Introduction

Despite its unpleasantness, pain serves a useful function. This is most clearly the case for phasic pain, where newly incoming nociceptive signals warn of impending tissue damage, elicit rapid defensive responses, and drive learning that reduces future chances of encountering harm. In this context, function can be objectively measured through conditioned responses and avoidance behaviour, which quantify pain in terms of its motivational value (Fields, 2006; Seymour, 2019). However, the adaptive function of tonic pain has been much harder to quantify. In ecological and ethological contexts, tonic pain is generally considered to serve protective and recuperative functions. For instance, in the context of injury, it has been proposed that tonic pain can aid recovery by reducing motivational vigour and hence suppress unnecessary activity until healing occurs (Bolles and Fanselow, 1980; Wall, 1979; Walters and Williams, 2019; Seymour et al., 2023). The goal of this study was to test whether we can formally and quantifiably dissociate these two distinct motivational functions of pain.

One of the challenges in studying adaptive functions of pain is the difficulty of embedding experiments within ecologically meaningful contexts. To solve this, we designed an immersive foraging task using virtual reality (VR), in which humans search a forest to collect fruits from the low-lying bushes at varying heights. A foraging paradigm provides a robust, free-operant framework that captures the core components of adaptive behaviour: it is goal-directed, involves complex movement, and requires the learning of an optimal strategy to maximise rewards. This allows us to computationally dissociate how different types of pain influence the control of action. Our first hypothesis was that phasic pain provides a distinct valuation signal that updates the value of specific actions within complex environments. In our task, this was implemented by associating specific fruit (distinguishable by colour) with a brief electrical stimulus to the grasping hand, emulating thorns. In our computational model, this was defined as an aversive utility term incorporated into the state-action value evaluation process. We predicted that this computational mechanism would manifest behaviourally as a reduction in choice probability for pain-associated targets and an increase in ‘choice distance bias’ (the willingness to travel further for pain-free options). Neurally and physiologically, we predicted that these aversive values would be tracked by skin conductance responses (SCRs) and the amplitude of nociceptive event-related potentials (ERPs), specifically the N1-P2 complex (Favero et al., 2023).

Second, we hypothesised that tonic pain acts as a coefficient modulating the trade-off between opportunity cost and vigour cost, thereby serving a recuperative function. To test this in Experiment 2, we delivered continuous tonic pressure to the non-dominant arm via an inflated cuff to emulate a background state of injury. Within our free-operant framework, tonic pain was modelled as a weighting factor that shifts the optimal balance toward reduced energy expenditure. Because the stimulus was applied to the non-task limb, we specifically predicted a global reduction in motivational vigour—operationalised as decreased movement velocities and foraging rates—rather than a direct mechanical impairment. By applying this formal computational approach, we move beyond exploratory observations to provide a rigorous, mechanism-based explanation for how distinct pain states adaptively govern choice and action.

Results

Experiment 1: Phasic pain avoidance in a free-operant foraging task

Twenty-five subjects performed a free-operant instrumental foraging task to accrue reward points and avoid pain. The task was embedded into a fully immersive VR context, in which subjects physically moved in a flat open space. A virtual boundary of 4 m × 4 m was displayed when the subject approached the edge of the area. Within the VR context, this space resembled a flat forest (Fig. 1), with trees and vegetation locations randomly generated. The task was structured into one-minute blocks. At the beginning of each block, virtual fruits of varying heights and locations were generated within the virtual space. The spatial coordinates of subjects’ heads and hands were tracked, requiring them to physically move towards the fruit, pick it up by reaching with their dominant hand, clicking, and holding the button on the handset. Only the dominant hand was enabled in the virtual environment. Subsequently, subjects could drop the fruit into one of the baskets to earn points. Subjects were told that the total points accrued would be rewarded with an incentive of up to £10. All fruits used in both experiments were pineapples, as their spiky shape naturally matches the pricking pain induced by the Wasp pain stimulation electrodes.

Figure 1.

Pineapples were categorised into two visually distinguishable types: 50% were green and 50% were yellow. The green pineapple was aversive, and picking one up immediately elicited a brief painful cutaneous electric stimulation to the proximal medial side of the middle finger of the (grasping) dominant hand. In contrast, picking up a yellow pineapple was always pain-free. The intensity of the electric stimulation was held constant within each block. Experiment 1 consisted of 20 blocks, with the initial 10 blocks designated for practice purposes with a fixed order of stimulation intensities. The subsequent 10 blocks used randomised pain stimulation intensities across five different levels, determined by a pain calibration procedure. After completing the 10 training blocks, all subjects were aware that picking up green pineapples resulted in painful shocks, as confirmed through verbal questioning. However, subjects did not know the exact pain level of a new block until they tried to pick up one green pineapple. Only data from the latter 10 non-training blocks were analysed.

Avoidance increases with increasing phasic pain intensity

We found that the probability of choosing a fruit decreased as the pain associated with it increased. For this analysis, both aversive choice probabilities and subjective pain ratings were estimated at the block level. Figure 3A shows an approximately linear decrease in aversive choice probability as a function of visual analogue scale (VAS) rating of pain stimuli, whereby the aversive choice probability is defined as the number of painful fruits collected divided by the total number of fruits collected.

Additional cost of effort associated with movement

Next, since movement itself may carry a small cost related to time and effort (incurring opportunity costs), we investigated whether there is a trade-off relationship between moving distances and pain intensity (i.e., is it worth moving further to avoid a certain amount of pain). We measured the distance from the fruit to the subject at the moment the subject first fixated on the fruit, as estimated from eye-tracking data, after collecting the previous one. We found a significant correlation between the egocentric distance differences between painful and non-painful fruits, and the VAS ratings. As shown in Figure 3B, the vertical axis represents the ‘choice distance bias’, calculated as the difference between the average egocentric distance to non-painful fruits and the average egocentric distance to painful fruits within each block. The egocentric distance is the fruit distance relative to the participant. This metric was computed to test whether subjects would trade off physical effort for pain avoidance; specifically, a positive bias indicates that subjects were willing to bypass closer painful fruits to reach more distant pain-free ones. As hypothesised, we found that as the pain intensity (VAS) of the aversive fruits increased, this distance bias grew significantly, confirming that subjects exerted greater movement effort to avoid higher levels of pain.

Choice-based computational modelling in RL framework

As noted above, these initial regression results were analysed at the block level. Within each one-minute block, each pickup, which could result in an immediate shock, can be considered an individual trial. By analysing the data at the trial level, we can extract more detailed information and provide a parameterised behavioural description for each individual. This approach laid the groundwork for a robust, objective measure of phasic pain value within an ecologically valid context. In each trial, despite the presence of numerous lower-level controls required to complete the action, the primary decision involved choosing which fruit to pick up. We simplified the problem by assuming that there exists a set of decision points 𝒮 as states, where the subject took an action at a state s ∈ 𝒮—choosing to pick up a fruit—and deterministically arrived at another state s^′ ∈ 𝒮if the block had not concluded. Based on the initial regression results, we assumed that the reward was a linear combination of an internal reward r for each fruit, a negative utility function u representing the phasic pain value, and an effort cost function d. Therefore, the Bellman optimality equation for the action-value function was

, where C_p and C_m were the parameters to be fitted for each subject.

As fruits were randomly distributed in our task, previous foraging studies suggested that animals may adopt a suboptimal greedy policy, opting for the closest fruit rather than solving the complex travelling salesman problem for a marginally better optimal solution (Anderson, 1983; Jeon et al., 2023). To initially demonstrate the basic effects, we further eschewed learning and future planning by setting the discount factor γ = 0. Figure 2 shows the conceptual diagram of the computational framework setup. One of the key simplifications was assuming the existence of a set of decision points in this free-operant context. To fit the model, we again utilised the eye-tracking data, setting the decision point as the first time point when the subject saw a new fruit. Nevertheless, the task’s available actions were partially observable, and subjects could explore for more actions if they were not satisfied with the currently available ones. This exploration action was not modelled in our formulation. To proceed with model fitting, we retrospectively identified the fruits that were picked up and selected only those time points when the fruits chosen afterwards were first seen as the decision points. We also kept track of the previously seen fruits, storing them in a memory queue, so the decision involved choosing the best option from these items in memory. Figure 3(C, D) shows the parameters fitted to the subjects’ behavioural data. This shows the negative utility associated with pain and vertical and horizontal components of movement, and it illustrates how the computational modelling framework can be used to quantify behaviour in a free-operant VR context.

Figure 2.

Figure 3.

Skin conductance analysis dissociated decision values and subjective ratings for phasic pain

As an objective measure derived from empirical behavioural data, the phasic pain utility function generated by the model can be abstract. To understand its physical implications, we further examined its relationship with subjective pain ratings and their correlations with physiological responses, specifically SCRs. Figure 3(E, F) presents the subjective ratings and decision values of phasic pain across different conditions, while Figure 3G demonstrates a significant correlation between decision values and subjective pain ratings.

Two types of evoked SCRs were analysed: those elicited by the pick-up action (which triggered a shock for painful fruit) and those associated with the fixation action (looking at either painful or non-painful fruit). Figure 4A compares evoked SCRs for painful and non-painful fruit. To quantify the magnitude of evoked SCRs which may overlap in the time domain, given that changes in skin conductance can be approximated by a linear time-invariant system, we fitted the skin conductance data to a constructed time series by convolving the event trigger with a known canonical response function (CRF) (Bach et al., 2010). We selected fixation and pick-up events for painful fruit and extracted their fitted coefficients. Analysis using a multilevel linear mixed-effects model revealed a clear dissociation in the relationship between physiological responses and motivational parameters. Fixation-evoked SCR coefficients were significantly associated with decision values, but not with subjective pain ratings (Fig. 4B). Conversely, shock-evoked SCR coefficients showed a significant association with subjective pain ratings, while the association with decision values was not significant (Fig. 4C). This double dissociation suggests a notable divergence between the physiological correlates of expected utility (at the decision level) and experienced utility (the actual pain experience). Taken together, these findings highlight the composite nature of the overall aversiveness of pain and underscore the benefit of combining subjective ratings with model-based measures to capture its distinct impacts on behaviour.

Figure 4.

Experiment 2: Modulation of free-operant foraging by tonic pain stimulation

In the second experiment, we introduced a tonic pain condition by applying an inflatable blood pressure cuff around the non-dominant arm (Graven-Nielsen et al., 2017). We reduced the number of phasic pain levels from five in the first experiment to three (no pain, low pain, and high pain). These three phasic pain levels were then combined with two fixed tonic pain conditions (with and without tonic pain) to form a factorial design experiment with a total of six different combinations. Similar to experiment 1, the experiment was divided into one-minute blocks, and the pain conditions were fixed within each block. The six combinations were initially presented in a fixed order as training blocks, then the order was randomised and repeated three times, resulting in a total of 24 blocks. To obtain a neural measure of pain behaviour, we introduced concurrent EEG recording alongside the VR set-up.

No effect found for tonic pain on phasic pain

We found no significant modulation of phasic pain ratings by tonic pain (Fig. 5(A, B)). We found a basic decrease in aversive choice probability, in terms of the probability of selecting a painful fruit, as in experiment 1. The average aversive choice probabilities were similar in conditions with and without tonic pain (Fig. 5(C-E)), providing evidence that punishment sensitivity was not affected by the presence of tonic pain.

Figure 5.

In the neural data, we found that phasic pain intensity modulated the amplitude of phasic pain ERPs, as would be expected. But as in the behavioural results, this was not modulated by the presence/absence of tonic pain (Fig. 6). That is, tonic pain neither facilitated nor inhibited the behavioural or neural responses to the phasic pain. We focused our neural analysis of phasic pain on ERPs as phasic stimuli are well characterised by these time-locked evoked potentials. Nevertheless, to ensure a comprehensive assessment of the neural response, we also examined induced oscillatory responses. These results were consistent with the ERP findings and are detailed in the Supplementary Materials (Fig. S4, S5).

Figure 6.

Tonic pain reduced motivational vigour

By analysing the motion tracking data, we found that tonic pain was associated with a reduction in task-related movement velocities (Fig. 7(A, C)) and an associated reduction in fruit collection rates (Fig. 7B).

Figure 7.

To quantify this effect in terms of motivational vigour, we extended the computational model to accommodate vigour effects. The model presented for experiment 1 does not consider time, and fitting the model to the collected data requires acausal simulation (see Methods - model-fitting section below). For experiment 2, we took account of the temporal factor and fitted the model causally with high temporal resolution. Niv et al. (2007) proposed that the delay between consecutive actions performed by animals can be formulated as a trade-off between opportunity costs and vigour costs. This framework was developed based on earlier works in average-reward reinforcement learning, showing that this trade-off approach maximises average reward (Niv et al., 2005). In this formulation, animals choose an action pair (a, τ). The first, a, is the common action in discrete decision-making models, and the separate τ represents the time delay between two actions. We built upon this model to suit the requirements of our study. Similar to the model in experiment 1, our action a represents the choice of which fruit to pick up. To account for the different effort costs incurred for each fruit, we again used a distance-based effort cost term C_m · d(a), as in experiment 1. Additionally, a time delay between actions is required to model vigour. We assumed that subjects estimated the time delay by estimating their speed V_speed and calculated the delay based on the distance to the fruit and the estimated speed. Hence, for a decision time point state s ∈ 𝒮, if the effort cost is inversely proportional to the time delay, the optimal differential value of an (a, τ) pair is

, where τ is the additional delay accounting for the time waited before committing to execute the action. s^′ ∈ 𝒮 is the deterministic subsequent state after choosing (a, τ). Q^* is the optimal differential state-action value function. C_v is the vigour constant that scales the vigour cost term. is the average reward and assumed to be constant (Fig. 8).

Figure 8.

We fitted vigour constants C_v and other parameters for no tonic pain and with tonic pain conditions separately. We found the fitted vigour constants were significantly higher in tonic pain conditions (Fig. 9A). We further fitted the model with additional separate vigour constants for different levels of phasic pain. Consistent with the fitting results over tonic pain conditions only, tonic pain showed a strong effect on fitted vigour constants, while phasic pain did not show a significant effect (Fig. 9B).

Figure 9.

A unified model for tonic and phasic pain for decision-making in free-operant task

This new decision model therefore includes both phasic pain and vigour costs, by which the tonic pain effects can be expressed as change in vigour. Experiment 2 replicated the phasic pain punishment sensitivity effect shown in experiment 1. Using this model, the fitted phasic pain parameters also showed no significant differences in punishment sensitivity (Fig. 9(C-E)), in line with the behavioural results presented above. The vigour constants and phasic pain function fitting results illustrate how we can combine two dissociable effects of pain motivation in a single model, consistent with the results from empirical regression.

Characterisation of neural activity during tonic pain

To probe the neural activity associated with the modulation of motivational vigour by tonic pain, we performed LMM-based time-frequency analysis on the EEG data. The data were epoched based on key decision points. Linear mixed models (LMM) were employed to account for random effects across subjects and to control for motion artefacts by including instantaneous head movement speed as an additional predictor. We found that alpha band power overlaying central-parietal and temporal regions (CP6 and T8), as well as beta power in parietal scalp regions (P8), was strongly associated with the presence of tonic pain (Fig. 10A).

Figure 10.

To investigate how tonic pain impacts vigour, the vigour constants separately fitted to tonic pain conditions (Fig. 9A) were fit to the EEG data. We found that the alpha band power overlaying the midline parietal region (Pz) was significantly negatively correlated with the vigour constants, but it did not survive the multiple comparison correction (Fig. 10B).

Discussion

The experiments show that tonic and phasic pain serve different motivational functions during adaptive behaviour, in line with ecological and evolutionary theories of pain (Bolles and Fanselow, 1980; Walters and Williams, 2019). Specifically, our findings point towards phasic pain providing a punishment teaching signal that directs avoidance through value-based learning, balancing the cost of future harm alongside potential reward. This is supported by the observation that increasing phasic pain intensity significantly reduced choice probability and increased distance bias between choices, whereby participants were willing to travel further to reach a pain-free fruit. In contrast, we found that tonic pain reduces motivational vigour, which supports energy conservation and recuperation in the context of bodily damage. This claim is directly evidenced by the reduction in task-related movement velocities and fruit collection rates during tonic pain blocks. The experiments are the first to show that these two functions can be formally distinguished and quantified during ongoing behaviour. By utilising a free-operant RL computational framework, we were able to dissociate these roles: phasic pain was quantified as a generally negative utility term affecting choice values, while tonic pain was formalised as a change in vigour constants that were significantly higher (increasing delays between actions) in tonic pain condition. This illustrates how pain simultaneously acts in different ways to serve self-protection.

One notable aspect of our results is that we did not see interactions between tonic and phasic pain at either the behavioural or neural level. Behaviourally, we observed that average aversive choice probabilities remained similar regardless of the presence of tonic pain, with no significant interaction effect on punishment sensitivity. Furthermore, our model-fitting confirmed that tonic pain did not significantly modulate the fitted phasic pain utility values. There are two contexts in which these might be predicted. First, in ‘conditioned pain modulation’ paradigms (Kennedy et al., 2016), a tonic pain stimulus is sometimes seen to reduce both the perceived intensity and the cortical evoked responses to phasic pain stimuli delivered somewhere else on the body (Höffken et al., 2017; Enax-Krumova et al., 2020). Although we utilised concentric ‘Wasp’ electrodes designed to selectively activate nociceptive A-delta fibres (Inui et al., 2002), and confirmed that the resulting ERPs (N1-P2) were significantly modulated by phasic intensity, we observed no such attenuation by tonic pain. Indeed, neither subjective pain ratings nor the N1-P2 amplitude showed a significant modulation by the tonic pressure pain stimulus. In contrast, our results were more compatible with a trend in the other direction.

This finding raises the (testable) question as to what extent conditioned pain modulation depends on the task setting: notably, here we utilise an active in contrast to a passive task, which is known to be an important distinction in other contexts of coping behaviour (Bandler et al., 2000). In an active behavioural context, the brain may prioritise the fidelity of phasic pain as a teaching signal for learning, potentially overriding modulation seen in passive CPM paradigms. A second context in which phasic-tonic interactions are possible is in punishment sensitivity, where it is proposed that tonic pain signals increase vulnerability to new phasic pain insults (Seymour et al., 2023). Punishment sensitivity is seen in terms of anxiety-like responses to novel threats in animals (Crook et al., 2014; Lister et al., 2020), and also in decision-making tasks in people with chronic pain (Mancini et al., 2024). However, we do not see an increase in the punishment value of phasic pain in relation to tonic pain here.

These two contexts highlight the potential difference between ratings and choice as measures of aversiveness, a concept that is sometimes referred to as experienced utility and decision utility (Kahneman and Tversky, 1984), or disliking and unwanting (the negative version of the liking-wanting distinction (Dayan and Seymour, 2009; Berridge et al., 2009), or put colloquially, ‘what you say’ versus ‘what you do’). In our task we get a further hint of this in the SCR measures in experiment 1, whereby a discrepancy exists between decision values and pain ratings in their respective associations with fixation-evoked SCRs and phasic pain-evoked (shock) SCRs. Taken together, this indicates the composite nature of overall aversiveness of pain, and highlights the benefit of combining subjective ratings with model-based measures of its motivational impact on behaviour.

The experiments show how probing objective motivational functions of pain is facilitated by VR. Conventionally, pain motivation is studied using desktop tasks in which decisions and physiological responses are made by a seated subject, responding via a keyboard with pain stimuli applied to a distant site. VR provides several useful advances. First, the ability to apply pain stimuli in a way that is appropriate to the action, i.e. phasic pain stimuli to a grasping hand, providing sensorimotor congruency. Second, the ability to embed pain within a free-operant and free-moving physical task allows evaluation of full-body responses to a topographic stimulus, and it is this in particular that allows assessment of motivational vigour. Third, the provision of a multisensory immersive environment entirely transforms a pain task into a naturalistic setting that matches the experience of pain in real-life. The importance of naturalistic contexts is well documented in fear learning, although less well studied for pain (Mobbs et al., 2015).

However, the use of VR requires an analysis framework that accommodates the complexity of free-operant and free-moving nature of behaviour. This challenge can be overcome with a computational approach and judicious use of simplifying assumptions where appropriate. The free operant model exploits a theoretical framework developed for animal studies that derives action delays from a trade-off between vigour cost and opportunity cost (Niv et al., 2007). This model has been applied to human studies (Nair et al., 2023), albeit in a much simpler setting involving perfect information and a single instantaneous action. In our task, the environment was partially observable and evolved as subjects interacted with it. By utilising eye-tracking to extract an abstract observation space and a modified version of the model that enabled simulation at high temporal resolution, we demonstrated that it is possible to predict free-operant behaviour in real-time. Compared to overall speed and collection rate, which can be influenced by multiple factors, such as different choice sets available to participants as the fruit locations are randomly generated, the model’s fitted parameters (e.g. vigour constant C_v) in theory serves as a direct, concrete estimate of that internal state. This approach revealed critical insights into a subject’s internal pain and motivational states.

A concern that is sometimes raised in pain experiments is to what extent the impact of pain on other behavioural and cognitive processes might be interpreted as an attentional phenomenon (e.g. distraction, or cognitive load), and therefore not specific to pain per se (Moont et al., 2010), specifically the reduction in vigour observed here. One argument to this point posits that attention is an intrinsic part of what pain is, and therefore it does not make sense to consider pain as an operational system without attention (Van Damme et al., 2010): i.e. tonic pain may mediate vigour effects by using distraction as a functional mechanism, and therefore attention is a core part of pain, and not a confound. One way to test this would be to add an additional condition comprising a tonic salient but non-painful stimulus (such as a continuous salient auditory stimulus), although this introduces other confounds related to the incommensurability of such dissimilar types of stimulation. Attentional arguments have also been proposed in CPM paradigms, but here it has been noted that CPM (and diffuse noxious inhibitory control) operates at least partly at a spinal/brainstem level, at ‘lower’ levels than those typically associated with attention (Youssef et al., 2016; Kucharczyk et al., 2023).

It is important to acknowledge that the signal-to-noise ratio in both our physiological and neural recordings is lower than that typically observed in conventional, stationary laboratory experiments (Gramann et al., 2011). This is primarily due to the motion artefacts inherent in an immersive and active virtual reality environment. Whilst we utilised robust cleaning and artefact-correction methods (Klug and Gramann, 2021), the elevated noise floor may limit our capacity to detect more subtle neural effects or interactions. These challenges highlight a critical area for future methodological research, particularly in the development of hardware and signal-processing tools designed to isolate neural signals during complex, mobile behavioural tasks.

Despite the greater levels of noise in EEG recordings due to our highly mobile task, an LMM-based approach (that incorporates head-movement speed into the model) revealed that suppression of oscillatory power in the alpha and beta bands in contralateral parietal and temporal regions during tonic pain was consistent with previous studies on tonic pain (Ferracuti et al., 1994; Huishi Zhang et al., 2016; Dowman et al., 2008; Schulz et al., 2015). When comparing how the neural correlates of tonic pain impacted motivational vigour, no oscillatory changes survived correction for multiple comparisons—although a trend towards reduced alpha power was observed at parietal scalp regions. Changes in alpha band power have been shown to affect neural activity in areas associated with sensory integration, attention, and motor planning (Hu et al., 2013; Palva and Palva, 2011), all of which are linked to motivational vigour.

It is also important to consider the spatial configuration of the stimuli used in this study. Phasic pain was delivered to the grasping hand to maintain spatial congruency with the virtual fruit, ensuring a coherent nociceptive feedback signal for the interactive task. Additionally, tonic pain was applied to the contralateral arm to prevent mechanical interference with motor execution, which would have occurred if pressure were applied to the ipsilateral limb used for grasping the controller. Whilst this design promotes spatial congruency and avoids mechanical confounds, future studies might explore how these effects generalise across different body parts, for which VR experiments serve as a promising tool to test relevant hypotheses (Hewitt et al., 2026).

Finally, the methodological framework has clinical applicability as a way to objectively measure pain behaviour in chronic pain. For instance, while conventional clinical trials often rely on Quantitative Sensory Testing (QST), which can deviate from real-world scenarios, our immersive VR task provides an ecologically valid and behaviourally sensitive evaluation of pain as a function of diagnosis or treatment. In addition, it could be implemented using real-time analysis to provide a feedback signal for closed-loop interventions, such as interactive VR rehabilitation or deep brain stimulation.

Methods

The free-operant foraging task

Foraging theory is a broad and well-studied subject within animal behaviour research. Various foraging study paradigms have been established in both laboratory and natural environments (Constantino and Daw, 2015; Charnov, 1976; Lihoreau et al., 2011). A natural foraging task, while simple for a healthy human to perform, encapsulates key components relevant to pain research. Firstly, foraging is goal-directed; subjects actively search for food under varying levels of motivation, allowing for the examination of the interaction between motivation and pain. Secondly, a natural foraging task involves full-body movement. Given that many types of pain are closely linked to movement and exercise, motion capture technology—a fundamental component of immersive VR—allows for precise recording and analysis of movement under different pain conditions. Finally, foraging involves exploring the environment and learning an optimal strategy to maximise rewards, naturally linking pain to cognitive functions. Motivated by these considerations, we designed the free-operant foraging task aimed at simulating realistic foraging behaviours.

At the beginning of each one-minute block, a total of 150 virtual pineapples of varying heights from 0.33 to 1 m were randomly generated in a circle centred around the participant with a diameter of 6.67 m. Five identical baskets were placed within the space. Spatial locations of trees and vegetation were generated using the game engine’s default tree painting tool (Unity Technologies, San Francisco, US).

While the colour association (green for painful, yellow for pain-free) was not counter-balanced across subjects, any inherent aversive value of green pineapples (e.g. as ‘unripe’ fruit) is expected to have a minimal confounding effect on the analysed data. In associative learning frameworks, while mild prior biases may influence initial value estimations, extensive training with a highly salient unconditioned stimulus (e.g. phasic pain) rapidly updates these values, driving them toward an asymptote determined entirely by the explicit task contingencies (Rescorla and Wagner, 1972; Sutton and Barto, 2018). Because participants underwent extensive training (10 blocks in Experiment 1 and 6 blocks in Experiment 2) to establish the explicit pain associations prior to the analysed sessions, the observed avoidance behaviour was predominantly driven by the learned phasic pain contingencies rather than baseline colour preferences.

We used the HTC Vive Pro Eye headset and controllers in combination with Valve’s SteamVR platform to track user motion in our Unity-based virtual environment. The system employs Lighthouse tracking, which uses infrared laser-emitting base stations and IR photodetectors to determine the position and orientation of tracked devices. Motion data from the headset and controllers were accessed in Unity via the SteamVR Unity plugin, which maps physical movements to virtual objects in real time. This setup enables accurate and low-latency motion representation in the virtual space, supporting immersive interaction during the experiment (Sansone et al., 2022). Data acquisition in Unity occurred at the rendering frame rate. The average delay between consecutive frames was 22.49 ± 5.82 ms.

Participants

A total of fifty-eight healthy subjects with no significant mobility issues were recruited from community and a pool of university staff and students. Twenty-five subjects were recruited for experiment 1; one subject was excluded due to technical failure. The final sample included 24 subjects (17 female, mean age 25.71 years). Thirty-three subjects were recruited for experiment 2. Two subjects were excluded due to technical failure and compliance failure, respectively. The final sample included 31 subjects (15 female, mean age 24.81 years). An a priori power analysis was not conducted due to the novelty of the investigation and the complexity of the analyses. Instead, we based our target sample size (N ≈ 30 per experiment) on previous studies using computational modelling of neurophysiological data (Mahajan et al., 2025), as well as EEG, SCR, and pain studies (Schulz et al., 2015; Zhang et al., 2018), and studies from our group using combined neurophysiological recordings and VR (Hewitt et al., 2026). This approach represents a pragmatic balance that ensures the credibility of the results and the stability of model estimates while accounting for the high per-subject cost and depth of data collected from each individual.

The procedure was approved by the Research Ethics Committee of the University of Oxford (CUREC approval reference R58778), and all participants gave written informed consent at the start of the experiment following the Declaration of Helsinki. Participants were reimbursed with £30-£45 depending on the duration of experiment, including a £10 performance incentive. Participants were informed at the start of the experiment that their total points would be rewarded with a monetary incentive of up to £10. To maintain a constant level of motivation throughout the task, the exact point-to-currency exchange rate was not specified. Upon completion of the session, all participants were awarded the maximum bonus of £10.

Statistical Analysis

Statistical analysis was carried out using custom scripts written in Python 3 and R 4.4.2. For LMMs, lme4 and lmerTest packages were used (Bates et al., 2015; Kuznetsova et al., 2017). For LMMs presented in Figure 3, Figure 4, Figure 5, and Figure 10, fixed effects, random effects for intercepts and independent variables (IV) were estimated. The model also estimated the correlation between the intercept deviations and IV effect deviations across subjects. Supplementary materials provide additional details of the formula of the LMMs and their detailed outputs.

Model description and fitting

Experiment 1

As we disregarded planning, the model in experiment 1 compared C_pu(a)+C_m · d(a) for each fruit at a particular time point. The first question was what fruits were available for subjects to choose from. Fruits were abundant in the subjects’ reachable area, however, many were not directly in subjects’ visual field or hidden behind the trees. Subjects also needed to identify the fruits’ colour to know whether the fruit was painful or not. Therefore, there was an additional search (or exploration) process that was not modelled here. We utilised eye-tracking technology (HTC Vive Pro Eye), and assumed only fruits that were in the central visual field were considered by the subject. Fruits that were captured by eye-tracking were saved into a first-in-first-out memory queue (the available choice set). We heuristically set the memory size to be 7, and the memory was cleared each time the subject chose a fruit. Next question was when the subject made a decision. The first time point that the subject looked at a pineapple not seen before could be used as the decision point in this free-operant task. However, we could not evaluate the model when the chosen fruit was not in the memory yet, as the model would never make a correct prediction out of a set of wrong candidates. We could only retrospectively identify the fruit that the subject chose and only fitted the model when the correct choice was in the choice set. This acausal problem is difficult to solve without taking time into consideration. We will show in the experiment 2 model fitting, by quantifying the effort and opportunity costs as a function of time delay, the model-fitting process was made causal.

We had converted the problem into finding a set of hyperparameters so that at each decision point, the maximum value of all choice differential values C_pu(a) + C_m · d(a) predicts the choice that the subject choose. Readers might find the problem similar to the multinomial logistic regression. However, as a decision model, we chose the objective function for the model-fitting to be the error rate (number of correct predictions divided by number of total predictions) rather than weighted least squares. The model is also flexible to take arbitrary parameters other than linear coefficients. As for this experiment, we chose u to be a sigmoid shape function with two additional x scale and translation hyperparameters. Grid search was used as a general method to find the hyperparameters that minimises the error rate. Grid search parameters can be found in supplementary materials.

To provide a detailed description regarding the computational framework, we summarise the properties of the first optimality equation and the underlying state-action space in Table 2.

Table 1.

Table 2.

Table 3.

The simulation and model-fitting process are conducted blockwise. Because the phasic pain intensity x_stim for green fruits is fixed for the duration of a single block, this information is treated as a known variable to the model rather than being explicitly encoded into the state space s. A discussion on the choice of constant x_stim for the duration of a single block is added in the Supplementary Material section: Discussion of Pain Intensity Information and Model Robustness.

Furthermore, we clarify the distinction between parameters and hyperparameters in this framework. Unlike traditional reinforcement learning models where internal parameters (such as action values or learning rates) may be continuously updated through trial-and-error, we have eschewed learning to maintain simplicity and focus on the steady-state motivational trade-offs. Consequently, there are no “online” parameters; the hyperparameters listed in Table 2 are the only variables optimised via the fitting procedure to describe the subject’s stable behavioural preferences across the experimental conditions.

Experiment 2

A key assumption in Niv et al. (2007) was that animals chose an action pair of an action and a delay. The delay included both the time waiting and the time executing the action. It was an approximation of real animal behaviour, and in fitting to real human subjects data in this task, we did not have direct access to subjects’ delay decision. Therefore, we tracked the delay and computed the differential value at each time point. The central assumption in this model fitting solution was all differential values for the choices that were not chosen should be less than the value for the final chosen choice. The model-fitting program tracked the model predicted value for each fruit that a subject fixated on. The model’s predicted choice was correct (i.e., fit the data) when the chosen fruit had the highest predicted value (compared to other fruits within the time interval before the last fixation on the chosen fruit). Otherwise, the model’s predicted choice was incorrect. γ is still set to 0 for model-fitting in this experiment.

Experimental pain stimulation

Electric phasic pain

The electric stimulation pulses were generated by a DS5 isolated bipolar constant current stimulator (Digitimer, Letchworth Garden City, UK) and delivered using a Wasp pain stimulation electrode. These electrodes preferentially activate nociceptive A-delta fibres, thereby eliciting ERPs that more accurately reflect nociceptive processing compared to standard bipolar stimulation (Inui et al., 2002; Mørch et al., 2011). The pulses were 200 Hz square waves with a 1 ms pulse width.

Calibration of stimulation intensity was conducted based on subjects’ pain ratings before the task commenced. To maintain consistency in pain ratings, additional calibration could be performed at the end of the 10th block in experiment 1 or at the end of every six blocks in experiment 2. Subjects rated their pain on a 0-10 numerical rating scale while researchers adjusted the stimulation intensities following a staircase procedure. The calibration process was considered complete when subjects provided consistent pain ratings within the desired range, as determined by the experimental design (8/10 in experiment 1 for the max pain condition; 3/10 in experiment 2 for the low phasic pain condition and 7/10 for the high phasic pain condition). At the end of each block, subjects provided a pain rating for that block. Ratings in both procedures were given verbally. During calibration, researchers recorded the subjects’ verbal ratings, whereas in post-block ratings, verbal responses were processed using speech recognition services (Microsoft, Redmond, US) and automatically converted into a visual analogue scale bar displayed in VR. In case of speech recognition failure, researchers wrote down the ratings and manually updated the data for analysis.

Pressure tonic pain

We used pressure cuffs (VBM, Sulz, Germany) to safely induce tonic pain (Graven-Nielsen et al., 2015; Joseph et al., 2022). The cuff was wrapped around the non-dominant arm’s biceps. The cuffs were inflated prior to the start of each required block and immediately deflated after the block ended. Subjects were asked to report verbally when they felt discomfort and when they felt pain as the researcher manually inflated the cuff. Inflation stopped once the subject reported pain. To control for variables associated with wearing the cuff other than pain, pressure cuffs remained wrapped around the subjects’ arms even in blocks without tonic pain.

Skin conductance

Data collection and processing

Experiment 1 was designed to establish the robust behavioural effects of the foraging task while ensuring the collection of reliable physiological data. We chose SCR as it is a well-validated index of autonomic arousal that we were confident would provide a clear peripheral measure of pain-related processing in this novel VR paradigm. In Experiment 1, SCRs were collected using a custom-built wireless sensor. The microcontroller was an Arduino Nano 33 IoT (Arduino SA, Chiasso, Switzerland). The skin conductance sensor was a Grove GSR sensor (Seeed Studio, Shenzhen, China). The firmware was written by the authors of this article with a designated sampling rate of 1000 Hz.

Data was segmented by experimental blocks (60 seconds long each). An entire block of data was removed from further analysis if the data loss was more than 20% of the expected total number of samples due to temporary wireless transmission failure. Each block of data was down-sampled to 10 Hz, and re-centred by the average baseline. A median filter with a window size of 5 was applied to the down-sampled data to further remove artefacts. Further small gaps of missing data were handled in the GLM fitting in R with na.action set to na.omit. As subjects would only know the intensities of pain stimulation in new blocks after picking up a painful fruit, for fixation events, only events after picking up the first painful fruit were analysed.

Effect coefficient estimate

Generalised linear models (GLM) were used to fit the skin conductance time series data to the constructed time series data. The time series was generated by convolving the event trigger with the CRF, with hyperparameters taken from (Bach et al., 2010). Fixation events that occurred during picking up painful fruit were not included as shock-evoked SCR was strong. For computational practicality, convolutions were truncated between 0 and 60 seconds (the length of an experimental block) (Bach et al., 2009). The GLMs employed a Gaussian distribution with an identity link function. Each GLM was fitted separately to individual blocks, generating its own set of coefficient estimates. Fitted coefficients were thresholded based on a maximum p-value of 0.05. Additionally, coefficients with z-scores greater than 3 were excluded. The remaining coefficients, including negative ones, were used for further correlation analysis, as our focus was on the relationship between decision values, ratings, and coefficients.

EEG

Data collection and processing

Whole-scalp EEG data was continuously recorded with 32-channel LiveAmp system (BrainProducts GmbH, Munich, Germany) at 500Hz. We chose FCz as the reference. To remove ocular artefacts, recordings were first cleaned with EEGLAB (v2023.1) by removing channel drift (linear filter (FIR) transition band from 0.25 to 0.75 Hz). It also automatically removes bad data periods by thresholding max acceptable 0.5 second window standard deviation of 20, max acceptable channel RMS range of 7, and maximum out-of-bound channels of 25%. The ICA was then performed on the cleaned data and the weight matrix was computed and saved separately. Ocular artefact components were identified manually, assisted by automated classifier ICLabel (Pion-Tonachini et al., 2019). We then cleaned ocular artefacts by removing these identified components back projected onto the raw data. By manually inspecting the data, we interpolated up to 10% (3 channels) of electrodes. The output data at this stage served as the input data for ERP and time-frequency analysis separately.

ERP analysis

ERP analysis was conducted on the preprocessed data in Python with MNE-Python 1.7.0 package. The data was re-referenced to the average amplitude of all electrodes. A bandpass filter from 1 to 30 Hz was applied. The high cutoff frequency was chosen because of the presence of strong motion artefacts in our mobile EEG data. Identical analysis was applied at 0.1Hz and 0.5Hz high-pass cutoff frequencies, and significant results remained unchanged (see Supplementary figures). Recordings were processed into 1s epoch with 0.3s baseline before the event trigger. Epochs with peak-to-peak amplitude exceeding 200 µV in Cz channel were excluded (average rate of excluded epochs was 2.81%). We searched for N1 amplitude peaks from 100-170ms and P2 amplitude peaks from 140-300ms at channel Cz (Favero et al., 2023; van den Broeke et al., 2010).

LMM-based time-frequency analysis

Time-frequency analysis was performed on preprocessed data in Python with custom scripts relying on MNE-Python 1.7.0 package. The data was re-referenced to the average amplitude of all electrodes. A notch filter with 4Hz width centred at 50Hz was applied. Data were epoched from 0-0.5s according to the decision points. The decision points were chosen when subjects fixated on the chosen fruit for the first time, or at a re-evaluation point. A re-evaluation point was a subsequent fixation on the chosen fruit when other fruit had higher value than the chosen fruit but the new fixation on the chosen fruit surpassed other fruits’ value. These time points were assumed to represent the key moments for decisions and thus chosen for analysing correlated neural activities.

The spectral power was estimated using Welch’s method. A Hamming-tapered window was used with the window length equal to the epoch length (250 samples, 0.5 second). The estimated spectral power was separated into four bands: theta 4-7 Hz, alpha 8-12 Hz, beta 13-29 Hz, gamma 30-100 Hz (Levy et al., 2023; Chrastil et al., 2022; Schulz et al., 2015), and averaged within each band range. These band-averaged values formed the basis for the LMM-based time-frequency analysis presented in the main text, which was designed to statistically account for the task’s complex noise profile. For each power band, averaged band power was inputted into an LMM for each channel as the dependent variable (Schulz et al., 2015). For each epoch, the z-score of the averaged band power was computed for each channel within the subject, and the epoch would be removed from LMM fitting for this band if any channel had a z-score greater than 3 (Nolan et al., 2010). For tonic pain effect, the binary condition of tonic pain presence was used as an independent variable. Random effects for tonic pain and intercepts were estimated. The model also estimates the correlation between the intercept deviations and tonic pain effect deviations across subjects. Crucially, a single head movement speed variable was added as an additional independent variable to reduce motion artefacts. The head movement speed for a particular epoch was calculated by averaging the speed between 10 Unity frames before the decision point and 10 Unity frames after the decision point. For vigour constants, the LMM had the same structure except replacing the binary tonic pain condition by fitted vigour constants presented in Figure 9A. Source analysis for spectral power was included in the Supplementary materials (section: Source Analysis). Furthermore, the analysis of induced oscillatory responses to phasic pain stimuli which supplement our ERP findings is provided in the Supplementary material (section: Induced oscillatory responses to phasic pain stimuli).

Data availability

Complete analysis code and raw data are released on GitHub: https://github.com/ShuangyiTong/Phasic-and-tonic-pain-serve-distinct-functions-during-adaptive-behaviour A series of software tools to replicate the experiment is available for download on GitHub: https://github.com/ShuangyiTong/PineappleStudy2025ReplicationSoftware

Acknowledgements

ST would like to thank Pranav Mahajan, Yijia Yan, Suyi Zhang, and Eoin Kelleher for their feedback on the content of this paper during the work-in-progress stage. ST also thanks Kirsty Bannister, Joseph Taylor, Kristian Hennings for sharing valuable knowledge on the safe use of pressure cuffs to induce tonic pain. Special thanks to Eoin Kelleher for providing an exceptional visual representation of the task (Fig. 1A). ST would like to thank the reviewers of IASP 2022, CCN 2023, and COSYNE 2024 for their feedback.

The work was funded by Wellcome Trust (214251/Z/18/Z, 203139/Z/16/Z and 203139/A/16/Z), IITP (MSIT 2019-0-01371, RS-2023-00233251), JSPS (22H04998), and EPSRC (EP/W03509X/1). This work was also funded by the EPSRC and MRC Research Grant Scheme under the reference number UKRI1970. This research was also partly supported by the NIHR Oxford Health Biomedical Research Centre (NIHR203316) and the Royal Academy of Engineering. The views expressed are those of the author(s) and not necessarily those of the NIHR or the Department of Health and Social Care. For the purpose of open access, the author has applied a CC BY public copyright licence to any Author Accepted Manuscript version arising from this submission.

Additional information

Author contributions

ST and BS designed the theory and experimental protocols. ST, BS, and TD developed the VR research platform used for the experiments. ST collected and analysed the data. ST and DH designed the EEG analysis pipeline and methodology. SL and TD provided critical feedback and theoretical insights throughout multiple iterations of the study. ST and BS wrote the first draft. ST, TD, DH, SL, and BS edited and approved the manuscript. TD, SL, and BS acquired funding for the study.

Funding

Wellcome Trust (WT) (214251/Z/18/Z)

• Ben Seymour

Wellcome Trust (WT) (203139/Z/16/Z)

• Ben Seymour

Wellcome Trust (WT) (203139/A/16/Z)

• Ben Seymour

MSIT | Institute for Information and Communications Technology Promotion (IITP) (2019-0-01371, RS-2023-00233251)

• Ben Seymour

• Sang Wan Lee

MEXT | Japan Society for the Promotion of Science (JSPS) (22H04998)

• Ben Seymour

UKRI | Engineering and Physical Sciences Research Council (SRC) (EP/W03509X/1)

• Ben Seymour

NIHR Oxford Health Biomedical Research Centre (NIHR203316)

• Ben Seymour

Royal Academy of Engineering (RAENG)

• Tim Denison

EPSRC and MRC Research Grant Scheme (UKRI1970)

• Tim Denison

• Ben Seymour

Additional files

Supplementary Materials.