Introspective Capacity of Pretrained Generative Language Models: A No-Go Theorem under Static Pretraining and Autoregressive Generation
Type: article
Abstract: This paper provides a mathematical proof of the following proposition P : within the technical paradigm Π, namely Transformer architecture + static pretraining + autoregressive generation, no pretrained generative large language model possesses introspective capacity in the human sense."Introspection in the human sense" is operationalized as five necessary criteria (A1-A5): accuracy (the self-report agrees with the system's true internal state), grounding (the self-report depends causally on the state it reports), internality (that causal dependence does not pass through the system's own sampled outputs), metacognitive representation (the system registers, before reporting, the secondorder fact "I am in state s"), and closed-loop plasticity (an introspective event can change the system's future internal state through a private channel, not merely through public text).Under explicit assumptions (Π1-Π5) this paper proves six theorems.Theorem 1 (absence of an introspective information channel): under the paradigm's sampling model, the report R and the hidden state H are conditionally independent given the public pair (θ, c), i.e., I(R; H | θ, c) = It follows that, within paradigm Π, any system can achieve at most weak functional self-report: accuracy unguaranteed, ungrounded in the native generative regime, no private closed loop, no privileged channel, and unable to drive self-model evolution.The empirical findings of the Anthropic concept-injection paper (arXiv:2601.01828),with a success rate of about 20%, strong context dependence, zero net benefit for base models, an unverifiable metacognitive-representation criterion, and an explicit denial of human-like self-awareness, agree completely with the theorems above and are, to a large extent, their necessary corollaries.The paper also draws its boundaries precisely: every impossibility theorem presupposes paradigm Π; once any pillar is relaxed (online plasticity, a non-static corpus, a non-textual interface, an explicit verifier), the corresponding theorems cease to hold.Accordingly, this paper does not support the absolutist claim "never", but only "impossible within the current paradigm", which is exactly the precise content of proposition P .
Author: Sheng, Cheng
Language: English
Subfield: Philosophy
Field: Arts and Humanities
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7222779579