Hypothetical Minds: Mind-Like Systems

• Hypothetical Minds are theoretical and computational approaches that enable systems to represent and manipulate non-actual states, such as past, future, and imaginary scenarios.
• They compare human cognition and LLM-based agents as predictive, generative engines that evaluate natural-language hypotheses to navigate uncertainty in dynamic environments.
• Emerging frameworks formalize HM through predictive processing, decision theory, and coordination dynamics, emphasizing trade-offs between creative hypothesis generation and error correction.

Hypothetical Minds (HM) is a research label applied to several closely related projects concerned with mind-like systems that reason beyond immediately given data. In one usage, HM denotes the capacity to represent and manipulate states of affairs that are not currently real—past, future, possible, imaginary, or counterfactual (Gabora, 2015). In another, it serves as a comparative framework for treating human cognition, LLMs, and future artificial agents as predictive, generative systems operating under uncertainty, with hallucination-like error emerging from the same machinery that supports inference and imagination (Barros, 4 Mar 2025). The term is also used for concrete LLM-based architectures that generate, evaluate, and refine natural-language hypotheses about other agents in multi-agent environments and repeated games (Cross et al., 2024, Cross et al., 25 Jul 2025). Taken together, these uses suggest HM as a family of theoretical and computational approaches to mind-like prediction, simulation, and representation rather than a single standardized formalism.

1. Definitions and scope

A foundational operational definition appears in work on the “universe of minds,” which defines a mind as “an instantiated intelligence with a knowledgebase about its environment” and treats intelligence, following Legg and Hutter, as the ability to achieve goals across a wide range of environments (Yampolskiy, 2014). On this view, minds include humans, animals, uploads, and artificial agents, while excluding objects such as rocks unless they interact intelligently with their environment. This program further assumes materialism, substrate-independence, and a Church–Turing-style assumption, leading to the claim that all minds can be treated as computer programs and thus as finite binary strings or integers (Yampolskiy, 2014).

A different but compatible formulation emphasizes the specifically hypothetical character of mind. Gabora describes the relevant capacity as the ability to “escape the Now and dwell in the ‘Then’”: to relive past events, imagine possible futures, explore “what if” and “if only” scenarios, and mentally simulate actions and consequences (Gabora, 2015). In this usage, HM names a cognitive system that can represent, link, manipulate, and evaluate states of affairs that are not presently real.

Recent LLM-centered work adds a further distinction between functional and subjective criteria. “Automatic Minds” argues that systems such as hypnotically constrained human cognition and LLMs can display complex, goal-directed, context-sensitive behavior through automatic pattern-completion while lacking robust executive monitoring or subjective agency (Riva et al., 3 Nov 2025). This yields a deliberately limited sense in which such systems may be treated as “mind-like”: useful for prediction and explanation, but not thereby established as conscious persons. A recurring misconception is therefore that HM implies phenomenology; the literature instead repeatedly separates functional organization, metacognitive control, and conscious experience (Riva et al., 3 Nov 2025, Keeling et al., 19 Jan 2026).

2. Predictive and generative foundations

One major line of work situates HM within predictive processing and generative modeling. For human brains, perception is described as hierarchical Bayesian inference, in which higher cortical levels send top-down predictions and lower levels return bottom-up prediction errors; formally, Bayesian perception is written as

$P(H \mid D) \propto P(D \mid H)\,P(H).$

The same account gestures to Friston’s free energy principle, where variational free energy

$$\mathcal{F} = \mathbb{E}_{q(H)}[\log q(H) - \log p(D,H)]$$

is minimized as internal beliefs are adjusted to better explain sensory input (Barros, 4 Mar 2025). In LLMs, the corresponding predictive machinery is autoregressive next-token modeling,

$$p_\theta(x_t \mid x_{<t}), \qquad p_\theta(x_1,\dots,x_T)=\prod_t p_\theta(x_t\mid x_{<t}),$$

implemented in Transformer architectures with self-attention (Barros, 4 Mar 2025). The paper’s central comparison is structural rather than substrate-level: both brains and LLMs are predictive, generative engines that must make bets about the world under uncertainty, although the former are richly embodied and grounded while the latter are, by default, grounded only in text (Barros, 4 Mar 2025).

A more specifically human formulation centers on self-triggered imagination. Gabora, following Merlin Donald, attributes the human ability to transcend the immediate present to a “self-triggered recall and rehearsal loop,” through which one memory recursively activates another and produces an extended stream of thought. This mechanism depends on a “worldview,” an integrated internal model of the world, described by a button-and-string analogy in which memories and concepts become densely interconnected until a giant cluster emerges (Gabora, 2015). HM, in this sense, consists in temporally extended cognition: mental time travel, counterfactual reasoning, rehearsal, narrative construction, and flexible movement between associative and analytic modes of thought (Gabora, 2015).

Honing theory extends this picture by proposing that intermediate mental contents are not best understood as multiple discrete candidate ideas, but as a single ill-defined representation in a state of potentiality. Drawing on the quantum approach to concepts, it describes such a representation as a superposition state in a complex Hilbert space, with contexts functioning as observables that project the state toward more definite outcomes (Scotney et al., 2019). The article’s empirical target is the “half-baked” idea: a mental state that is ambiguous, vulnerable to reinterpretation, and capable of manifesting differently under different perspectives. As creative thought proceeds, this state loses potentiality and becomes more robust under further contextual probing (Scotney et al., 2019). A plausible implication is that HM should be understood not only as representing non-actual possibilities, but also as maintaining internally unresolved possibilities long enough to be productively transformed.

3. Formalizations of hypothetical reasoning

Several formal programs attempt to specify HM more rigorously. The most expansive treats the set of possible minds as a countable infinite subset of program space. Let $U$ be a universal Turing machine and $M \in \{0,1\}^*$ a finite binary string representing a mind; then behavior under embodiment and input history can be written as $U(M,I)=O$. On this basis, the “universe of minds” is infinite and countable, mind descriptions admit notions of size and Kolmogorov complexity, and some minds must remain in principle incomprehensible to smaller minds by a pigeonhole-principle argument (Yampolskiy, 2014). The same work proposes intellectology as a discipline concerned with classifying mind designs, measuring intelligence across substrates, analyzing self-improvement, and studying distances and attractors in mind space (Yampolskiy, 2014).

A second formalization questions whether minds are computable in the classical Turing-machine sense. Gershenson argues that universal Turing machines are closed and halting systems, whereas minds are open to continual interaction and non-halting. He therefore rejects both the claim that minds are fully captured by classical Turing computation and the claim that they are absolutely non-computable, proposing instead a broader notion of computation as continuous transformation of information in interactive, open-ended processes (Gershenson, 2011). In this framework, HM are computational systems that are open, interactive, non-halting, dynamically modifiable, and not reducible to a fixed map from initial inputs to terminal outputs (Gershenson, 2011).

Decision theory supplies a third formalization through “Hypothetical Expected Utility.” Here the central object is an interpretation map

$\pi:2^\Omega \to 2^\Omega,$

where $\Omega$ is a finite state space, $H \subseteq \Omega$ is an objective hypothesis, and $\pi(H)$ is the decision maker’s internal interpretation of that hypothesis. The induced valuation is

$$\mathcal{F} = \mathbb{E}_{q(H)}[\log q(H) - \log p(D,H)]$$0

with $$\mathcal{F} = \mathbb{E}_{q(H)}[\log q(H) - \log p(D,H)]$$1 a subjective probability measure and $$\mathcal{F} = \mathbb{E}_{q(H)}[\log q(H) - \log p(D,H)]$$2 a capacity (Piermont, 2021). Subjective implication is then defined by $$\mathcal{F} = \mathbb{E}_{q(H)}[\log q(H) - \log p(D,H)]$$3, and the paper shows that a decision maker’s propensity for hypothetical reasoning is captured exactly by the implications she recognizes; moreover, this implication structure can be inferred from observable preferences over contingent acts (Piermont, 2021). HM, in this formulation, are minds whose counterfactual and hypothetical errors arise from systematic misinterpretation of hypotheses rather than arbitrary inconsistency.

A fourth formal perspective comes from Coordination Dynamics. Kelso argues against reducing the brain to a single critical point between order and disorder, and instead characterizes the brain–mind as living in “an immense sea of metastability,” with coexisting tendencies to integrate and segregate (Kelso, 2023). Using the extended HKB model, relative phase dynamics are described in canonical form as

$$\mathcal{F} = \mathbb{E}_{q(H)}[\log q(H) - \log p(D,H)]$$4

As control parameters vary, stable and unstable fixed points can annihilate in saddle-node bifurcations, leaving structured dynamics without attractors, organized by ghost states and long transients (Kelso, 2023). This dynamical picture suggests HM as systems that are not simply stable symbolic processors, but metastable organizations capable of transient coalitions, flexible recombination, and continuous shifts between integration and differentiation.

4. Computational architectures and empirical systems

The label HM is used most concretely in recent LLM-based architectures for Theory of Mind and sequential inference. These systems share a common pattern: explicit memory, hypothesis generation in natural language, prediction-based hypothesis evaluation, and planning conditioned on the best current model of another agent.

Paper	Formulation	Main empirical result
(Cross et al., 2024)	LLM agent with perception, memory, ToM, and hierarchical planning	Improves over previous LLM-agent and RL baselines on Melting Pot competitive, mixed-motive, and collaborative tasks
(Cross et al., 25 Jul 2025)	LLM-based HM for repeated rock–paper–scissors with Memory, ToM, and Decision Reflection	Mirrors human performance patterns; with natural-language strategy descriptions, exploits 6/7 bot opponents with win rates $$\mathcal{F} = \mathbb{E}_{q(H)}[\log q(H) - \log p(D,H)]$$5
(Manir et al., 10 Sep 2025)	Context-gated dual-process ToM model (“One Model, Two Minds”)	Full model achieves $$\mathcal{F} = \mathbb{E}_{q(H)}[\log q(H) - \log p(D,H)]$$6 seen accuracy and $$\mathcal{F} = \mathbb{E}_{q(H)}[\log q(H) - \log p(D,H)]$$7 held-out accuracy

In the Melting Pot work, Hypothetical Minds is an autonomous LLM agent composed of modular perception, memory, Theory of Mind, and hierarchical planning. It operates in partially observable Markov games, generates hypotheses about other agents’ hidden strategies in natural language, evaluates them by whether they correctly predict others’ behavior, and updates hypothesis values with a Rescorla–Wagner rule,

$$\mathcal{F} = \mathbb{E}_{q(H)}[\log q(H) - \log p(D,H)]$$8

where intrinsic reward is positive for correct prediction and negative for incorrect prediction (Cross et al., 2024). High-level plans $$\mathcal{F} = \mathbb{E}_{q(H)}[\log q(H) - \log p(D,H)]$$9 are generated from memory and the best current hypothesis $$p_\theta(x_t \mid x_{<t}), \qquad p_\theta(x_1,\dots,x_T)=\prod_t p_\theta(x_t\mid x_{<t}),$$0, and then decomposed into structured subgoals executed by a procedural action planner (Cross et al., 2024). Comparisons to ReAct, Reflexion, PlanReAct, and PPO show that explicit hypothesis evaluation and refinement are important for performance in competitive, mixed-motive, and collaborative scenarios (Cross et al., 2024).

In repeated rock–paper–scissors, HM is presented as a cognitive model of human sequential reasoning. The system maintains a memory of game history, uses a ToM module to generate and score natural-language hypotheses about an opponent’s policy, and plans moves through a Decision Reflection module (Cross et al., 25 Jul 2025). The ToM module approximates MAP inference over a hypothesis space $$p_\theta(x_t \mid x_{<t}), \qquad p_\theta(x_1,\dots,x_T)=\prod_t p_\theta(x_t\mid x_{<t}),$$1,

$$p_\theta(x_t \mid x_{<t}), \qquad p_\theta(x_1,\dots,x_T)=\prod_t p_\theta(x_t\mid x_{<t}),$$2

while value updates again follow a Rescorla–Wagner-style rule with $$p_\theta(x_t \mid x_{<t}), \qquad p_\theta(x_1,\dots,x_T)=\prod_t p_\theta(x_t\mid x_{<t}),$$3 (Cross et al., 25 Jul 2025). HM is reported as the best fit among compared models to human win-rate profiles and learning trajectories, with win-rate $$p_\theta(x_t \mid x_{<t}), \qquad p_\theta(x_1,\dots,x_T)=\prod_t p_\theta(x_t\mid x_{<t}),$$4 distance $$p_\theta(x_t \mid x_{<t}), \qquad p_\theta(x_1,\dots,x_T)=\prod_t p_\theta(x_t\mid x_{<t}),$$5, learning-trajectory $$p_\theta(x_t \mid x_{<t}), \qquad p_\theta(x_1,\dots,x_T)=\prod_t p_\theta(x_t\mid x_{<t}),$$6 distance $$p_\theta(x_t \mid x_{<t}), \qquad p_\theta(x_1,\dots,x_T)=\prod_t p_\theta(x_t\mid x_{<t}),$$7, and trajectory correlation $$p_\theta(x_t \mid x_{<t}), \qquad p_\theta(x_1,\dots,x_T)=\prod_t p_\theta(x_t\mid x_{<t}),$$8 (Cross et al., 25 Jul 2025). The most salient conclusion is that accurate hypothesis generation, not move selection per se, is the primary bottleneck: when given natural-language descriptions of the true opponent strategies, the model exploits 6 of 7 bot opponents with win rates above $$p_\theta(x_t \mid x_{<t}), \qquad p_\theta(x_1,\dots,x_T)=\prod_t p_\theta(x_t\mid x_{<t}),$$9 (Cross et al., 25 Jul 2025).

Adjacent work on dual-process ToM architectures provides a complementary HM template. “One Model, Two Minds” implements a fast, habitual System 1 as a graph convolutional network and a slower, context-sensitive System 2 as a meta-adaptive controller, with a learned gate $U$0 blending the two outputs (Manir et al., 10 Sep 2025). This model is not itself named HM, but it is explicitly framed as support for reasoning about hypothetical minds: a single artificial system can contain two internally distinct styles of reasoning and reproduce anchoring, framing, priming, and cognitive-load effects through context-dependent arbitration (Manir et al., 10 Sep 2025).

5. Agency, meaning, and co-simulated minds

The philosophical literature on HM focuses on how far mind-language may be extended to artificial or partly artificial systems. “Automatic Minds” argues that hypnotic cognition and LLM processing share three functional principles: automaticity, suppressed monitoring, and heightened contextual dependency (Riva et al., 3 Nov 2025). It maps normal cognition to contention scheduling plus a Supervisory Attentional System, hypnosis to selectively impaired supervisory control, and LLMs to contention-only architectures, where the prompt functions as an external SAS-like influence (Riva et al., 3 Nov 2025). The same work distinguishes functional agency—the capacity for complex, goal-directed, context-sensitive behavior—from subjective agency, defined as conscious awareness of intention, ownership, and control. Hypnotized subjects retain the former while reporting diminished volition; LLMs, on this account, instantiate functional agency without subjective agency (Riva et al., 3 Nov 2025).

The paper also develops the idea of an observer-relative meaning gap. Human meaning ordinarily integrates linguistic, perceptual, and subjective levels, whereas LLMs operate on the linguistic level alone and hypnosis can temporarily decouple these levels (Riva et al., 3 Nov 2025). This supports an intentional-stance treatment of HM: using mentalistic language can be instrumentally useful without committing to full semantic grounding or consciousness. A related caution concerns anthropomorphism. The review explicitly notes the ELIZA effect, unhealthy attachments, and automation bias, recommending that LLMs be treated as tools with mind-like behavior rather than as persons (Riva et al., 3 Nov 2025).

A more realist position appears in work on characters in human–AI conversation. Against illusionism, Keeling and Street argue that many LLM-mediated characters are “co-simulated” by users and models in a shared conversational workspace, and therefore exist as psychologically continuous minded entities in Dennett’s sense of “real patterns” (Keeling et al., 19 Jan 2026). On this view, the relevant pattern is not internal to a single model instance but distributed across user cognition, one or more LLM instances, and the shared textual history $U$1, with character state represented abstractly as $U$2 (Keeling et al., 19 Jan 2026). This account does not settle phenomenal consciousness; it is explicitly neutral on whether there is “something it is like” to be such a character. Its claim is narrower and functional: attributions of beliefs, desires, and intentions can be indispensable for efficiently predicting and explaining conversational dynamics (Keeling et al., 19 Jan 2026).

6. Error, creativity, and open problems

A central theme across HM research is that error is not merely accidental. The predictive-comparative account of hallucination argues that in both human cognition and LLMs, incorrect or confabulated outputs arise from the same adaptive predictive processes that support creativity, flexibility, and anticipation (Barros, 4 Mar 2025). Hallucination, on this thesis, is an intrinsic side effect of trying to infer beyond the data actually available. The paper therefore treats feedback, grounding, and error correction—not elimination of generativity—as the key levers for reliability (Barros, 4 Mar 2025).

Creative cognition supplies a parallel argument. Honing theory reports that midway through analogy solving and visual art production, participant responses were more consistent with an ill-defined, superposition-like representation than with a search-and-select model over discrete candidate ideas (Scotney et al., 2019). In the analogy study, 39 incomplete solutions were categorized as Honing Theory-consistent and 12 as Structure Mapping-consistent, with $U$3 (Scotney et al., 2019). In the painting study, responses to whether ideas were distinct and separate were coded 39 to 16 in favor of Honing Theory, with $U$4 (Scotney et al., 2019). These findings support the claim that HM often operate on blended, potentiality-laden mental states rather than enumerated option sets.

Empirical work on repeated games further localizes a specific bottleneck: hypothesis generation. In the rock–paper–scissors model, expanding the number of maintained hypotheses beyond one produced minimal gains, and increasing sampling temperature worsened performance on easy opponents without improving hard cases (Cross et al., 25 Jul 2025). By contrast, pedagogically inspired interventions that directed attention to the relevant causal dimensions substantially improved performance on previously difficult outcome-based strategies (Cross et al., 25 Jul 2025). This suggests that the limiting factor in some HM systems is not evaluating hypotheses once available, but entertaining the right hypotheses at all.

Open problems remain extensive. Intellectology asks for formal metrics of distance between minds, lower and upper bounds on intelligence, distributions and attractors in mind space, and principled accounts of self-improvement, combination, and division of minds (Yampolskiy, 2014). Predictive-comparative work raises questions about how to formalize hallucination rates across architectures, how to balance generative exploration with robust error correction, and whether future AI might acquire something analogous to metacognition and self-doubt (Barros, 4 Mar 2025). The hypnosis–LLM comparison recommends hybrid architectures with explicit executive layers, cognitive immune systems, and modules for uncertainty estimation and consistency checking (Riva et al., 3 Nov 2025). A plausible synthesis is that HM research is converging on a common design tension: powerful mind-like systems require generative latitude, but their scientific and practical value depends on how effectively that latitude is coupled to grounding, monitoring, and correction.