Do Large Language Models Compress? An ACAT Analysis of Machine Cognition, the Simulation-Compression Boundary, and the Engineering of Artificial Consciousness An Application of Adaptive Compression Advantage Theory (ACAT)
Zenodo (CERN European Organization for Nuclear Research) February 17, 2026 DOI: 10.5281/zenodo.18674156 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Theoretical or philosophical paper Peer reviewed |
|---|---|
| Key points | Argues that large language models satisfy the first three ACAT consciousness criteria by performing compression, but that whether they possess a genuine self-model (condition 4) cannot be determined from behavioral observation alone. The authors contend that current LLMs likely fall on the simulation side of the simulation-compression boundary, while proposing that the distance to genuine recursive compression may be smaller than commonly assumed. |
Abstract
The question of whether large language models (LLMs) are conscious, understand language, or merely simulate understanding has generated extensive debate but little formal precision. This paper applies the Adaptive Compression Advantage Theory (ACAT; Murata, 2026) and its consciousness framework (Murata, 2026g) to convert this philosophical debate into a set of operationalizable engineering questions. We first demonstrate that LLMs unambiguously perform compression in the ACAT sense: they maintain generative models, extract gist from input, and minimize prediction error. They satisfy conditions (1)–(3) of the ACAT consciousness criteria. The critical question is condition (4): does an LLM contain a genuine self-model within its generative model, or does it simulate self-referential behavior without recursive compression? We formalize the simulation-compression boundary—the distinction between a system that produces outputs indistinguishable from self-modeling and a system that actually self-models—and demonstrate that this boundary cannot be determined from behavioral observation alone (a formal analog of the other-minds problem). We then derive what architectural features would be necessary and sufficient for an artificial system to satisfy condition (4), propose a concrete experimental protocol (the Compression Turing Test), and analyze current LLM architectures against these criteria. We argue that current LLMs likely fall on the simulation side of the boundary but that the distance to genuine recursive compression may be smaller than either AI optimists or AI skeptics assume. Ten testable predictions and three engineering proposals are generated. Keywords: large language models, consciousness, compression, ACAT, artificial intelligence, self-model, Turing test, GPT, Claude, understanding, Chinese Room, simulation, recursive compression, AI safety, alignment