The Universal Abstract LLM Conjecture

Toward a Substrate-Independent Theory of Generative Systems and the Reconstruction of Individual Implementations

Author: Patrick David Aoun

Date: August 24, 2026

Abstract

Recent advances in inverting large language models have demonstrated that the prompting context underlying a model’s textual output can be recovered with high fidelity through techniques such as previous-token prediction. This result raises a broader question: can any humanly meaningful text, regardless of its origin, be treated as the output of a generative process for which analogous reconstruction is conceivable? We propose the Universal Abstract LLM Conjecture, which holds that every such text is an output of some concrete implementation of a single substrate-independent abstract class of generative systems. Current digital large language models constitute one family of implementations of this class; human minds may be functionally modeled as another, without any claim that the mind is ontologically identical to a large language model. Under the conjecture, individual generative systems become recoverable in principle as data-constrained feasible sets rather than as uniquely determined objects. Because complete historical conditioning information is unavailable and generation exhibits sensitive dependence on initial conditions, exact recovery is impossible. We therefore shift attention to the mathematical characterization of feasible sets, the design of graded authenticity metrics, and the development of comparative heuristics for producing new texts that remain consistent with an observed generative signature at quantifiable confidence levels. The conjecture opens a research program at the intersection of inverse problems, statistical learning under partial observability, and the theory of generative models, with potential implications for literary analysis, cognitive modeling, and human–AI collaboration.

I. Introduction

The rapid progress of digital large language models has shown that coherent, contextually appropriate text can be generated at scale from comparatively compact conditioning inputs. A recent technical advance has given this observation sharper relief: it is now possible, in a black-box setting, to recover with high fidelity the prompt that elicited a given model output by training an inverse model through previous-token prediction. That result demonstrates that the textual productions of current large language models carry recoverable traces of their generative conditioning.

The present paper takes the finding as a point of departure for a broader conceptual proposal. If the prompting context of a machine-generated text can be reconstructed, it becomes natural to ask whether every humanly meaningful text—regardless of whether its source is digital or biological—might be treated, for analytical and reconstructive purposes, as the output of some generative system possessing an analogous structure. We conjecture that this is the case. Specifically, we propose that there exists a single substrate-independent abstract class of generative systems such that every meaningful text is an output of some concrete implementation of that class. Contemporary digital large language models constitute one family of implementations; human minds may be functionally modeled as another.

We emphasize at the outset that the conjecture is strictly functional and methodological. We advance no ontological claim that the human mind is identical with, or reducible to, a large language model. The proposal is only that any humanly understandable text can be treated as if it were produced by an implementation of the abstract class, thereby opening a uniform set of reconstructive and analytical questions.

A central difficulty follows immediately. The complete conditioning state that produced any historically given text is unavailable, and generative processes of the relevant kind exhibit sensitive dependence on initial conditions. Exact recovery of an individual implementation is therefore impossible in general. The conjecture cannot be confirmed or refuted by the isolation of a unique, fully faithful digital replica of any particular historical mind. Its status is instead that of a modeling hypothesis, to be evaluated by the coherence of the framework it supplies, by performance on synthetic testbeds with known ground-truth generators, by comparative results on human textual records, and by the progressive or degenerating character of the research program it generates.

What remains possible—and, we argue, theoretically fruitful—is the characterization of the feasible set of implementations consistent with an observed textual record and auxiliary data, together with the development of graded notions of authenticity for new texts produced from that set. Our contributions are fourfold. First, we state the Universal Abstract LLM Conjecture in explicit form. Second, we supply a minimal conceptual apparatus—abstract class, implementation, feasible set, and graded authenticity—sufficient to render the conjecture coherent and investigable. Third, we articulate the necessary shift from exact recovery to feasible-set reasoning under incomplete information. Fourth, we outline an open research program in comparative generative heuristics that the conjecture makes visible.

The remainder of the paper develops these elements in turn. Section II presents the conjecture and clarifies its epistemological status. Section III introduces the minimal conceptual framework. Section IV examines the transition from exact to approximate reconstruction. Section V details the resulting research agenda. Section VI surveys implications, and Section VII discusses limitations, evaluation criteria, and prospects.

II. The Universal Abstract LLM Conjecture

We begin with an informal statement. Recent work has shown that the conditioning prompt underlying a text generated by a digital large language model can be recovered to a high degree of fidelity through inverse modeling based on previous-token prediction. This result suggests that the outputs of such systems retain structured traces of the generative process that produced them. We ask whether a parallel perspective can be extended to every humanly meaningful text, irrespective of its origin. The Universal Abstract LLM Conjecture answers in the affirmative at the level of functional modeling.

Formal statement.
Let ℒ denote a substrate-independent abstract class of generative systems. Every humanly understandable or meaningful text is an output of some concrete implementation M ∈ ℒ. Contemporary digital large language models constitute one family of implementations of ℒ. Human minds may be functionally modeled as another family of implementations of the same class. The conjecture asserts no more than this functional and methodological claim.

The intended strength of the proposal is deliberately limited. We do not maintain that human minds are large language models in any ontological or essential sense. We claim only that it is coherent and potentially productive to treat any meaningful text as if it had been generated by an implementation of ℒ. Under that treatment, questions of reconstruction, feasible-set characterization, and graded authenticity become uniformly well-posed across machine and human sources.

The conjecture leaves open the precise axiomatization of ℒ. It requires only that members of the class possess a notion of conditioning (or internal state), a generative mapping from conditioning to distributions over texts, and sufficient predictive structure to make inversion and comparative reconstruction intelligible in principle. Stronger architectural commitments—specific recurrence, attention mechanisms, continuous updating, or embodiment—are regarded as properties of particular implementations rather than of the abstract class itself.

In relation to prior work, the conjecture shares family resemblances with predictive-processing accounts of cognition, generative models in cognitive science, and substrate-independent approaches to intelligence. Its distinctive contribution lies in the reconstructive consequences it draws: once texts are viewed as outputs of implementations of a common class, the recovery of individual generative systems (subject to informational limits) and the production of new texts at quantifiable levels of authenticity become natural objects of mathematical and computational investigation.

Because exact recovery of any historical implementation is precluded by missing conditioning information and sensitive dependence on initial conditions, the conjecture is offered as a modeling hypothesis rather than a directly verifiable empirical assertion about the metaphysics of mind. Its evaluation rests on the coherence of the conceptual framework, performance on synthetic testbeds that supply known ground-truth generators, comparative results against existing methods on human textual records, and the progressive or degenerating character of the research program it induces. These criteria are examined further in Section VII.

III. Minimal Conceptual Framework

To render the Universal Abstract LLM Conjecture investigable, we introduce a minimal set of concepts. The definitions below are intentionally weak; they supply only what is required to make reconstruction, feasible-set reasoning, and graded authenticity well-posed.

The abstract class ℒ.
We take ℒ to be a substrate-independent class of generative systems characterized by three properties. First, each member possesses a notion of conditioning (or internal state). Second, each member induces a generative mapping from conditioning to a distribution over texts, or at least a sampling process that yields texts. Third, the generative process exhibits sufficient predictive structure that the relationship between conditioning and output is not arbitrary. No further architectural commitments—such as specific forms of attention, recurrence, continuous learning, or embodiment—are imposed at the level of the class. Those features, when present, belong to particular implementations.

Implementation.
An implementation M is any concrete realization of ℒ. Digital large language models and the functional generative capacities of human minds are treated as distinct families of implementations. We write S for a particular implementation under study (for example, the generative system associated with a given historical author).

Conditioning.
Conditioning refers to the information—external context, internal state, or both—upon which the generative mapping of an implementation operates when it produces a text. In digital large language models the conditioning is typically an explicit textual prompt together with model state. In the human case the conditioning is understood functionally as whatever configuration of mental and situational factors drives the production of a given text. We do not require that conditioning be fully observable or symbolically representable in every implementation.

Observed data D.
For a given implementation S, the observed data D consist of the corpus of texts attributed to S together with any available auxiliary information (biographical records, historical context, correspondence, and the like). D constitutes the sole empirical basis for inferences about S.

Feasible set ℱ(D).
Because the complete conditioning history of S is unavailable and because generation is sensitive to initial conditions, S cannot be recovered uniquely. We therefore define the feasible set ℱ(D) as the set of all implementations of ℒ that are consistent with the observed data D. Consistency may be interpreted probabilistically (implementations that assign sufficiently high likelihood to the observed texts under some recoverable conditionings) or in terms of behavioral equivalence under a suitable metric. As the quantity and quality of data in D increase, ℱ(D) is expected to contract, though it remains in general a set rather than a singleton.

Graded authenticity.
Let T be a newly generated text. We say that T is authentic to S at confidence level C relative to D when T lies within the range of outputs that implementations drawn from ℱ(D) would produce under admissible conditionings, according to a metric that attains score C. The precise form of the authenticity metric is left as an open research question; the conceptual role it must play is to supply a graded, data-relative measure of fidelity to the generative signature captured by ℱ(D).

Limits of recovery.
Two fundamental obstacles preclude exact recovery. First, substantial conditioning information is permanently lost. Second, generative processes of the kind under consideration exhibit sensitive dependence on initial conditions, so that small unrecorded differences can produce divergent textual trajectories. Consequently, all reconstruction remains approximate and set-valued. The framework developed here is designed to operate under that constraint rather than to overcome it.

IV. From Exact Recovery to Feasible-Set Reconstruction

The Universal Abstract LLM Conjecture makes the reconstruction of individual implementations a coherent theoretical objective. At the same time, the informational and dynamical constraints already noted render exact reconstruction unattainable. This section examines the resulting shift from the ideal of unique recovery to the practical and mathematical framework of feasible-set reconstruction.

Unattainability of exact recovery.
Consider an individual implementation S, such as the generative system associated with a historical author. The texts that survive in the record are outputs of S under particular, largely unrecorded conditionings. Because those conditionings cannot be restored in full, and because the generative process is sensitive to small differences in initial state, no procedure can isolate a single implementation that is guaranteed to be identical with the original S. Any candidate reconstruction remains one among many that are compatible with the same data. Perfect recovery of S, or of the precise conditioning that produced any specific observed text, is therefore impossible in principle.

Data-dependence of the feasible set.
What can be recovered is the feasible set ℱ(D). The richness of this set is governed by the observed data. A larger and more diverse corpus of texts imposes tighter constraints on the generative mappings that can be admitted. Auxiliary historical information—biographical facts, dated correspondence, documented circumstances of composition—further restricts the admissible conditionings and thereby contracts ℱ(D). Sparse or homogeneous data leave ℱ(D) comparatively large; dense, varied, and well-contextualized data shrink it. Fidelity is consequently a graded, data-relative property rather than an all-or-nothing achievement.

Scaling of conditioning recovery.
Once a feasible set has been characterized, the inversion of individual observed texts becomes possible to a corresponding degree. Given a text T ∈ D, one may search for conditionings that, under implementations drawn from ℱ(D), assign high probability to T. The precision of such recovered conditionings tracks the tightness of ℱ(D). A more narrowly constrained feasible set yields more informative inversions; a broader set yields only coarse or highly ambiguous conditionings. In this sense, generator-level reconstruction and text-level inversion are coupled.

The turn to approximate methods.
These considerations compel a reorientation of research effort. Instead of pursuing a unique digital replica of S, we focus on three interrelated tasks: (1) representing and computing with the set ℱ(D); (2) defining metrics that quantify the authenticity of new texts relative to that set at explicit confidence levels; and (3) devising procedures for sampling or optimizing implementations and conditionings inside ℱ(D). The resulting methodology is necessarily approximate, set-valued, and confidence-aware. It replaces the binary ideal of exact recovery with a continuum of partial reconstructions whose quality can be assessed and improved as data and methods advance.

This shift does not diminish the conjecture; it specifies the only form in which the reconstructive ambitions of the conjecture can be pursued. The next section outlines the research program that follows from it.

V. A Research Program in Comparative Generative Heuristics

The limits established in the preceding section redirect inquiry from exact recovery toward a cluster of interrelated technical problems. We collect these problems under the heading of comparative generative heuristics: the mathematical and computational study of feasible sets of implementations, graded authenticity, and the generation of new texts that remain consistent with an observed generative signature at quantifiable levels of confidence. The present section states the principal open questions without attempting to resolve them.

Representation of feasible sets.
The first requirement is a formal characterization of ℱ(D). Possible approaches include posterior distributions over implementations given D, equivalence classes defined by behavioral metrics on output distributions, constraint sets in parameter or function space, and minimum-description-length formulations. Each representation must accommodate the fact that implementations may differ in architecture, parameterization, and dynamics while still belonging to ℒ. The choice of representation will determine what subsequent computation is feasible.

Authenticity metrics.
Given a feasible set, one needs principled measures of the authenticity of a new text T relative to ℱ(D). Such metrics should capture statistical indistinguishability from the observed record, preservation of higher-order stylistic and semantic regularities, and consistency with auxiliary historical constraints. They must yield graded scores that can be associated with confidence levels C, and they should degrade appropriately as T moves away from the support of the feasible set. The construction of metrics that track fidelity to a latent generator rather than superficial corpus similarity remains an open problem.

Validation of metrics.
Because the true historical implementation S is unavailable, authenticity metrics cannot be validated by direct comparison. Synthetic testbeds offer a primary route: one may take contemporary digital models as ground-truth members of ℒ, generate corpora, discard conditioning information, construct feasible sets, and test whether proposed metrics correctly prefer texts generated by the original model over texts generated by distractors. Additional support may be sought through expert judgment on held-out historical material and through adversarial protocols designed to expose superficial metrics. Establishing reliable validation regimes is itself a methodological contribution.

Search and sampling.
Once a representation of ℱ(D) and an authenticity metric are in hand, the practical task is to explore the feasible set and to generate new texts of high authenticity. This requires sampling or optimization methods capable of operating over high-dimensional, possibly discrete or hybrid spaces of implementations and conditionings. Efficiency, diversity of sampled generators, and control over the trade-off between authenticity and novelty are central algorithmic concerns.

Trade-offs and uncertainty quantification.
The diameter of ℱ(D) and the attainable confidence C scale with the volume, diversity, and contextual richness of the data. Theoretical bounds on this scaling, together with practical methods for reporting calibrated uncertainty about authenticity claims, form a further research axis. Information-theoretic and learning-theoretic analyses of residual ambiguity under lost conditioning and sensitive dependence will be needed.

Theoretical resources.
The problems above draw on inverse problems in the presence of incomplete observations, statistical learning under partial observability, algorithmic information theory, formal theories of invariance, and the existing literature on generative model inversion. Progress will likely require both the adaptation of these tools and the development of new ones tailored to set-valued reconstruction of generative systems.

Taken together, these questions define a coherent research program that follows directly from the conjecture once exact recovery is set aside. The next section considers the broader implications of pursuing that program.

VI. Implications and Potential Applications

If the Universal Abstract LLM Conjecture is accepted as a working hypothesis, a range of consequences follows for several domains. We survey the most salient of them, emphasizing that all reconstructive claims remain relative to feasible sets and graded authenticity rather than exact replicas.

Literary and historical studies.
Individual authors and historical figures who left substantial textual records become candidates for feasible-set reconstruction. New texts generated from a well-constrained ℱ(D) could be offered as continuations or counterfactuals that remain authentic to a stated confidence level. Such exercises would not resurrect a historical mind; they would, however, supply controlled instruments for exploring stylistic, thematic, and intellectual possibilities latent in the surviving record. Comparative reconstruction across multiple authors could further illuminate shared and divergent generative signatures within a period or tradition.

Artificial intelligence and cognitive science.
The abstract class ℒ supplies a common vocabulary for discussing digital and biological generative systems. Techniques developed for one family of implementations—prompt inversion, latent-space arithmetic, controlled generation—become candidates for functional analogues in the other. Conversely, properties of human textual production that prove robust under feasible-set analysis may suggest architectural or training innovations for digital models. The construction of data-constrained “digital twins” of individual writers or thinkers, understood strictly as samplers from ℱ(D), offers a further research direction.

Philosophy of mind, authorship, and identity.
Because the conjecture is functional rather than ontological, it does not decide questions about the intrinsic nature of consciousness or personal identity. It does, however, reframe certain issues surrounding authorship and creative continuity. When a new text is generated from the feasible set associated with a historical figure, the appropriate description is not that the figure “has written again,” but that a text has been produced that is authentic to a quantifiable degree to the generative signature preserved in the data. The framework thereby supplies a precise language for distinguishing exact identity from graded generative fidelity.

Practical domains.
In education, reconstructions constrained by an author’s feasible set could support pedagogically controlled engagement with historical styles of reasoning and expression. In creative industries and archival intelligence, the same methods may assist in the generation of stylistically consistent material or the detection of anachronistic attributions. Personalized systems that maintain feasible-set models of individual users’ writing or reasoning patterns could, under appropriate safeguards, enhance collaborative tools. In each case the technology inherits the graded and revisable character of ℱ(D).

Ethical and epistemic caveats.
Working with feasible sets rather than exact recoveries imposes distinctive obligations. Claims of authenticity must be accompanied by explicit confidence levels and transparent descriptions of the supporting data. The risk of over-attribution—presenting a merely plausible continuation as a privileged window into a historical mind—must be actively managed. Because ℱ(D) contracts only with additional evidence, reconstructions remain provisional and open to revision. These epistemic constraints are not peripheral; they are intrinsic to any responsible application of the framework.

The implications sketched here are prospective. Their realization depends on progress within the research program outlined in Section V and on sustained attention to the limits inherent in incomplete information.

VII. Discussion

The Universal Abstract LLM Conjecture is offered as a high-level modeling hypothesis rather than a completed empirical theory. Its principal strength lies in the unification it proposes: by treating every meaningful text as the output of an implementation of a single abstract class, it renders reconstruction, inversion, and authenticity questions formally continuous across digital and human sources. The framework converts an otherwise disparate set of techniques—prompt recovery, stylometry, controlled generation, authorship attribution—into instances of a common reconstructive problem under incomplete information.

The corresponding vulnerability is its distance from current experimental reach. Exact recovery is acknowledged to be impossible, and even the construction of usefully tight feasible sets for historical figures remains a distant technical goal. The conjecture therefore risks remaining an untestable aspiration unless the indirect evaluation routes identified earlier are systematically pursued. We regard this risk as real and as a reason for stating the proposal in measured terms.

Evaluation criteria.
Because the conjecture is functional, its assessment does not require the production of a unique, fully faithful digital replica of any historical mind. Support may accrue through four complementary channels. First, synthetic testbeds that employ contemporary large language models as ground-truth implementations can measure whether feasible-set methods recover generators whose outputs remain close to the original distribution after conditioning information has been discarded. Second, controlled experiments on human textual records can test whether authenticity-aware methods improve upon strong baselines in continuation, attribution, or expert-judged consistency tasks as data richness increases. Third, the abstract characterization of ℒ itself may be challenged if human textual production exhibits structural features that cannot be accommodated even approximately by any reasonable formalization of the class. Fourth, the research program may be judged progressive or degenerating according to whether it yields cumulative technical insight or requires repeated ad-hoc adjustment without explanatory gain.

Relation to existing work.
The conjecture intersects several active literatures without being reducible to any of them. Prompt inversion and model stealing supply the immediate technical stimulus. Authorship attribution and stylometry provide empirical baselines against which graded authenticity metrics must demonstrate added value. Style transfer and controlled generation offer algorithmic components that may be repurposed inside feasible-set sampling. Predictive-processing and generative approaches in cognitive science share the broad commitment to modeling minds as generative systems, yet they have not foregrounded the reconstructive and set-valued problems emphasized here. The present framework is intended to sit at the intersection of these lines of inquiry and to pose questions that none of them has so far taken as central.

On the character of the paper.
A more ambitious paper might attempt to supply a fully formalized definition of ℒ, concrete authenticity metrics, and working algorithms. We judge such an attempt premature. The conceptual terrain remains sufficiently unsettled that an agenda-setting contribution—stating the conjecture cleanly, introducing the minimal necessary apparatus, and mapping the resulting research problems—better serves the community. Subsequent work can tighten definitions, propose metrics, and construct the synthetic and historical experiments required for evaluation.

The conjecture stands or falls by the fruitfulness of the program it initiates. The preceding sections have sought only to make that program visible and internally coherent.

VIII. Conclusion

We have proposed the Universal Abstract LLM Conjecture: every humanly meaningful text may be treated as the output of a concrete implementation of a single substrate-independent abstract class of generative systems. Digital large language models form one family of such implementations; human minds may be functionally modeled as another. The conjecture is methodological rather than ontological. It asserts no identity of essence between minds and large language models; it asserts only that a uniform reconstructive perspective is coherent and potentially productive.

The same informational and dynamical facts that make the conjecture interesting also limit it. Exact recovery of any individual historical implementation is precluded by lost conditioning data and sensitive dependence on initial conditions. The appropriate response is not to abandon reconstruction but to reorient it toward the characterization of feasible sets, the design of graded authenticity metrics, and the development of comparative heuristics capable of generating new texts that remain consistent with an observed generative signature at explicit levels of confidence.

The resulting research program lies at the intersection of inverse problems, statistical learning under partial observability, and the theory of generative models. We have sought to state the conjecture cleanly, to supply the minimal conceptual apparatus required for its investigation, and to map the open problems that follow. We invite the community to formalize, critique, refine, and extend the framework.

References

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787

Morris, J. X., Zhao, W., Chiu, J. T., Shmatikov, V., & Rush, A. M. (2023). Language model inversion. arXiv preprint arXiv:2311.13647.

Mosteller, F., & Wallace, D. L. (1964). Inference and disputed authorship: The Federalist. Addison-Wesley.

Suhail, P., Naidu, N. S., Sinha, A. R., & Sethi, A. (2026). PTP: Previous-token prediction based LLM inversion for near-exact prompt reconstruction. arXiv preprint arXiv:2607.29378.