Decision-makers face a fundamental challenge: should they exploit familiar, rewarding options or explore uncertain alternatives that might yield better long-term outcomes? (Cohen et al., 2007; Cogliati Dezza et al., 2017). This exploration - exploitation dilemma arises across species and contexts, from animal foraging to human strategic planning (Monk et al., 2018; Berger-Tal et al., 2014; Rich and Gureckis, 2018), yet questions persist regarding the psychological and neurobiological mechanisms underlying these competing drives (Berger-Tal et al., 2014; Mehlhorn et al., 2015).

Adaptive gain theory (AGT) provides a neurobiological framework for understanding the explore–exploit balance by linking behavioral modes to activity patterns in the locus coeruleus–norepinephrine (LC- NE) system (Aston-Jones and Cohen, 2005). According to AGT, the LC-NE system operates in two functionally distinct modes. In the phasic mode, the LC exhibits transient bursts of activity in response to task-relevant events, supporting focused attention and goal-directed behavior, conditions optimal for exploiting known reward sources. In the tonic mode, elevated baseline LC firing rates, reflecting a state of exploratory control - heightened arousal and promotion of exploration of alternative options, while lower tonic rates are associated with exploitative control, which is focused on choosing the known best option. Critically, these modes are not merely different levels of activation but correspond to qualitatively different patterns of neural activity that bias behavior toward exploitation or exploration, respectively.

Pupil diameter provides a noninvasive window into these LC-NE dynamics. Changes in pupil size closely track LC activity: baseline (average) pupil diameter reflects tonic LC firing, while transient pupil dilations index phasic LC responses (Joshi et al., 2016; Murphy et al., 2014). This correspondence allows pupillometry to test AGT’s core prediction that exploration and exploitation categorically distinct neuromodulatory states. Specifically, AGT predicts that exploration should be accompanied by larger tonic pupil size, reflecting sustained arousal in the tonic LC mode, whereas exploitation should be characterized by smaller baseline pupils with transient phasic dilations time-locked to decision-relevant events (Jepma and Nieuwenhuis, 2011). Consistent with this view, recent work of Fan et al. (2023), has demonstrated that baseline pupil size tracks total uncertainty in the environment and is associated with increased random exploration, suggesting that tonic arousal reflects a global state of uncertainty rather than specific decision demands (Fan et al., 2023).

While AGT framework links neuromodulatory activity to behavioral flexibility, previous findings revealed that exploration is not a unitary phenomenon. Wilson et al. (2014) distinguished directed exploration- deliberate information-seeking guided by uncertainty, from random exploration, which reflects stochastic choice variability unrelated to specific information goals. This functional distinction challenges AGT’s arousal-based account: directed exploration requires effortful goal maintenance and value integration, whereas random exploration may arise from diffuse arousal or decision noise. Yet these two forms of exploration exhibit fundamentally different behavioral pattern: directed exploration is modulated by information structure and planning opportunities, whereas random exploration scales with overall uncertainty regardless of whether the environment affords effective information use (Gershman, 2019).

This dissociation indicates that exploration cannot be fully captured by arousal fluctuations alone but depends on strategic control processes selectively recruited according to decision context. The Expected Value of Control (EVC) framework provides a complementary account, proposing that cognitive control is allocated when anticipated benefits, such as reducing uncertainty or optimizing future outcomes- outweigh intrinsic effort costs (Shenhav et al., 2013). Critically, EVC predicts that control deployment should scale with the prospective utility of investing effort in each decision rather than with exploration per se. This perspective suggests that exploration and exploitation behaviors may recruit control levels depending on strategic context, specifically, whether the decision environment affords sufficient opportunity to capitalize on acquired information (Sadeghiyeh et al., 2020; Frömer et al., 2021). Empirical evidence supports this prediction: Dubois and Hauser (2022) demonstrated that the planning horizon, the number of choices remaining in a task, dictates exploration strategy. They found that participants strategically increased random exploration only when the horizon was long (e.g., six choices), affording an opportunity to capitalize on the information, but relied less on this strategy when the horizon was short. This cost–benefit logic has been formalized computationally by Musslick et al. (2015), who demonstrated that the cognitive mechanisms governing goal-directed response selection determines when to deploy effortful control versus rely on stochastic processes.

Integrating AGT and EVC thus suggests that arousal and control jointly regulate adaptive decision-making: tonic arousal sets background readiness for behavioral flexibility, while strategic control is selectively recruited when decision context renders effortful processing worthwhile. Recent neurophysiological evidence supports this integrated view. Kozunova et al. (2022) using a continuous probabilistic learning task, provided direct support for this distinction. They operationalized exploitation as choosing the known high-payoff (HP) option after learning and directed exploration as choosing the known low-payoff (LP) option. They found that directed exploration elicited significantly larger phasic pupil dilations and slower response times compared to exploitation. This effect was absent during “random exploration” (choices made before learning), suggesting the pupil response tracked more than just uncertainty. Instead, the authors compellingly attribute the dilation to the cognitive conflict arising from intentionally overriding a strong, learned bias to exploit. However, both exploration and exploitation decisions happened within a continuous learning task, where exploitation reflects low expected value of control, while directed exploration reflects higher value of control. Shorter choice horizon, where the current choice cannot be used to inform future decisions, may reflect different values of control to exploration and exploitation.

This finding isolates a key component of strategic processing, linking pupil dilation to conflict-driven control. A key open question, however, is whether this pupillary response is specific to this form of internal conflict, or if it reflects a broader component of exploration, such as decision uncertainty. This distinction is central to integrating the EVC and AGT frameworks, particularly as other findings suggest pupil-linked arousal can be dissociated from effective control deployment. Frömer et al. (2021), for example, demonstrated that while neural markers of control (ERPs) scaled with reward and efficacy, pupillary responses showed a contrasting pattern: they were largest under low efficacy conditions. In that context, uncertainty was high, but strategic control was inefficacious. This suggests that pupil responses can track uncertainty-driven arousal even when strategic control is not effectively deployed.

The present study examines how pupil-indexed arousal relates to exploration and exploitation under varying conditions of uncertainty, information structure, and planning opportunity. Using the Horizon Task, a multi-trial decision task in which participants choose between two options with different reward distributions (Figure 1). In this task, participants complete four forced-choice trials that provide initial information about each option’s payoff, followed by either one or six free-choice trials during which they can explore or exploit. We manipulate value differences, information asymmetries, and horizon length to differentially promote directed exploration, random exploration, and exploitation. AGT predicts that explorative control elicits elevated tonic arousal while exploitative control engages phasic responses. EVC predicts arousal scales with the strategic value of control deployment rather than choice type. In the horizon task, directed exploration is associated with the choice of low value option under long horizon conditions, while random exploration is associated with choosing the low value option under short horizon conditions. Exploitation is related to choosing the high-value option, but in this task the importance of exploitation changes. Specifically, when value differences are small (high uncertainty), exploitative choices require greater discriminative effort and carry higher impact on outcomes, increasing their expected value of control. Similarly, when horizons are short, exploitative decisions become more critical as there are no future opportunities to correct errors, further elevating their value of control. In contrast, exploration choices under long horizons may carry lower control demands because errors can be compensated for in subsequent trials, while exploratory choices under short horizons (random exploration) may reflect low-cost stochastic sampling rather than effortful strategic processing. Thus, integrating EVC consideration provides novel predictions for pupil diameter in exploration and exploitation decisions, predicting that pupil size should be larger (reflecting higher arousal and control investment) when the expected value of control is high, specifically, during exploitative choices under small value gaps or short horizons, and smaller when the expected value of control is low, as in exploratory choices where strategic demands are minimal. By measuring pupil dynamics across these conditions, this study tests whether arousal tracks exploratory versus exploitative choices or the contextual factors determining when effortful processing is warranted.

All models included random intercepts for participants. More complex random-effects structures including random slopes for the experimental predictors were evaluated but did not converge reliably. Therefore, the reported results are based on the most parsimonious models that yielded stable estimates.

The present investigation examined the relationship between pupillary responses and different types of control demands, associated with exploration and exploitation behavior. We used a decision-making task where the value of control was manipulated in different ways, including the level of uncertainty about the best option, the future consequences of the decision and the importance of current outcome. These allowed us to track subtle differences within exploration and exploitation decisions, and to examine how they bear common theories of task engagement. Behavioral analyses confirmed the importance of these factors in influencing participants’ choices. Participants adjusted their choices based on value uncertainty, consistently chose options with less prior information in asymmetrical-information conditions, and exhibited greater information-seeking behavior when more free choices remained (six vs. one). These patterns replicated Wilson et al. (2014) findings and confirmed the effectiveness of the experimental manipulations in eliciting exploration behavior.

Pupillary responses, however, revealed a more selective pattern. In line with AGT predictions, pupil size was somewhat larger in exploratory choices, i.e., choices of lower value options, compared with exploitations. However, when examining patterns within each type of choice, we found that pupil size was affected by task parameters only during exploitative choices. Both choice horizon and value-gap significantly predicted average pupil size: pupils were larger when only one decision remained, reflecting how importance and value of control shape arousal, and when value differences in values were small, reflecting directed exploitation. This dissociation reveals that pupil-indexed arousal selectively tracks the strategic demand for control, beyond differentiating simply between exploitative and exploratory decisions.

These findings prompt a refinement of AGT’s predictions regarding exploration and exploitation. AGT proposes that exploration engages elevated tonic arousal while exploitation engages transient phasic bursts (Aston-Jones and Cohen, 2005; Jepma and Nieuwenhuis, 2011). Our findings partially support this framework: we observed modulation of average (tonic) pupil size in exploratory and exploitative choices, consistent with AGT’s emphasis on sustained arousal states. However, further task-related modulations were observed during exploitation rather than exploration, and specifically during high uncertainty, where exploitative decisions are still uncertain and provide information, similar to directed exploration. This suggests that exploration and exploitation control strategies should be defined according to task demands and not just based on choice value. Moreover, we did not observe discrete phasic peaks time-locked to decisions but instead sustained elevation throughout the decision period. This pattern is consistent with the view that pupil signals reflect context-dependent integration of multiple neuromodulatory and cognitive processes, rather than a one-to-one mapping onto phasic versus tonic LC activity (Joshi and Gold, 2020). Notably this stable temporal profile contrasts with the discrete phasic bursts observed in visual- attentional tasks (Gabay et al., 2011; Preuschoff et al., 2011) and may reflect prolonged evaluative processing under strategic constraints, as was tested in early study in working memory domain (Granholm et al., 1996).

The absence of pupil modulation during exploration warrants consideration alongside prior evidence. Kozunova et al. (2022) found that directed exploration elicited larger pupil dilations compared to exploitation. Critically, their operationalization involved deliberately choosing a known low-payoff option after learning, thereby overriding an established exploitation bias- a process likely engaging conflict-driven control (Shenhav et al., 2013). In contrast, the Horizon Task manipulates environmental factors that promote exploration without imposing strong conflict or immediate costs. Participants can explore with relatively low cost, particularly when horizons are long and future opportunities remain available (Wilson et al., 2014; Zajkowski et al., 2017). This difference in the way exploration is elicited by the task may explain why we observed no pupil modulation during exploration despite systematic exploration strategies, as pupil-linked arousal in such contexts may primarily reflect uncertainty at the level of the environment rather than decision-specific control demands (Fan et al., 2023).

This interpretation aligns with the Expected Value of Control (EVC) framework (Shenhav et al., 2013), which proposes that control is allocated when anticipated benefits outweigh costs. Exploitation under short horizons represents high-importance decisions where errors cannot be corrected, and exploitation with small value differences requires enhanced processing for accurate discrimination. Both conditions increase the expected value of control, warranting greater resource investment. Importantly, pupillary responses have been shown to index the amount of cognitive effort actually invested, rather than objective task difficulty per se, highlighting the distinction between task demands and effortful control allocation (van der Wel and van Steenbergen, 2018). This finding is also consistent with resource-rational principles (Lieder and Griffiths, 2020), which argue that people allocate cognitive resources only when the computational costs are justified. Our current adapted paradigm, characterized by relatively low-cost exploration; initial uncertainty is reduced by forced trials and the stakes of any single choice are low. This creates a condition where exploration is not computationally demanding. Thus, the absence of pupil modulation during exploration may not indicate a failure of strategic processing, but rather an efficient adaptation where costly cognitive control was unnecessary. This account is also in line with recent works that similarly shown that pupillary responses track the efficacy and reward value of control deployment (Frömer et al., 2021) and vary systematically with task demands (Nassar et al., 2012).

Our findings should be interpreted in light of several limitations. The Horizon Task structure provides initial outcome information via forced trials, reducing baseline uncertainty and potentially lowering the strategic demands of directed exploration relative to paradigms requiring information-seeking under complete uncertainty. This limits generalizability to contexts involving substantial opportunity costs or cognitive conflict. Additionally, pupillometry provides an indirect measure of arousal, and individual differences in exploration strategies and cognitive control capacity were not examined. Future work combining pupil measurements with direct neural recordings (e.g., fMRI) across diverse decision context, particularly those where exploration and exploitation vary in strategic importance beyond immediate outcomes- will clarify how pupil-linked arousal indexes context-dependent control allocation. Moreover, from a modeling perspective, although we evaluated more complex random-effects structures, models including random slopes did not converge reliably, likely reflecting overparameterization relative to the experimental design and the limited number of levels per predictor (Bates et al., 2015). We therefore report parsimonious models with random intercepts for participants, highlighting the need for future studies with greater within-subject variability to more fully characterize individual differences in context-dependent control demands.

To conclude, our results provide novel insights into how different styles of control shape pupil-indexed arousal, integrating AGT and EVC theoretical frameworks. We demonstrate that arousal does not uniformly distinguish exploration from exploitation but instead tracks contextual factors determining when control is strategically warranted.