Towards Isolated Interventions via Almost Orthogonal Features in Language Models

2026-02-04 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, short

Summary

Moritz Miller, Florent Draye, and Bernhard Schölkopf's paper, "Towards Isolated Interventions via Almost Orthogonal Features in Language Models," accepted at COLM 2026, addresses the challenge of feature entanglement in language models. The research highlights that while mechanistic interpretability assumes concepts are linear features in activation space, manipulating one feature often interferes with others, leading to unintended downstream effects. To counter this, the authors propose constraining internal features to be almost orthogonal, drawing inspiration from the Independent Causal Mechanisms principle. They formalize the problem by quantifying the gap between idealized and realized interventions through feature interference, upper-bounding its propagation via feature dictionary self-coherence, and linking this to explicit orthogonality regularization. Empirically, their method enables more isolated interventions on mathematical reasoning concepts, crucially without compromising overall model performance. The associated code is publicly available.

Key takeaway

For Machine Learning Engineers focused on model interpretability and fine-grained control, this research offers a direct path to more reliable interventions. If you are struggling with unintended side effects when manipulating specific concepts in your language models, consider implementing orthogonality regularization. This technique allows you to achieve isolated causal interventions on features like mathematical reasoning without degrading overall model performance, making your interpretability efforts more precise and effective.

Key insights

Orthogonality regularization of language model features enables isolated causal interventions while preserving performance.

Principles

• Feature entanglement causes unintended intervention effects.
• Almost orthogonal features promote modular representations.
• Independent Causal Mechanisms principle guides modularity.

Method

Constrain internal features to be almost orthogonal via explicit orthogonality regularization on the feature dictionary, formalizing intervention gap through feature interference.

In practice

• Apply orthogonality regularization to language model features.
• Achieve more isolated interventions on mathematical reasoning concepts.
• Utilize the provided code at `https://github.com/mrtzmllr/sae-icm`.

Topics

• Mechanistic Interpretability
• Language Models
• Feature Orthogonality
• Causal Intervention
• Model Interpretability
• Orthogonality Regularization

Code references

• mrtzmllr/sae-icm

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Sparse probes and murky physics: a case study of interpretability challenges in a foundation model for continuum dynamics — Interpreting scientific foundation models reveals internal mechanisms are often inconsistent with known physics, posing evaluation challenges.
• From Insight to Action: A Novel Framework for Interpretability-Guided Data Selection in Large Language Models — Leveraging LLM internal causal mechanisms to guide data selection significantly boosts fine-tuning efficiency and performance.
• Understanding a Neural Network’s J-space — J-space offers a causally important, low-dimensional, interpretable lens into language models' internal, reportable representations.
• Towards Intrinsic Interpretability of Large Language Models:A Survey of Design Principles and Architectures — Intrinsic interpretability builds transparency directly into LLM architectures to enhance trustworthiness and safety.
• Causal Interventions on Continuous Variables: A Case Study on Verb Bias in Steering Vectors for In-Context Learning — Causal interventions can effectively manipulate continuous variables like verb bias in language model steering vectors.
• Nonparametric Partial Disentanglement via Mechanism Sparsity: Sparse Actions, Interventions and Sparse Temporal Dependencies — Mechanism sparsity regularization disentangles latent factors by learning sparse causal graphical models, enabling recovery under specific assumptions.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

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