations increase language modeling loss by up to 42.9 times more than magnitude perturbations, while magnitude perturbations cause a disproportionately larger drop in syntactic processing accuracy (20.4% vs. 1.6% on subject-verb agreement at \delta=5). Causal intervention experiments further show that angular damage primarily flows through attention pathways (28.4% loss recovery via attention repair), whereas magnitude damage is partly mediated by LayerNorm pathways (29.9% recovery via LayerNorm repair). These patterns replicate across scales within the Pythia architecture but show different behaviors in RMSNorm-based architectures like TinyLlama-1.1B, suggesting architectural choices influence these functional roles.

Key takeaway

For research scientists working on Transformer interpretability or model editing, you should consider the distinct functional roles of vector direction and magnitude. When modifying model behavior, preserve direction for language modeling quality and factual knowledge, and preserve magnitude for syntactic coherence. This approach enables more precise interventions with fewer unintended side effects, especially in LayerNorm-based architectures.

Key insights

Direction and magnitude of Transformer hidden states have distinct computational roles in LayerNorm-based architectures.

Principles

• Direction governs attentional routing.
• Magnitude modulates processing intensity for syntax.
• Normalization type affects geometric property roles.

Method

L2-matched perturbation analysis ensures angular and magnitude perturbations achieve identical Euclidean displacements, enabling controlled comparison of their functional importance in Transformer hidden states.

In practice

• Analyze direction and magnitude separately for interpretability.
• Target edits to direction for factual knowledge.
• Target edits to magnitude for syntactic coherence.

Topics

• Transformer Representations
• L2-Matched Perturbation Analysis
• Language Model Interpretability
• Layer Normalization
• Syntactic Processing

Best for: Research Scientist, AI Researcher, AI Scientist, NLP Engineer

Related on AIssential

• Give it Space! Explicit Disentangling of Positional and Semantic Representations in Encoders — Explicitly disentangling positional and semantic representations in Transformers improves positional encoding and linguistic understanding.
• Algebraic Dead Directions in LayerNorm Transformers: A Forward-Pass-Only Diagnostic at LLM Scale — A novel diagnostic identifies algebraic dead directions in LayerNorm transformers from parameters alone, without forward/backward passes.
• An expressivity analysis of hierarchical modelling in deep transformers via bounded-depth grammars — Deep transformers can formally represent hierarchical structures using bounded-depth grammars, encoding states in linearly separable subspaces.
• Understanding Cross-Modal Contributions in Continual Vision-Language Models: A Theoretical Perspective — A new theoretical perspective clarifies cross-modal contributions in continual vision-language models, addressing catastrophic forgetting.
• Trajectory Geometry of Transformer Representations Across Layers — Trajectory geometry offers a principled, probe-free lens to understand Transformer representation evolution across layers.
• Bölüm 1: Yapay Zekada Bir Devrin Kapanışı — Transformer Öncesi “Sequential” Kısıtlar — The Transformer architecture revolutionized NLP by enabling parallel processing and global dependency modeling through attention, overcoming sequential constraints.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.