Sparse Auto-Encoders and Holism about Large Language Models

arXiv cs.AI / 3/30/2026

💬 OpinionSignals & Early TrendsIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper asks whether large language models imply a “meta-semantic” view of how words and complex expressions acquire meaning.
It reviews prior arguments that LLMs adopt a holistic conception of meaning via distributional semantics, but notes that mechanistic interpretability may challenge that view.
It introduces recent findings from sparse auto-encoders showing a large number of interpretable latent features in LLM embedding spaces, motivating a more decompositional interpretation of meaning.
The author then analyzes the nature of these features and concludes that holism can still hold if the relevant features are countable.

Abstract

Does Large Language Model (LLM) technology suggest a meta-semantic picture i.e. a picture of how words and complex expressions come to have the meaning that they do? One modest approach explores the assumptions that seem to be built into how LLMs capture the meanings of linguistic expressions as a way of considering their plausibility (Grindrod, 2026a, 2026b). It has previously been argued that LLMs, in employing a form of distributional semantics, adopt a form of holism about meaning (Grindrod, 2023; Grindrod et al., forthcoming). However, recent work in mechanistic interpretability presents a challenge to these arguments. Specifically, the discovery of a vast array of interpretable latent features within the high dimensional spaces used by LLMs potentially challenges the holistic interpretation. In this paper, I will present the original reasons for thinking that LLMs embody a form of holism (section 1), before introducing recent work on features generated through sparse auto-encoders, and explaining how the discovery of such features suggests an alternative decompositional picture of meaning (section 2). I will then respond to this challenge by considering in greater detail the nature of such features (section 3). Finally, I will return to the holistic picture defended by Grindrod et al. and argue that the picture still stands provided that the features are countable (section 4).