AlignSAE: Concept-Aligned Sparse Autoencoders

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Yang, Minglai, Guo, Xinyu, Shi, Zhengliang, Bi, Jinhe, Bethard, Steven, Surdeanu, Mihai, Pan, Liangming
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866909988353474560
author Yang, Minglai
Guo, Xinyu
Shi, Zhengliang
Bi, Jinhe
Bethard, Steven
Surdeanu, Mihai
Pan, Liangming
author_facet Yang, Minglai
Guo, Xinyu
Shi, Zhengliang
Bi, Jinhe
Bethard, Steven
Surdeanu, Mihai
Pan, Liangming
contents Large Language Models (LLMs) encode factual knowledge within hidden parametric spaces that are difficult to inspect or control. While Sparse Autoencoders (SAEs) can decompose hidden activations into more fine-grained, interpretable features, they often struggle to reliably align these features with human-defined concepts, resulting in entangled and distributed feature representations. To address this, we introduce AlignSAE, a method that aligns SAE features with a predefined ontology through a "pre-train, then post-train" curriculum. After an initial unsupervised training phase, we apply supervised post-training to bind specific concepts to dedicated latent slots while preserving the remaining capacity for general reconstruction. This separation creates an interpretable interface where specific concepts can be inspected and controlled without interference from unrelated features. Empirical results demonstrate that AlignSAE enables precise causal interventions, such as reliable "concept swaps", by targeting single, semantically aligned slots, and further supports multi-hop reasoning and a mechanistic probe of grokking-like generalization dynamics.
format Preprint
id arxiv_https___arxiv_org_abs_2512_02004
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle AlignSAE: Concept-Aligned Sparse Autoencoders
Yang, Minglai
Guo, Xinyu
Shi, Zhengliang
Bi, Jinhe
Bethard, Steven
Surdeanu, Mihai
Pan, Liangming
Machine Learning
Computation and Language
Large Language Models (LLMs) encode factual knowledge within hidden parametric spaces that are difficult to inspect or control. While Sparse Autoencoders (SAEs) can decompose hidden activations into more fine-grained, interpretable features, they often struggle to reliably align these features with human-defined concepts, resulting in entangled and distributed feature representations. To address this, we introduce AlignSAE, a method that aligns SAE features with a predefined ontology through a "pre-train, then post-train" curriculum. After an initial unsupervised training phase, we apply supervised post-training to bind specific concepts to dedicated latent slots while preserving the remaining capacity for general reconstruction. This separation creates an interpretable interface where specific concepts can be inspected and controlled without interference from unrelated features. Empirical results demonstrate that AlignSAE enables precise causal interventions, such as reliable "concept swaps", by targeting single, semantically aligned slots, and further supports multi-hop reasoning and a mechanistic probe of grokking-like generalization dynamics.
title AlignSAE: Concept-Aligned Sparse Autoencoders
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2512.02004