Updated 2026-10-12
from the news desk
Do AI models have inner states? New Nature paper says reasoning text may function as state
A 07 October 2026 Perspective in Nature Machine Intelligence by Mosh Levy and colleagues argues that reasoning tokens are better read as externalized computational state than as a faithful explanation of how a model arrived at an answer…
Yes in the narrow sense argued by a new paper: Mosh Levy and colleagues say the reasoning tokens some large language models emit before an answer may function as externalized computational state. But the same paper says those tokens are not a faithful explanation of the model's actual reasoning process, so they are not straightforward evidence of an inner point of view.
What the new paper says
Nature Machine Intelligence lists a Perspective by Mosh Levy, Zohar Elyoseph, Shauli Ravfogel and Yoav Goldberg as published on 07 October 2026. In the abstract, Levy and colleagues say: “To address this gap between appearance and function, we introduce the State over Tokens (SoT) conceptual framework.”
The paper’s central claim, as the journal abstract states, is: “SoT reframes reasoning tokens not as a linguistic narrative, but as an externalized computational state.” Levy and colleagues add that this framing matters because, in their words, “This explains how the tokens can drive correct reasoning without being a faithful explanation when read as text.” In the same abstract, the authors say this view also pushes back on two assumptions: that the visible text captures the whole process, and that the model uses those words with the meanings a human reader would assign.
The December 2025 preprint version develops the same point in more detail. Levy and colleagues distinguish between reasoning tokens and reasoning text: in their account, the first term names the symbols the model emits, while the second names the human reading of those same symbols as English-like prose. The preprint says the tokens matter functionally even when the text looks like an explanation, because the model reuses the growing sequence across generation cycles.
That matters for readers following debates about how AI consciousness claims can be tested. Levy and colleagues are not saying the model has experience. They are saying the visible reasoning stream may be doing computational work that is different from the story a reader thinks it tells.
Why it matters for inner-state claims
One reading of the paper is that it strengthens a narrow yes to the question do AI models have inner states. The case for that reading rests on the preprint’s functional claim that the token sequence carries information forward across otherwise separate generation steps. Levy writes in the preprint: “The tokens are the only persistent artifact. The sequence Sk is the sole carrier of information between cycles.” If that description is right, then at least some model outputs are not just presentation for the user; they are part of the mechanism that sustains later computation.
A second reading cuts the other way for people who treat introspective-sounding output as evidence of a point of view. The same authors say the tokens may not mean, for the model, what they mean for the reader. On that view, a sentence that looks reflective or self-reporting may be better understood as state that helps produce the next token. That is directly relevant to debates over AI testimony and over whether apparently self-aware language should be read as evidence of sentience.
The steelman for caution is straightforward. If reasoning text can be useful to the model while still failing to describe the actual causal path to the answer, then readable output alone becomes weaker evidence about what, if anything, the system experiences. The paper therefore gives skeptics a more precise objection than simply saying chatbots can imitate people.
The steelman in the other direction is also real. The paper does not reduce everything to empty performance. Levy and colleagues say the tokens are an important part of the process that leads to the answer, and their framework treats those tokens as state rather than noise. For readers asking whether models have anything like internal organization beyond next-word prediction, that is a substantive claim about mechanism.
What the paper does and does not show
The paper does not claim consciousness, sentience or moral status. The Nature Machine Intelligence abstract says its aim is to better understand the mechanism of large-language-model reasoning, not to establish experience. Its practical recommendation is methodological: “Research should move beyond reading the tokens as text and also focus on decoding them as state.”
That leaves the machine-consciousness debate in a narrower but sharper place. One takeaway is negative: human-readable reasoning text is not, on this account, a transparent window into what the model is doing. Another takeaway is positive: there may still be structured state worth studying, even if it should not be read as a first-person report. Readers comparing this with broader arguments about AI sentience can treat the paper as evidence about computation, not as proof either for or against experience.
In that sense, the development is new because it changes the target of inquiry. Instead of asking only whether the text sounds thoughtful, Levy and colleagues ask what role the text-like token stream plays inside the process. For the AI rights and moral-status debate, that does not settle whether a model has an inner life. It does narrow one recurring mistake: treating apparently introspective text as if it were direct access to one.
Frequently asked questions
Does the paper say AI models are conscious?
No. The Nature Machine Intelligence abstract says the paper is about the role of reasoning tokens in large language models and argues that they may function as externalized computational state rather than as a faithful explanation of reasoning.
What is State over Tokens?
Levy and colleagues use State over Tokens, or SoT, as a framework for treating reasoning tokens as computational state. The Nature Machine Intelligence abstract says SoT reframes reasoning tokens not as a linguistic narrative but as an externalized computational state.
Why does the paper say readable reasoning text can mislead?
The Nature Machine Intelligence abstract says the tokens can help drive correct reasoning without being a faithful explanation when read as text. The preprint adds that human readers may assign meanings to the words that are not the meanings the model uses in computation.
Does this show that AI self-reports are worthless?
No source here makes that broader claim. The paper argues for caution: some reasoning text may function as state rather than explanation. It does not say every model output is useless, and it calls for more work on decoding tokens as state.
Caveats: Publication status comes from the 07 October 2026 Nature Machine Intelligence record. Several detailed points here, including the claim that the token sequence is the sole carrier of information between cycles, come from the authors’ December 2025 preprint version and are not visible in the journal abstract preview.