Superscopes: Amplifying Internal Feature Representations for Language Model Interpretation

2025-03-03 · Jonathan Jacobi, Gal Niv

Understanding and interpreting the internal representations of large language models (LLMs) remains an open challenge. Patchscopes introduced a method for probing internal activations by patching them into new prompts, prompting models to self-explain their hidden representations. We introduce Superscopes, a technique that systematically amplifies superposed features in MLP outputs (multilayer perceptron) and hidden states before patching them into new contexts. Inspired by the "features as directions" perspective and the Classifier-Free Guidance (CFG) approach from diffusion models, Superscopes amplifies weak but meaningful features, enabling the interpretation of internal representations that previous methods failed to explain-all without requiring additional training. This approach provides new insights into how LLMs build context and represent complex concepts, further advancing mechanistic interpretability.

📄 PDF Abstract BibTeX arXiv:2503.02078

Code (1)

GalNivs/Superscopes 공식 구현 pytorch

Tasks

Language ModelingLanguage Modelling

Methods 이 논문이 사용한 방법론

Diffusion Diffusion models generate samples by gradually removing noise from a signal, and their training objective can be expressed as a reweighted variational lower-bound…

Patching Activation patching studies the model's computation by altering its latent representations, the token embeddings in transformer-based language models, during the inference process

Sparse Autoencoders Reveal Interpretable and Steerable Features in VLA Models

2026-03-19 · Aiden Swann, Lachlain McGranahan, Hugo Buurmeijer, Monroe Kennedy 외 arxiv

Vision-Language-Action (VLA) models have emerged as a promising approach for general-purpose robot manipulation. However, little research has mechanistically explored when and why they generalize across objects, scenes, …

Robot Manipulation

The Amplifying Mirror: Locating and Steering the Partisan Direction inside a Large Language Model

2026-06-07 · Wendy K. Tam arxiv

Large language models are rapicly replacing search engines as the primary interface between people and information. Unlike search engines, which retrieve existing content, LLMs generate novel text shaped by internal repr…

Auditing Asset-Specific Preferences in Financial Large Language Models: Evidence from Bitcoin Representations and Portfolio Allocation

2026-06-01 · Wenbin Wu arxiv

Large language models now power robo-advisors and trading agents, yet whether they carry built-in biases toward specific assets is largely untested. We ask three questions: do LLMs systematically prefer certain financial…

FiLoRA: Focus-and-Ignore LoRA for Controllable Feature Reliance

2026-02-02 · Hyunsuk Chung, Caren Han, Yerin Choi, Seungyeon Ji 외 arxiv

Multimodal foundation models integrate heterogeneous signals across modalities, yet it remains unclear whether their predictions can be controlled by explicitly modulating reliance on different internal feature pathways.…

Why Does Reinforcement Learning Generalize? A Feature-Level Mechanistic Study of Post-Training in Large Language Models

2026-04-27 · Dan Shi, Zhuowen Han, Simon Ostermann, Renren Jin 외 arxiv

Reinforcement learning (RL)-based post-training often improves the reasoning performance of large language models (LLMs) beyond the training domain, while supervised fine-tuning (SFT) frequently leads to general capabili…

Reinforcement Learning