Representationalism and Enactivism in Cognitive Translation Studies: A Predictive Processing Perspective

Abstract

Representational Theories of Mind have long dominated Cognitive Translation Studies, typically assuming that translation involves the manipulation of internal representations (symbols) that stand in for external states of affairs. In recent years, classical representationalism has given way to more nuanced, inferential, interpretive, context-sensitive, and modern representational models, some of which align naturally with probabilistic and predictive approaches. While these frameworks remain broadly compatible with one another, radical enactivism offers a more disruptive alternative: it denies representational content altogether, viewing translation instead as an affectively grounded, context-sensitive, self-evidencing activity shaped by the translator’s embodied engagement with text, context, and sociocultural norms. From an enactivist standpoint, translation emerges not from static symbolic mappings, but from situated, embodied, and affectively modulated inference processes that dynamically negotiate meaning across languages. The paper provides a theoretical synthesis, arguing that the Free Energy Principle under Predictive Processing and Active Inference provides a suitable mathematical framework amenable to representational and enactive accounts.

1. Introduction

For much of the history of Translation Studies (TS), representational theories of mind formed the dominant backdrop for modelling translation processes. In classical cognitive frameworks, understanding and producing translations are conceived as operations on internal, symbolic representations of linguistic meaning. These representations were thought to be relatively stable mental structures that the translator accessed, manipulated, and transformed in order to produce an equivalent text in the target language. The translator’s mind was thus likened to an information processor: sensory input (the source text (ST)) was encoded into an internal format, transformed according to linguistic and cultural rules, and then decoded into output (the target text (TT)).

Around 2010, however, embodied, embedded, enacted, and extended approaches gained traction in TS [1,2]. Drawing from developments in cognitive science, these perspectives suggested that translation is not merely an internal, representational/computational activity, but a process situated in the translator’s body, physical environment, and social–technological context. They challenged the idea that cognition is confined to the manipulation of abstract symbols inside the head. Instead, they stressed the bodily basis of cognition (embodied), its reliance on environmental scaffolds (embedded), the idea that cognition is something we perform rather than something that merely happens inside the mind (enacted), the extension of cognitive processes into tools and artifacts (extended), and the centrality of emotion in shaping thought (affective). These perspectives resonated with empirical findings in cognitive TS, showing the role of translators’ physical engagement with their working environment, their interaction with translation tools, and the affective states shaping decision-making [2] (Risku 2020).

It is helpful to distinguish classical representationalism from modern representationalism [3]. Classical representationalism, inspired by symbolic AI, posited static, language-like mental representations manipulated by formal rules (e.g., Fodor’s Language of Thought, [4]). Meaning would be stored and processed in a relatively context-independent way. Modern representationalism, instead, emphasizes probabilistic, context-sensitive internal models [3]. These representations are not fixed symbols but inference-driven states that integrate sensory evidence with prior expectations.

For example, an embodied view highlights the role of perceptual-motor skills, such as the fluid coordination between eye movements, keystrokes, and attention. An extended view includes the translator’s interaction with computer-assisted translation (CAT) tools and online resources as integral parts of their (representational) cognitive system.

Importantly, these approaches are not necessarily incompatible with classical representationalism. Consider a translator working with a complex CAT interface where the translator’s cognitive processes rely heavily on the layout, memory suggestions, and interactive features of the interface. Mental representationalism may be considered central. A translator using a termbase in a CAT tool might still internally represent the term’s meaning and possible equivalents; the novelty in modern representationalism lies in acknowledging that these representations are dynamically supported and constrained by external artifacts, bodily routines, and environmental affordances. As the activation and deployment of translation units, probabilistic matches, and linguistic context models are shaped by environmental features, the representational content is here contextually embedded, but not eliminated.

Relevance Theory (RT, [5]) is a case in point that enjoys increasing popularity in cognitive TS. As a theory of communication, RT is an example of modern representationalism, which defines translation as an instance of “interpretive language use” [6]. Interpretive resemblance is understood not as propositional equivalence but as similarity in contextual effect; the translator seeks to establish interpretive resemblance between the ST and the TT. Interpretive resemblance is assumed to be achieved through mind-reading: inferring the speaker’s beliefs and intentions, and by reconstructing the original communicative goal. RT presupposes a Higher-Order Theory of Mind (HOToM, [7,8]), which posits that understanding others—necessary for translation—requires forming metarepresentations of mental states. RT also aligns with a Higher-Order Theory of Thought (HOToT, [9,10]), a model of consciousness that assumes mental states become conscious only when accompanied by higher-order thoughts about those states. Higher-Order Theories rely on representational assumptions, whereby cognition is defined by the manipulation and nesting of contentful mental representations [3].

More recently, enactivist approaches have been suggested in TS [11]. Enactivists claim that cognition does not first of all depend on internally stored representations of external state of affairs [12], but emerges directly from the dynamic interaction between an organism and its environment [13]. For example, a translator improvising a solution to a novel idiom might do so not by consulting an internal representation, but by exploring possible phrasings interactively with their tools, the emerging text, and online corpora—treating the translation as a situated, lived activity rather than an internal computational problem. Here, “knowledge” exists in the translator’s skills and dispositions to act rather than in stored information structures.

Radical enactivism [14,15,16] rejects the view that cognition requires contentful internal representations. Instead, cognition is seen as emerging from embodied, affective, and context-sensitive interactions with the environment. Mental states, including those about oneself (HOToT) and others (HOToM), need not be constructed via propositional inference; they can be enacted through embodied social engagement, emotional attunement, and pragmatic responsiveness. This suggests a distributed and interactive model of cognition where HOToM and HOToT are scaffolded by cultural practices, narrative engagement, and emotional resonance rather than by abstract metarepresentational processing.

Kiverstein [17] argues that “body enactivism” is incompatible with representationalism. From this perspective, any appeal to internal content-bearing structures undermines the enactivist claim that cognition is inherently pragmatic, world-involving, and non-representational. Kiverstein’s position emphasizes that meaning is enacted in real time through bodily action and environmental engagement, rather than pre-stored and then retrieved.

In the translation contexts, computational enactivism might thus model translation competence as emerging from the dynamics of translator–text coupling rather than from accessing stored linguistic knowledge. Instead of representing ST meanings that are mapped into TT equivalents, an enactive computational model might simulate the temporal dynamics of meaning emergence through embodied engagement with textual materials.

Although representationalism and enactivism seem to mutually exclude each other, Predictive Processing (PP)—grounded in Active Inference (AIF) and the Free Energy Principle (FEP)—offers a framework to account for these opposed positions. PP [18,19] and AIF [20,21] are suited frameworks to model these embedded translation processes within a Bayesian approach (cf. also [11]). Together they open a shared formalism that allows for different interpretation possibilities, thus amenable to a weak integration.

PP conceptualizes the brain to constantly generate predictions about sensory input, comparing them with incoming signals, and updating its internal model to minimize prediction error. The framework accommodates modern representationalism, i.e., probabilistic generative, dynamic, and/or embodied models that are shaped through sensorimotor engagement. The PP framework is also compatible with enactive insights in which action and perception are continuous, reciprocal (or mutually inseparable) processes aimed at bringing about sensory states that match predictions. Translation becomes then a “self-evidencing” activity [18,20,22], in which translators enact their own cognitively and affectively biased predictions.

The PP/AIF framework thus allows for modelling embodied, embedded representational approaches but also non-representational enactive ones. Similarly, Nikolova [23] stipulates that “4E Cognition and PP complement each other, providing an explanation of cognitive processes through different but compatible conceptual apparatuses. ”

From this vantage point, translation can be seen as a form of epistemic action—actively seeking information to refine predictions about meaning and form—and pragmatic action—shaping the emerging target text to align with communicative goals. The translator’s body, environment, tools, working conditions, and affective states are all integral parts of the process, as well as the internal probabilistic models that enable anticipation, adaptation, and learning.

2. Representational View

Relevance Theory (RT) is a modern representational theory of communication, developed by Sperber and Wilson [5], that explains how listeners infer a speaker’s intended meaning by relying on the principle of cognitive relevance—namely, the expectation that communication will yield maximal informational benefit for minimal processing effort. Central to RT is the notion that communication is ostensive-inferential, that individuals mutually recognize each other’s communicative intentions. For successful communication to occur, RT holds that the content must be mentally represented, cognitively accessible, and relevant to the audience. Speakers engage in deliberate acts of ostensive communication, intentionally making aspects of their mental states manifest to others.

RT conceptualizes both cognition and communication as inherently representational processes, rooted in the mind’s capacity to form and manipulate internal models of states of affairs in an external world. As Sperber (p. 3, [24]) notes, cognitive systems are defined by their ability to “construct and process mental representations.” These representations serve as internal stand-ins for external states of affairs. Beyond first-order representations, RT also posits metarepresentations—mental representations of other mental states. For instance, the belief “John believes it will rain” exemplifies our ability to attribute mental content to others, a capacity foundational to Theory of Mind and higher-level pragmatic understanding.

2.1. Theory of Mind

Theory of Mind (ToM, [25]) is our ability to attribute mental states—such as beliefs, desires, intentions, and emotions—to others and to ourselves. It is the capacity to understand that other people have their own beliefs, desires, and intentions, which can differ from ours, and to use this understanding to explain and predict their behavior. According to Premack and Woodruff[25], an individual has a theory of mind if he imputes mental states to himself and to others.

ToM suggests that mental processes fundamentally involve internal representations—structured, contentful symbols or analogues that stand for things in the world. Rosenthal [9], Carruthers [10], and other ToM theorists generally support a representational model of the mind, rooted in cognitive science and analytic philosophy. Some theorists distinguish between two modes of representations: cognitive—concerning mental states, beliefs, thoughts, and intentions—and affective—concerning the emotions of others [26,27].

However, Gutt’s [28,29] conception of interpretive language use already hints at a deviation from representationalism. Rather than prioritizing the preservation of truth-conditional content or nested propositions, Gutt emphasizes the normative and functional dimension of translation: translation establishes, according to him, a relation of resemblance between “bodies of thought,” [29], i.e., culturally and pragmatically situated intentions, rather than propositional truths. Gutt suggests two translation modes, a stimulus mode and an interpretive mode, but neither mode assumes that a source utterance’s content can be fully captured by propositional structure alone. The I-Mode, especially, focuses on interpretive resemblance of thoughts, which can involve norms, emotions, and cultural embeddings—none of which are exhaustively representational. This opens the door to a more dynamic and interactive perspective that will be developed in this paper.

2.2. S-Mode and I-Mode Translation

Gutt [28] characterizes translation as a communicative act that engages with a prior communicative act. According to him, “any act of communication concerned with another act of communication… can aim at providing information about either of its two key elements: the stimulus used or the interpretation intended in the original act” (p. 4, [28]). In this framework, translation can focus either on the stimulus—that is, the linguistic form or “what is said”—or on the interpretation, the intended meaning or “what is meant.”

When a translator reproduces surface features of the original utterance—such as its form, syntax, or stylistic elements—he operates in what Gutt terms the S-Mode (stimulus mode). Conversely, when the goal is to render the speaker’s intended meaning, especially where it diverges from surface form, the translator shifts into the I-Mode (interpretation mode). As Gutt (p. 35, [28]) further explains, “the stimulus is the perceptible evidence and the intended meaning is the thoughts of the communicator it provides evidence for.”

Both S-Mode and I-Mode translation are shaped by affective and conceptual priming, but in qualitatively different ways and at different levels of awareness. S-Mode translation is primarily driven by direct phenomenal experience: the translator (or reader) responds to the stimulus—such as word choice, rhythm, or imagery—through embodied and affective reactions, often without conscious inference. Such responses are frequently guided by form-based attention and may involve structural or semantic priming, operating largely below the threshold of reflective awareness. While S-Mode relies on access to structure, pattern, or semantic association, I-Mode engages phenomenal awareness, where the translator attends to the speaker’s intention, tone, and experiential meaning. As we will discuss below, these distinctions become elements of a hierarchically organized Behavioral Translation Style Space.

3. Enactivist View

Understanding (higher-order) metacognition through enactive, non-representational lenses offers a radically different perspective from classical and modern cognitive models. In an enactive framework, metacognition is not about internal representations or mental models being monitored by some higher-order cognitive system. Instead, it emerges from the dynamic coupling between the translator and their translation environment. The translator’s metacognitive awareness arises through their embodied engagement with the text and an imagined target audience, not through accessing stored representations of their knowledge. It amounts to self-regulation through coupling with the environment and the reorganization of patterns of engagement.

Consider how a translator might experience uncertainty while working with an ambiguous passage. Rather than consulting internal representations of their linguistic knowledge, their uncertainty manifests as a felt sense of resistance or instability in the translation process itself. The text does not “flow” in the usual way; its typical patterns of meaning-making encounter obstacles. This uncertainty is not represented as propositional knowledge (“I am uncertain about X”) but emerges as a qualitative shift in their embodied engagement with the translation task.

The non-representational aspect becomes clear when we consider how translators often report that solutions “feel right” or that certain renderings create a sense of coherence that others lack [30]. These metacognitive judgments are not based on comparing current states to stored templates, but arise from the translator’s skillful attunement to the affordances present in the translation situation. Their expertise manifests as a refined sensitivity to what the text-in-translation is calling for, rather than as accessing comprehensive mental representations of translation rules or strategies.

In this enactive view, the translator’s metacognitive awareness is fundamentally relational and temporal. It emerges through the ongoing interaction between their embodied skills, the source text’s constraints and possibilities, the target language’s resources, and the broader communicative context. The translator does not step outside this process to observe it representationally; rather, their metacognitive sensitivity is woven into their skilled engagement with the translation task itself.

This perspective suggests that developing translational expertise involves cultivating increasingly sophisticated forms of embodied attunement rather than building more comprehensive representational mental models. The skilled translator becomes more sensitive to subtle variations in meaning potential, more responsive to the emerging demands of specific translation situations, and more capable of modulating their approach based on the qualitative feedback arising from their engagement with the text.

S-Mode and I-Mode Revised

Under an enactive interpretation, S-Mode translations, just like other perceptual modalities, are not brain-bound representations but skills of sensorimotor interaction [31]. Knowing how sensory inputs change as a result of previous action is more important than static internal images or symbolic content. Sensorimotor contingencies or patterns of readiness-to-act encode the relationship between action and sensory input that support embodied interaction with the world. In a representational view, the translator retrieves sensorimotor or linguistic representations to map the source to the target. The enactivist position would maintain that a translator is engaged in a real-time coordination process, where meaning is enacted through habitual attunement to linguistic and contextual cues.

Under an enactive interpretation, I-Mode translation, like other forms of sense-making, is not the retrieval of pre-formed conceptual structures but the skillful coordination of interpretive action in context. Understanding arises from the translator’s ongoing engagement with linguistic, situational, and cultural affordances, rather than from manipulating static semantic representations. Patterns of interpretive readiness—built from past encounters with texts, genres, and discourse practices—shape how attention is directed and how emerging meanings are integrated. In a representational account, the translator would access stored conceptual or propositional content to decode the source and recode it into the target. The enactivist position instead views the process as real-time co-constitution of meaning, in which the translator navigates and refines interpretive possibilities through continuous attunement to the evolving dynamics of the communicative situation.

4. Predictive Processing

Contemporary accounts of cognition grounded in Predictive Processing (PP) challenge classical representational views by reconceptualizing the mind not as a passive mirror of the external world, but as an active prediction engine. Scholars such as Clark [32], Seth [33], or Hohwy [18,34] argue that cognition arises from the continuous minimization of prediction error through reciprocal top–down and bottom–up interactions within the brain’s hierarchical generative models. Rather than constructing internal representations that reflect the world, the brain actively generates predictions about the likely causes of sensory input and updates these in light of incoming data. As Seth [33] explains, perceptual experience is not to “read off” sensory data, but emerges from the brain’s best guess—that is, approximate posterior beliefs resulting from a probabilistic integration of prior expectations and sensory evidence (likelihoods). This process, fundamentally Bayesian in structure, is implemented not via exact computation but through prediction error minimization, which enables the system to dynamically adjust its expectations in response to environmental feedback.

Importantly, these predictions are structured across nested hierarchical levels, such that high-level generative models constrain lower-level perceptual processes. Prediction errors flow upward in the system, allowing the brain to continuously refine its models. Moreover, this process is not limited to perception: it extends to action as well. Under the Active Inference (AIF) framework [22,35], organisms can also reduce prediction error by acting upon the world to make sensory input conform to expectations—closing the loop between cognition, perception, and embodied behavior.

Yet, as Clark [32] emphasizes, the predictive brain does not function in isolation. Rather, it is situated within dense bodily, social, and technological contexts that scaffold and shape its predictive capacities. The mind, on this view, is not confined to the brain but can extends into the environment when tools, linguistic structures, cultural artifacts, and interpersonal interaction become integrated parts of cognitive systems. These external resources contribute to reducing cognitive load and minimizing uncertainty, effectively becoming integrated components of the predictive machinery. As such, PP bridges the gap between brain-bound computation and extended, embodied cognition, portraying the mind as a dynamically enacted system that spans neural, bodily, and environmental domains.

4.1. Active Inference (AIF)

According to Deane (p. 3, [36]), “The notion of a hierarchical generative model lies at the centre of the Active Inference Framework.” AIF builds upon the principles of PP by integrating perception and action into a unified account of cognition, grounded in the Free Energy Principle (FEP) [20,22,35]. PP posits that the brain is organized hierarchically, with higher levels generating predictions about lower-level sensory input, thereby minimizing prediction errors through a continuous process of updating internal models. AIF extends this framework by emphasizing that agents not only passively receive information but also actively shape their sensory input through goal-directed interaction with the environment.

Under the FEP, biological systems sustain their viability by minimizing variational Free Energy—a formal measure of surprise or uncertainty relative to the system’s generative model of the world. AIF formalizes this through Bayesian inference, wherein agents continuously revise their beliefs based on prior expectations and new data. Importantly, agents also engage in epistemic actions—actions that reduce uncertainty by selectively sampling the environment. In the context of translation, this means the translator is not merely reacting to linguistic stimuli but is actively engaged in a dynamic cycle of hypothesis generation, testing, and refinement in response to unfolding communicative, cultural, and contextual cues.

From a representational perspective, the hierarchical generative model can be understood as an internal structure—a probabilistic map encoding beliefs about the external world and its communicative dynamics. In this view, translation involves updating and manipulating these internal models to best match the source input and generate an appropriate target output.

By contrast, an enactivist interpretation reframes the inferential machinery not as internal modeling, but as embodied engagement. Here, the generative model does not refer to stored content but to the agent’s skillful regulation of sensorimotor and affective interaction with the environment [37]. Cognition, including translation, is thus not a matter of inferring pre-given meanings from abstract representations but a process of enacting meaning in situ, through adaptive and context-sensitive coupling with linguistic and cultural affordances. Enactive Inference in translation is embodied, affectively modulated, and context-sensitive sense-making, where the translator actively brings forth meaning in interaction with the text and context, rather than passively computing correspondences.

4.2. Surprise and Prediction

AIF stipulates that expected input to the brain is unsurprising and may not be fully processed [38]. Hence, only surprising inputs are informative. In the translation situation, surprising inputs are likely to trigger transitions from fluent automatized processing to cognitive interventions and may result in conscious awareness. Similar assumptions have also been put forward in bilingualism studies [39] and interpreting (e.g., [40]), suggesting that language speakers anticipate upcoming linguistic input based on contextual cues, prior knowledge, and probabilistic expectations. This anticipatory processing allows speakers to pre-activate likely words and structures, facilitating faster and more efficient language comprehension and production. In translation and interpreting, this involves the translator predicting the speaker’s message, which allows him to manage cognitive load and maintain fluency.

It is controversial, however, how much planning ahead is required, what the size of the process-based translation units is, and whether and how these planning units constitute full linguistic entities [41]. Some schools (e.g., Fillmore [42] , and construction grammar) suggest that readers need to activate entire frames when engaging in language production. It is further unclear to what extent these assumptions are compatible with S-Mode translations, which stipulate that translators can work on a much smaller window. Up to date, there exists no integrated framework for testing and evaluating such hypotheses and their interactions in a formalized way.

4.3. Free Energy

Figure 1 provides a sketch of the translation process demonstrating the basic ideas of the Free Energy Principle (FEP) and Active Inference (AIF). It illustrates the dynamic interaction between the translator and their working environment, conceptualized as a continuous cycle of prediction, perception, and adjustment. The Free Energy equation in Figure 1 consists of two parts—Divergence and Evidence—which relate to different kinds of activities. Based on these terms, Free Energy F(Q,o) will decrease in the following manner.

Upon entering a translation task, a translator holds a set of prior predictions, beliefs, or expectations, denoted as P(b). These beliefs may include representational and/or action-based assumptions about the source text, the communicative intent, available translation resources, stylistic conventions, or contextual constraints imposed by the client, domain, or purpose of the translation, etc. These priors form the translator’s initial generative model—a probabilistic understanding of how the task should unfold.

As the translation progresses, the translator gathers sensory data through contextual observations, “$o$”, by consulting reference materials, interpreting pragmatic cues, or receiving feedback. Since computing the true posterior P(b|o) is intractable, the translator uses variational inference to compute an approximate posterior Q(b).The divergence between the approximate posterior Q(b) and the true posterior P(b|o) can be quantified by the KL divergence, $D_{KL}[Q\left(b\right) || P(b|o)]$. The translator's goal is to minimize this divergence, effectively minimizing prediction error between expected and actual sensory input. The approximate posterior would approximate the true posterior $P\left(b|o\right)$ if the complete information was available. The divergence measures the information-theoretical cost (in bits) of how much information is lost by using Q(b) instead of the ideal $P\left(b|o\right)$. Minimizing this divergence is equivalent to reducing the system’s epistemic uncertainty or Bayesian surprise.

Note that, when the variational approximation perfectly matches the true posterior, that is, Q(b) = P(b|o), the divergence is zero, and the variational Free Energy equals the negative log model evidence, as shown in Equation (1):

$$if: D_{KL}\left[\left(Q\left(b\right)\parallel P\left(b|o\right)\right) \right]= 0, \mathrm{then} F\left(Q, o\right)=-\mathit{ln}P\left(o\right)$$

(1)

While an ideal match between approximate posterior beliefs Q(b) and true posterior $P\left(b|o\right)$ is theoretically possible, such perfect congruence is rare in practice. In most real-world scenarios, the translator’s initial assumptions are partially or entirely challenged by new observations encountered during the task—such as unexpected textual features, ambiguous constructions, contextual inconsistencies, etc.

Rather than decreasing the Divergence, a translator can also aim at increasing the Evidence. Instead of passively updating beliefs to accommodate unexpected input, the translator may engage in active interventions, denoted by a, that deliberately reshape the translation environment. These interventions may involve insertions, deletions, rephrasing, or structural adjustments, intended to optimize the communicative fit of the translation with the translator’s expectations, task goals, or contextual constraints.

The consequences of actions are captured in the form of future observations, o(a,e), which enclose both the earlier state of the environment (e) and the sequence of translator-initiated actions (a). The observations are not arbitrary: they are shaped by the translator’s own attempts to bring about a more predictable and coherent state of interaction between the evolving translation and the translator’s internal model.

Crucially, the translator selects actions not at random, but in accordance with the principle of expected model evidence. That is, translators are more likely to choose actions that increase the anticipated congruence between forthcoming linguistic input and their current belief state Q(b). From this perspective, action becomes a means of epistemic self-evidence: the agent (translator) acts not merely to express, but to actively generate a world (text) that confirms and stabilizes their internal generative model.

In this enactive view, agents—including translators—do not passively receive information but shape their sensory and cognitive landscape in ways that minimize surprise and uncertainty [22,34,37]. By intervening in the environment, they help construct predictable futures, futures that are intelligible and actionable under their current model. Thus, action serves both a pragmatic function (reshaping communicative outcomes) and an epistemic function (reducing inferential load), grounding behavior in both goal-directed regulation and model-confirming engagement.

Thus, FEP and AIF propose that all cognitive agents—including translators—strive to minimize Free Energy over time in order to maintain adaptive coherence with their environment. The Free Energy equation in Figure 1, here reproduced as Equation (2), consists of two parts which account for two complementary strategies.

$$F\left(Q, o\right)=\stackrel{\mathrm{Divergence}}{\overbrace{D_{KL}\left[Q\left(b\right)\parallel P\left(b\left|o\right(a,e\right)\right]}}-\stackrel{\mathrm{Evidence}}{\overbrace{lnP\left(\mathrm{o}\left(a,e\right)\right)}}$$

(2)

The Divergence term represents Bayesian surprise—the divergence between what the translator initially believed and what they now infer based on evidence. The second term, log model Evidence, corresponds to Shannon surprise—a measure of how unlikely or unexpected the observations were under the model. From a cognitive translation perspective, this suggests two possible modeling strategies:

• Inference-driven translation, where the translator tends to updates expectations based on new linguistic input. This view aligns with representational approaches, in which translation is seen as a decoding of meaning based on evolving mental representations.
• Action-oriented translation, in which the translator actively shapes the communicative situation—through strategic choices, recontextualization, or adaptive stylistic shifts—to render the translation environment more predictable and congruent with prior knowledge. This process-oriented view resonates with enactivist theories, emphasizing the translator’s role as an agent embedded in, and acting upon, a dynamic communicative environment.

As Parr et al. [21] summarize, this unifying framework allows us to understand perceptual inference, action planning, and learning as different manifestations of the same underlying principle—the minimization of variational Free Energy. Within translation studies, such a framework offers a powerful lens for integrating representational and enactive accounts of translation cognition under a shared formal model.

4.4. Complexity and Accuracy

Note that $P(o$)—the observation evidence—is often computationally intractable, because it requires summing over all possible hidden states b, which becomes infeasible in high-dimensional spaces:

$$P\left(o\right) ={{\displaystyle \sum }}_{b}P\left(o|b\right) P\left(b\right)$$

(3)

However, it is possible to refactor Equation (2) as Equation (4):

$$F\left(Q,o\right)=\stackrel{\mathrm{Complexity}}{\overbrace{D_{KL}\left(Q\left(b\right)\parallel P\left(b\right)\right)}}-\stackrel{\mathrm{Accuracy}}{\overbrace{{\mathbb{E}}_{Q\left(b\right)}\left[lnP\left(o\left(\mathit{\alpha },e\right)│b\right)\right]}}$$

(4)

In Equation (4), the Complexity term quantifies how expectations coincide with evidence, and the Accuracy term tells us how actions are in tune with expectations. In Equation (4), the Complexity term, on the one hand, avoids computing P(o) (or P(o($a,e$))) by using an optimization that does not require the full normalizing constant. The Accuracy term, on the other hand, is conditioned on a particular, limited set of internal belief states, $b$, which are known as part of the generative model Q(b). These computations are thus computationally tractable.

Both Equations (2) and (4) are mathematically equivalent (pp. 28,29, [21]). However, Equation (4) foregrounds that agents infer explanations that have minimal complexity. While Equation (2) is used to show why FEP works, in practice, the “complexity–accuracy trade-off” as in Equation (4) may be more practical. The complexity cost is incurred when maintaining a sophisticated internal model, while the accuracy benefit comes with better models, which enable more precise predictions and actions. Equation (4) provides an “interpretation of Free Energy minimization as finding the best explanation for sensory data, which must be the simplest (minimally complex) explanation that is able to accurately account for the data.” (ibid.) It shows that Free Energy can be reduced in two different ways, by maximizing Accuracy and/or by minimizing Complexity. Thus, Hohwy (p. 87, [34]) stipulates that “one could maximise accuracy under complexity constraints, by updating beliefs; however, one could also maximise accuracy, without adding complexity, by acting upon the world to change the sensations being explained” According to Hohwy, “[m]inimising complexity can be read as keeping posterior beliefs close to prior beliefs,” however, “in the absence of a complexity constraint there would be no characteristic states towards which an agent would self-evidence.”

5. Hierarchical Organization of Translation States

Crucial for PP is the system’s hierarchical organization and the ability to construct higher-order generative models of its own lower-level mental states. These higher-level processes generate predictions about their embedded states, such as, for instance, the sensory inputs they receive. When these predictions are accurate, we experience a sense of familiarity or certainty. However, when a prediction is incorrect (i.e., a prediction error occurs), it signals that something unexpected has happened, and, on a higher level, this can lead to conscious experience. The mechanism is regulated by a construct referred to as precision weighting, indicating how confident the brain is in a prediction. Precision weighting refers to the brain’s ability to assess and weigh the reliability of different sources of information, including sensory inputs, internal predictions, and prior beliefs. It is a mechanism for determining how much weight to give to different pieces of evidence when making a decision or forming a belief [36]. Conscious states are then those that receive high-precision estimates, making them dominant in shaping behavior. (e.g., [18,22,43]).

5.1. Self-Evidencing

According to PP, the brain constructs generative models of the environment that are continually updated in response to sensory input. These models are self-evidencing in the sense that they actively shape perception and action to sample data that confirm their own predictions [18]. The concept of self-evidencing stipulates that agents minimize Free Energy by optimizing their internal models and regulating their sensory engagements. For this, they effectively infer their own existence, maintaining structural and functional coherence through anticipatory interaction with the world. Hohwy [34]maintains that “If the model we are looking at manages to persist, it must be doing something in addition to passive updating, which makes the model accurate, and this can only be to engage in active, selective sampling of its expected observations.”

The translation process, when viewed through the lens of PP and AIF, can be seen as a self-evidencing activity: the translator enactively regulates their interaction with linguistic, cultural, and pragmatic constraints by forming and testing predictions about meaning and communicative intention. When determining interpretive resemblance (I-Mode translation), translators draw on affective and emotional cues [11,44,45,46], supporting the notion that translation is an embodied, affectively anchored practice of active sense-making. The translator’s felt alignment with textual and contextual cues serves as a form of prediction validation.

Higher-order internal states regulate/interact with lower-order ones, not by mirroring them, but by providing functionally useful predictions. These higher-order states do not represent the lower-level states in a pictorial or symbolic way but model them as being a certain way, typically via structured, functional mappings.

5.2. The Behavioral Translation Style Space (BTSS)

The Behavioral Translation Style Space (BTSS, [47]) maps translators’ behavior—including keystrokes, gaze patterns, pauses, revisions, and cognitive–affective states—onto a multi-dimensional embedded state space. This space captures the microdynamics of the translation process over time. The BTSS consists of several processing strata that form the nested hierarchical architecture of the translating mind. Some of the layers are described in

Table 1. Layers of the Behavioral Translation Style Space (BTSS).

	Layer	Focus
1	AU (Activity Units)	Minimal coordination of pausing, gazing, and typing.
2	PU (Production Units)	Chunks of planned typing activity; PUs consist of sequences of AUs, separated by production breaks.
3	OHRF (Orientation, Hesitation, Revision, Flow)	Affective aspects of epistemic/pragmatic affordances, OHRF states consist of sequences of PUs, and/or PU breaks.
4	TP (Translation Policies)	Perception Action loops are sequences of O(HRF)+ states.

, and Figure 2 and Figure 3 plot a graphical description of those layers; a full description of the BTSS is provided in [48,49].

The four temporally embedded layers in Table 1 are also marked in Figure 2. Each of the four layers has characteristic features:

(1)

AUs (Activity Units) fragment the behavioral data into six categories that specify whether the translator is involved in ST or TT reading, whether s/he is typing, or simultaneously reading and writing. AUs unfold on the smallest temporal layer. They can be as short as a single fixation or a few keystrokes, up to the length of a linear reading pattern.

(2)

PUs (Production Units) consist of sequences of Keystroke Bursts and are interrupted by longer “willful” PU breaks (>3 *median between-word keystrokes, see [49]). While Keystroke Bursts are learned sensorimotor contingencies based on habits and embodied skills, PU breaks reflect moments of heightened cognitive–affective regulation, where the translator engages in deliberate sense-making, planning, or decision-making. They mark transitions from fluent embodied action to reflective control, revealing the temporal segmentation of the translation process into phases of automaticity and conscious modulation.

(3)

OHRF states are translation states that indicate the following:

• Orientation (O): prolonged ST reading, an epistemic affordance, enabling the translator to actively sample information from the ST, update internal belief states, or reduce uncertainty about meaning and intention.
• Hesitation (H): unexpected challenges, leading to longer typing pauses, rereading, or revisions, indicating cognitive uncertainty or conflict.
• Revision (R): prolonged TT reading, reflecting the translator’s attempt to understand/revise the TT. In our previous research [48,49], we only used three states: H, O, and F. Here, we introduce the new revision (R) state.
• Flow (F): fluent, uninterrupted production, a pragmatic affordance, characterized by minimal reading and short breaks, signifying full cognitive immersion.

(4)

TP (Translation Policies) are Perception–Action loops that consist of sequences of OHRF states. A TP starts out within an epistemic affordance (O), followed by one or more H, R, or F states. A TP typically ends in a pragmatic affordance (F), where well-aligned predictions and motor plans enable fluent production with minimal monitoring. A TP may also contain states of hesitation (H) that signal moments when predictive models are challenged and require revision, often triggering exploratory actions such as rereading or tentative reformulations, or revision (R), which involves deliberate restructuring of the target text, integrating new insights into the evolving translation. Through these cycles, TPs adaptively shape the translator’s interaction with the task environment, progressively minimising prediction errors and guiding the translation toward communicative adequacy.

5.3. An Example

The graphs in Figure 2 and Figure 3 plot the behavioral data for the translation of an English sentence into Arabic. It is a small fragment of the CRITT TPR-DB, a large database of behavioral translation process data that consists of several thousand translation sessions . While Figure 2 shows the raw logging data (keystrokes and gaze data with fixation mapping onto ST and TT words), Figure 3 illustrates the segmentation of this data into a nested hierarchical structure, the Behavioral Translation Style Space (BTSS).

The Figures plot a stretch of 110 s (timestamp 175–285) fragmented into five OHRF states. These five states are clustered into two TPs (separated by a dotted line in

Table 2. Sequence of OHRF states in and their contribution to TPs as plotted in Figure 2 and Figure 3.

	OHRF	Time (s)	Description
1	O	175–193	Read the entire ST sentence
2	F	193–230	Type the translation of an initial fragment of the sentence
3	O	230–248	Reread the last part of the sentence
4	R	248–252	Read/confirm the previously produced TT segment
5	F	252–285	Type the translation of a second fragment of the sentence

), each spanning approximately 55 seconds, specifically (175–230) and (230–285). The duration of the initial orientation (epistemic affordance) is approximately 18 seconds in both TPs, (175–193) and (230–248), respectively.This is followed in the second TP by a short revision, of approximately 4 s, in which the translator apparently reconfirms her previous writing before continuing with the next segment..

The relation between epistemic affordances (orientation, revision, i.e., reading ST and/or TT) and pragmatic affordances (typing/text modifications) amounts to the Accuracy–Complexity trade-off. Epistemic affordances may temporarily increase complexity, since gathering new information may strain memory resources or increase immediate uncertainty. However, it may also lead to more sophisticated plans and models, with the benefit of reducing future prediction errors. Extended epistemic actions (which include, in this case, also the Revision state, (R), in the second TP, as no modifications are produced) can be considered investments in model complexity for better contextual understanding.

Pragmatic affordances, in contrast, tend to reduce model complexity, since they reduce the unpredictability of the environment, thereby making the models less complex. The Accuracy–Complexity trade-off creates an explore–exploit dynamic where translators alternate between epistemic actions (explore) by actively gathering sensory or linguistic input to increase model accuracy and reduce prediction error. Pragmatic actions (exploit), on the other hand, directly move the translator toward the communicative goal with minimal new information gathering. In this trade-off, a translator must continually balance epistemic and pragmatic affordances to manage time, cognitive load, and quality.

Representational interpretations of this process have been suggested in which the internal state space is considered a representation of some sort of states of affairs of an external reality. Alternatively, in the enactive view, translators regulate their interaction in a translation task as dynamically enacted, embodied trajectories through linguistic, affective, and cultural affordances. This latter view is, for instance, endorsed by Ramstead et al. (p. 227, [37]) who introduce the notion of Enactive Inference, maintaining that, under FEP, “generative models are [enactive] control systems, and they are not structural representations.”

6. Conclusions

This paper lays out approaches of modern representationalism and enactivist views of translation and suggests Predictive Processing under the Free Energy Principle as a framework to accommodate both views. While modern representationalism holds that perception arises from an inferential process over internal symbols revealing structural regularities in the world, enactivism rejects the notion of such internal stand-ins. Instead, it views perception as emerging from the dynamic coupling of an embodied agent and its environment, where meaning is enacted through sensorimotor skills, affective engagement, and pragmatic action. From this perspective, structural regularities are not abstracted internal constructs but lived, pragmatic patterns disclosed through embodied interaction.

Predictive Processing (PP) remains agnostic to these competing interpretations of internal states, functioning as a generic framework for hierarchical inference and belief updating. Within a representational view, translation involves multiple strata of internal models: from sensorimotor representations of orthography and phonology, through lexical-semantic and syntactic mappings, to pragmatic and meta-level beliefs about genre and communicative intent, etc. These hierarchies integrate top–down predictions with bottom–up error signals, making translation a multi-level inferential process of continual model refinement.

In contrast, the enactive view would conceive translation not as manipulation of symbolic content, but as a situated activity distributed across diverse strata of embodied and affective engagement, co-regulating perception and action in real time. Within PP, these processes can be modeled as hierarchically organized generative models, appealing to sensorimotor skills and affective modulation rather than amodal symbol manipulation.

Active Inference thereby plays a central role in translation. Translators do not merely update beliefs to fit the text (perceptual inference) but actively intervene—through reformulation, restructuring, or selective emphasis—to shape the unfolding translation environment in ways that align with prior expectations or stylistic aims. Each intervention changes the structure of future or expected observations, linking the evolving state of the translation environment with the translator’s embodied actions. The goal, then, is to maximize model evidence—producing a target text that is not only faithful to the source but also coherent, contextually attuned, and self-consistent under the translator’s preferred generative model.

By framing translation as a dynamic cycle of prediction, evaluation, and adaptive action, this paper demonstrates how Predictive Processing and Active Inference (PP/AIF) can accommodate representational and enactivist accounts within the same mathematical framework. However, this formal unification does not amount to a full philosophical reconciliation between the two paradigms. Moreover, it remains unclear at this point how representational and enactivist interpretations of empirical translation process data lead to different conclusions on the different levels of the Behavioral Translation Style Space. Future investigations will show whether and how symbol-oriented and skill-based approaches to meaning-making in translation will consolidate.