arXiv cs.CL· Minglai Yang, Xinyu Guo, Zhengliang Shi, Jinhe Bi, Steven Bethard, Mihai Surdeanu, Liangming Pan·· 2 天前AI 评分24
AlignSAE：概念对齐的稀疏自编码器
AlignSAE: Concept-Aligned Sparse Autoencoders
AI 导读
AlignSAE 通过“先预训练、后后训练”的两阶段方法，将稀疏自编码器（SAE）特征与预定义本体对齐，在 GPT-2 上实现精确因果干预，1-hop 概念替换成功率达 94.8%。其解耦效果较标准 SAE 提升 9 至 ∞ 倍（以互信息差衡量），并支持多跳推理与类 grokking 泛化动态的机制探针分析。该方法与模型无关，已发表于 TMLR。
正文 · 原文
Abstract: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 on GPT-2 across synthetic biographical QA (1-hop, 6 concepts) and compositional relational reasoning (2-hop, 20 concepts) demonstrate that AlignSAE enables precise causal interventions, such as reliable concept swaps (94.8% success in 1-hop), by targeting single, semantically aligned slots. AlignSAE achieves disentanglement improvements of $9$--$\infty\times$ over standard SAEs (measured via Mutual Information Gap), and further supports multi-hop reasoning and a mechanistic probe of grokking-like generalization dynamics. Although our evaluation focuses on GPT-2, AlignSAE itself is model-agnostic.
| Comments: | Published in Transactions on Machine Learning Research (08/2026). 36 pages, 18 figures, 12 tables. v2: adds an Errata section clarifying a few issues in the TMLR version |
| Subjects: | Machine Learning (cs.LG); Computation and Language (cs.CL) |
| Cite as: | arXiv:2512.02004 [cs.LG] |
| (or arXiv:2512.02004v4 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2512.02004 arXiv-issued DOI via DataCite |
|
| Journal reference: | Transactions on Machine Learning Research (TMLR), 2026. https://openreview.net/forum?id=I9UjKxW4nq |
Submission history
From: Minglai Yang [view email]
[v1]
Mon, 1 Dec 2025 18:58:22 UTC (2,207 KB)
[v2]
Sat, 10 Jan 2026 19:34:02 UTC (7,015 KB)
[v3]
Tue, 13 Jan 2026 02:52:14 UTC (7,015 KB)
[v4]
Thu, 15 Oct 2026 16:55:22 UTC (8,892 KB)
来源：arXiv cs.CL · arxiv.org