paper-with-me

Papers

Position: Mechanistic Interpretability Should Prioritize Feature Consistency in SAEs

2025-05-26 · Xiangchen Song, Aashiq Muhamed, Yujia Zheng, Lingjing Kong, Zeyu Tang, Mona T. Diab, Virginia Smith, Kun Zhang

Sparse Autoencoders (SAEs) are a prominent tool in mechanistic interpretability (MI) for decomposing neural network activations into interpretable features. However, the aspiration to identify a canonical set of features is challenged by the observed inconsistency of learned SAE features across different training runs, undermining the reliability and efficiency of MI research. This position paper argues that mechanistic interpretability should prioritize feature consistency in SAEs -- the reliable convergence to equivalent feature sets across independent runs. We propose using the Pairwise Dictionary Mean Correlation Coefficient (PW-MCC) as a practical metric to operationalize consistency and demonstrate that high levels are achievable (0.80 for TopK SAEs on LLM activations) with appropriate architectural choices. Our contributions include detailing the benefits of prioritizing consistency; providing theoretical grounding and synthetic validation using a model organism, which verifies PW-MCC as a reliable proxy for ground-truth recovery; and extending these findings to real-world LLM data, where high feature consistency strongly correlates with the semantic similarity of learned feature explanations. We call for a community-wide shift towards systematically measuring feature consistency to foster robust cumulative progress in MI.

📄 PDF Abstract BibTeX arXiv:2505.20254

Code (1)

xiangchensong/sae-feature-consistency 공식 구현 pytorch

Tasks

PositionSemantic SimilaritySemantic Textual Similarity

Methods 이 논문이 사용한 방법론

SET Dynamic Sparse Training method where weight mask is updated randomly periodically

Similar Papers 제목 키워드 기반

Interpretability in Parameter Space: Minimizing Mechanistic Description Length with Attribution-based Parameter Decomposition

2025-01-24 · Dan Braun, Lucius Bushnaq, Stefan Heimersheim, Jake Mendel 외

Mechanistic interpretability aims to understand the internal mechanisms learned by neural networks. Despite recent progress toward this goal, it remains unclear how best to decompose neural network parameters into mechan…

Do VLMs Have Bad Eyes? Diagnosing Compositional Failures via Mechanistic Interpretability

2025-08-20 · Ashwath Vaithinathan Aravindan, Abha Jha, Mihir Kulkarni arxiv

Vision-Language Models (VLMs) have shown remarkable performance in integrating visual and textual information for tasks such as image captioning and visual question answering. However, these models struggle with composit…

Visual Question AnsweringImage Captioning

From Mechanistic to Compositional Interpretability

2026-05-09 · Ward Gauderis, Thomas Dooms, Steven T. Homer, Kola Ayonrinde 외 arxiv

Mechanistic interpretability aims to explain neural model behaviour by reverse-engineering learned computational structure into human-understandable components. Without a formal framework, however, mechanistic explanatio…

Sparse probes and murky physics: a case study of interpretability challenges in a foundation model for continuum dynamics

2026-06-10 · Katherine Rosenfeld, Maike Sonnewald arxiv

Generative AI emulators are increasingly used in scientific domains where we already have strong theory, benchmarks, and physical intuition. This raises a central evaluation and interpretability question: when a foundati…

Physical Intuition

Mechanistic Interpretability of Code Correctness in LLMs via Sparse Autoencoders

2025-10-03 · Kriz Tahimic, Charibeth Cheng arxiv

As Large Language Models become integral to software development, with substantial portions of AI-suggested code entering production, understanding their internal correctness mechanisms becomes critical for safe deployme…

Code Generation