Reshaping restriction: virtual embodiment modulates eating biases in anorexia nervosa Authors: M Dimakopoulou, T Ciorli, M Pyasik, M Da Col, F Boggio, C Andriulli, F Bevione, M Romanisio, M Martini, M Guerraz, G Abbate-Daga, L Pia Published: 2026 ABSTRACT AbstractAnorexia nervosa (AN) involves extreme food restriction and body image disturbances, partly sustained by altered responses to food cues, including implicit and behavioral avoidance, especially toward high-calorie foods. We tested whether altering body ownership in virtual reality could modulate such biases. Female participants with restricting-subtype AN (AN-R; n = 29) and healthy controls (HC; n = 31) completed three sessions: a baseline and two full-body illusion (FBI) sessions in which they embodied slimmer or larger avatars than their own body size. Across sessions, they performed a binocular rivalry task and a food-specific approach-avoidance task with high- and low-calorie foods, together with body-size estimation and symptom measures. At baseline, AN-R participants showed greater perceptual salience and stronger avoidance of high-calorie foods than HC. Relative to the baseline, embodying a larger avatar attenuated avoidance in AN-R, whereas HC showed the opposite pattern. These findings suggest that the FBI can transiently modify implicit food-related avoidance in AN-R.Subject terms: Neuroscience, Psychology, Psychology Introduction Anorexia nervosa (AN) is a complex and severe mental disorder marked by an intense fear of gaining weight, a distorted body image, and extreme restriction of food intake1. Maladaptive eating behaviors in AN are supported by heightened cognitive engagement and strong, often distressing emotional responses to food stimuli2,3. Individuals with AN typically avoid high-energy-dense foods and assign heightened value to low-calorie alternatives4,5. A recent study6 showed that these food attitudes are rooted in lower-level cognitive mechanisms. Specifically, high-calorie foods were perceptually prioritized in individuals with the restricting AN (AN-R) subtype but were subsequently avoided at an implicit response initiation level. This paradoxical pattern of initial hypervigilance followed by automatic avoidance may temporarily reduce anxiety, but over time, it likely reinforces restrictive eating behaviors and maintains the disorder7. Given that such distorted behaviors present significant challenges to effective treatment8,9, approaches targeting earlier stages of visual and behavioral processing are of considerable clinical relevance. In recent years, Virtual Reality (VR) has offered new possibilities for both studying and potentially modulating distorted behaviors that maintain eating disorders10–12. An emerging technique is the Full-Body Illusion (FBI) paradigm, which elicits a sense of illusory body ownership by inducing participants to experience inhabiting a virtual avatar13–15. This is typically achieved by presenting a three-dimensional avatar from a first-person perspective (1PP) in VR, visually substituting the participant’s real body. In healthy individuals, it has been demonstrated that when the avatar’s body size is slimmer or larger than the participant’s actual body, individuals often experience a corresponding shift in perceived body dimensions and a compelling feeling of owning that virtual body16–21. Beyond changes in perceived body size, FBI and related VR body-exposure paradigms can induce broader cognitive and emotional changes22. For example, embodying larger virtual bodies has been associated with increased preference for energy-dense foods23, whereas slimmer avatars have been linked to avoidance of high-calorie food stimuli21. This evidence supports the idea that the embodiment of a specific body size may dynamically activate eating-related cognitive and behavioral tendencies associated with that body representation by default (see ref. 22 for an extensive discussion). In clinical samples, VR body-exposure and embodiment of larger or healthy-weight bodies are used as exposure-based techniques to deliberately elicit fear of gaining weight and body anxiety in a controlled setting, and, over repeated sessions, can reduce fear of weight gain and other AN symptoms24–27. This emotional activation indicates that such procedures need to be carefully framed and integrated into treatment25–27. Thus, while embodiment-based techniques are promising, their therapeutic application requires careful consideration of both potential benefits and risks. Much less is known about whether FBI procedures can influence earlier stages of food processing in AN, such as visual awareness and automatic approach-avoidance tendencies toward high- and low-calorie foods. These mechanisms may be less accessible to introspection yet still shape everyday eating behavior6,28. Specifically, we focused on two implicit aspects of visual processing: awareness and automatic associations. Visual awareness was assessed using the Binocular Rivalry paradigm (BR29, which measures conscious prioritization between competing stimuli, such as high- and low-calorie food images30. Implicit associations were examined using the Food Preferences Approach-Avoidance Test (FP-AAT)31, which captures automatic approach-avoidance tendencies toward different food categories. Based on previous findings, we expected that, at baseline, AN-R participants would show greater perceptual salience and stronger avoidance of high-calorie foods than HC6. For AN-R, we further hypothesized that embodying a larger avatar would attenuate these restrictive food-related biases, thereby transiently disrupting the rigid avoidance patterns that typically characterize this group. In contrast, drawing on VR work in healthy samples, where embodying slimmer avatars has been linked to stronger restraint and avoidance of high-calorie food21, we expected a different embodiment profile in HC: slimmer avatars were expected to reinforce avoidance of high-calorie foods in line with common weight- and shape-related concerns, whereas the effects of embodying a larger avatar were considered more exploratory, with the potential to modify these biases. The slimmer avatar condition was therefore included for both its theoretical relevance and to allow comparison across body sizes in the two groups. Finally, we explored whether implicit food-related biases were associated with body-size estimation and clinical measures of eating-disorder symptoms. Methods Participants Twenty-nine individuals diagnosed with restrictive-type Anorexia Nervosa (AN-R) were recruited between January 2021 and December 2024 from the Eating Disorder Center at the “Città della Salute e della Scienza” Hospital, University of Turin. The sample size was chosen to match that used in Dimakopoulou and colleagues6, which adopted a comparable design and analytic approach. Inclusion criteria for the AN-R group were: (a) female gender; (b) a formal diagnosis of AN-R, assessed by a trained psychiatrist using the Structured Clinical Interview for DSM-5 Disorders (SCID-5); and (c) provision of written informed consent. The study was approved by the hospital’s ethics committee (Protocol No. 0006434) and conducted in accordance with the 2013 revision of the Declaration of Helsinki. The healthy control (HC) group included thirty-one females recruited from the University of Turin’s volunteer database. Inclusion criteria for the HC group were: (a) female gender; (b) no personal or family history of eating disorders or other psychiatric conditions; (c) a body mass index (BMI) within the normative range (18.5-24.9); and (d) written informed consent, approved by the University of Turin’s ethics committee (Protocol No. 100960). All participants were native Italian speakers, aged 18-30 years, with normal or corrected-to-normal vision, and no history of substance use or neurological disorders. Although both groups met the same age inclusion criteria, natural sampling variability resulted in a significant age difference between groups (t(58) = 2.88, p = 0.006) and significant differences in educational level (t(55) = 4.92, p < 0.001). Significant differences in BMI (t(57) = 12.84, p < 0.001) were consistent with the clinical characteristics of the AN-R population. As expected, AN-R participants also showed significantly higher scores on all EDE-Q subscales and on the global score (all p < 0.001), indicating more pronounced eating-disorder psychopathology relative to HC (see Table 1). Table 1Summarizes demographic and EDE-Q characteristics of the HC and AN-R groupsVariableHC (n = 31)AN-R (n = 29)Mean difference [95% CI]tpAge (years)24.29 ± 2.9122.28 ± 2.462.01 [0.61, 3.41]2.880.006Education (years)16.26 ± 2.1113.65 ± 1.832.61 [1.55, 3.67]4.92<0.001BMI (kg/m²)20.83 ± 1.7215.10 ± 1.705.73 [4.84, 6.62]12.84<0.001EDE-Q Restraint1.02 ± 1.013.23 ± 1.72−2.21 [−2.96, −1.46]−5.95<0.001EDE-Q Eating concern0.70 ± 0.843.16 ± 1.27−2.46 [−3.03, −1.89]−8.65<0.001EDE-Q Shape concern1.73 ± 1.464.53 ± 1.07−2.80 [−3.46, −2.14]−8.47<0.001EDE-Q Weight concern1.29 ± 1.253.89 ± 1.30−2.60 [−3.27, −1.93]−7.80<0.001EDE-Q Global score1.18 ± 0.993.70 ± 1.11−2.52 [−3.07, −1.97]−9.14<0.001Values are reported as mean ± SD. Mean differences were calculated as HC-AN-R and are presented with 95% confidence intervals. Negative mean differences indicate higher scores in the AN-R group. Summarizes demographic and EDE-Q characteristics of the HC and AN-R groups Values are reported as mean ± SD. Mean differences were calculated as HC-AN-R and are presented with 95% confidence intervals. Negative mean differences indicate higher scores in the AN-R group. The Eating Disorder Inventory (EDI-2) ED-related psychopathology was evaluated with the validated Italian version of the Eating Disorder Inventory-2 (EDI-2)32. EDI-2 is a self-report measure employed to assess cognitive aspects and behaviors associated with Eating Disorders (EDs) such as AN, Bulimia Nervosa (BN), and Binge Eating Disorder33,34. It has shown good reliability and validity for assessing eating-disorder psychopathology and related psychological traits and is widely used in clinical and research settings33,34. It consists of 91 questions distributed across 11 subscales, with items rated on a 6-point Likert scale ranging from 1 (never) to 6 (always). The first three scales (drive for thinness, BN, and body dissatisfaction) index the main symptoms of EDs, and the remaining scales measure psychological aspects such as low self-esteem, perfectionism, interpersonal distrust, fear of adulthood, disturbed interoceptive awareness, ascetics, distorted impulse regulation, and social insecurity. The Eating Disorder Examination Questionnaire (EDE-Q) The Italian version of EDE-Q35 was used to assess the occurrence and frequency of the key behavioral features of EDs. EDE-Q is a 28-item self-report questionnaire derived from the Eating Disorder Examination (EDE) interview. It demonstrates strong internal consistency, test-retest reliability, and convergent and discriminant validity36,37. It includes four subscales: restraint, eating concern, shape concern, and weight concern, as well as an overall global score. It employs a 6-point Likert scale (0-6) with scores of 4 or higher, indicating the clinical range. Multidimensional Assessment of Interoceptive Awareness-2 (MAIA-2) MAIA-238 is a recently developed revised version of MAIA39, a multi-dimensional instrument designed to measure interoceptive sensibility40. It consists of thirty-seven items designed to assess eight concepts: (1) noticing: awareness of uncomfortable, comfortable, and neutral body sensations; (2) not-distracting: tendency not to ignore or distract oneself from sensations of pain or discomfort; (3) not-worrying: tendency not to worry or experience emotional distress with sensations of pain or discomfort; (4) attention regulation: ability to sustain and control attention to body sensations; (5) emotional awareness: awareness of the connection between body sensations and emotional states; (6) self-regulation: ability to regulate distress by attention to body sensations; (7) body listening: active listening to the body for insight; and (8) trusting: experience of one’s body as trustworthy and safe. Participants must rate each item on a 6-point Likert scale, ranging from never (0) to always (5). An Italian version was employed. Body Shape Questionnaire-34 (BSQ-34) The Italian BSQ-3441 is used for assessing cognitively based body dissatisfaction. It demonstrates strong test-retest reliability, good internal consistency, and convergent validity compared to other body dissatisfaction measures in clinical and nonclinical populations42. It is a 34-item measure in which participants rate the frequency of negative body-related thoughts on a 6-point Likert scale ranging from 1 (never) to 6 (always) over the last 4 weeks. State-Trait Anxiety Inventory (STAI) The STAI Italian43 is a psychological assessment tool used to measure anxiety, possibly influencing interoception sensitivity. It consists of 40 self-report items rated on a 4-point Likert scale. The inventory evaluates two types of anxiety: state anxiety (STAI-X1), which measures anxiety related to specific events, and trait anxiety (STAI-X2), which assesses the general tendency to experience anxiety. For state anxiety, participants rate their feelings of anxiety “right now” on a scale from 1 (not at all) to 4 (very much so). For trait anxiety, participants indicate the frequency of their anxiety experiences on a scale from 1 (rarely) to 4 (always). Beck Depression Inventory-II (BDI-II) The Italian version of the BDI-II44 is a 21-item inventory designed to assess depressive symptoms and their severity, classified as representing “none, mild, moderate, and severe” depression. Each item is scored on a scale from 0 to 3. It aligns with the DSM-IV criteria for depression and is widely considered one of the most effective tools for distinguishing between non-depressed, moderately depressed, and severely depressed individuals. The BDI-II is known for its strong psychometric properties, including high test-retest reliability and internal consistency45. Experimental tasks Stimuli were obtained from the Food-pics Database46. We selected six high-calorie (pistachio nuts, donuts, chocolate biscuits, hamburger, pizza, lasagna) and six low-calorie (zucchini, tomato, mushrooms, green beans, mixed steamed vegetables, apple) items, which were standardized to a circular frame. Stimuli were then superimposed onto an image of a plate contour to ensure symmetrical binocular fusion. Although the Food-pics database provides standardized images, we additionally extracted low-level visual features (brightness, contrast, and spatial frequency) and compared them between high- and low-calorie categories using one-way ANOVAs. No significant differences emerged for these visual features (all ps > 0.05), indicating that the two stimulus categories were comparable on low-level visual properties. The experiment was created using MATLAB (Release, 2021b) and Psychtoolbox. Stimuli were presented on a BenQ Monitor (1920 × 1080 pixels resolution, 120 Hz, 24”) positioned 57 cm in front of the participants. Participants’ heads were placed on a homemade chinrest connected to a stereoscope, adjusted to maintain stable binocular vision. Binocular rivalry (BR) BR measures perceptual dominance when different stimuli are presented to each eye and compete for conscious awareness. Perception typically alternates between the stimuli because the visual system cannot simultaneously fuse two dissimilar images into a single coherent percept. The rationale is that stimuli reported for a longer dominance time are prioritized within perceptual awareness. In each trial, a high-calorie stimulus was presented to one eye, whereas a low-calorie stimulus was presented to the other (counterbalancing eye-calorie pairs). Stimuli were displayed within squared frames (7.2° × 7.2°) on a black background for 60 s, made of white and black pixels (width 0.2°), positioned 5° from the center (one on the left, one on the right), with a black central fixation cross. Participants were instructed to fixate on the cross, avoiding eye blinking or closing one eye at a time, and to continuously report the dominant stimulus throughout the trial. After four practice trials, the experiment began. Participants pressed the corresponding arrow key to indicate their perception each time it changed: the right arrow if they perceived more than 50% of the low-calorie stimulus (low-calorie predominance), the left arrow for more than 50% of the high-calorie stimulus (high-calorie predominance), or the up arrow if no stimulus was dominant. The trial continued automatically for 60 s without interruption if no key was pressed. Periods with no reported dominant percept (i.e., no response) were excluded from the computation of perceptual dominance durations. Each condition included 24 randomized trials (12 with each stimulus presented to the left and 12 to the right eye). A 10-s break followed each trial, and a 60-s eye rest was provided after 12 trials (see Fig. 1). Schematic representation of a Binocular Rivalry (BR) trial. Participants viewed a high-calorie stimulus with one eye and a low-calorie stimulus with the other eye for 60 s. They were instructed to maintain central fixation and continuously report their dominant percept by pressing the corresponding arrow key: left arrow for high-calorie predominance, right arrow for low-calorie predominance, and up arrow for no clear predominance. After each trial, participants had a 10-s break, with a 60-s eye-rest break after every 12 trials. Food preferences approach-avoidance test (FP-AAT) We utilized the FP-AAT to assess the implicit attitudes toward different food categories by quantifying participants’ attraction or repulsion to specific stimuli. The stimuli included eight images of hypercaloric foods (e.g., pizza, French fries, chips, chocolate), eight images of hypocaloric foods (e.g., fruits, white meat), and eight neutral stimuli (numerical digits such as 7035 and 7004). Each stimulus category was presented within a square or round frame, with the frame shape always serving as the target. Before each trial, a red “X” fixation point appeared at the center of the screen. Participants clicked it to initiate the trial and were instructed to move the mouse toward or away from themselves, depending on the frame shape (circle or square). The underlying assumption was that the semantic properties of the task-irrelevant stimulus (e.g., calorie content) carried a specific valence (avoidance or attraction), potentially influencing response times depending on whether the response action (pushing/pulling) aligned or conflicted with the stimulus category. For instance, participants who preferred hypercaloric foods were expected to pull the mouse faster than push it away. Participants were instructed to move the mouse cursor to the upper or lower edge of the screen (i.e., to “touch” the top or bottom screen border). As they performed the movement, the stimulus progressively expanded (pull) or shrank (push) to enhance the visual feedback associated with the motor action. During the practice session, if an incorrect response was given, the word “error” appeared at the center of the screen. Participants were instructed to respond exclusively to the shape of the frame (circle vs square) by pulling or pushing the mouse, while ignoring the food content of the images. RT (ms) were measured from stimulus onset to the time the mouse cursor reached the upper or lower edge of the screen. Trials with incorrect responses or RTs <200 ms or >3000 ms were discarded. For each participant, condition, and calorie category (high vs low), mean RTs were computed separately for push and pull movements. To quantify approach-avoidance tendencies, we calculated a D-score by subtracting the mean push RT from the mean pull RT and dividing the difference by the pooled standard deviation across push and pull trials. Positive D-scores, therefore, indicate relatively faster pushing (greater avoidance) of that food category. The experiment consisted of 20 practice trials and 80 experimental trials. Each of the 24 stimuli (8 hypercaloric, 8 hypocaloric, 8 neutral) was presented twice, once in a square frame and once in a circle frame, for a total of 48 trials. Additional stimulus repetitions were included to reach 80 trials, with approximately equal distribution across the hypercaloric, hypocaloric, and neutral categories (see Fig. 2). Schematic representation of a food preferences approach-avoidance task (FP-AAT) trial. At the start of each trial, a red “X” appeared at the center of the screen, which participants clicked to initiate the trial. A stimulus image (hypercaloric food, hypocaloric food, or neutral digit) appeared within a square or a circular frame. Participants were instructed to pull the mouse toward themselves or push it away based on the frame shape (circle or square), ignoring the image’s content. As participants moved the mouse, the stimulus progressively expanded (pull) or contracted (push) to reinforce the perception of movement. Each trial ended once the movement was completed, with reaction time recorded from the onset of the trial. Procedure At admission, participants completed a battery of clinical assessments, including the EDI-232, EDE-Q35, MAIA-238, BSQ-3441, STAI43, and BDI-II44. The experimental procedure comprised three sessions (approximately 1.5 h each), structured in a mixed-subjects design, and spaced 1 week apart. During the first session, participants provided written informed consent and demographic information. Anthropometric measurements (height, weight, bust, waist, hip, thigh, and foot size) were recorded to calculate BMI and to create individualized avatars for the VR embodiment. Avatar height was adjusted to match each participant’s measured height, while body width dimensions were subsequently manipulated by ±15% for the Slim and Large conditions. Foot size was used as a control measure in subsequent tasks. Following these preliminary procedures, participants completed the BR and FP-AAT tasks to assess the visual processing and implicit attitudes toward food. This session constituted the Baseline condition. In the second and third sessions, all participants underwent the VR-FBI in which they embodied two avatars of different sizes: one slimmer (Slim condition) and one larger (Large condition) than their actual body. The order of these conditions was randomized and counterbalanced across participants, and sessions were spaced 1 week apart to minimize potential practice and habituation effects. The Body Size Estimation Task (BSET) was administered before the first FBI exposure (Slim/Large) and again at the end of each FBI session to quantify participants’ subjective perceptions of their body size. Following each embodiment, the Body Ownership Questionnaire (BOQ) was administered to evaluate the strength of the illusion. At the end of each session, the BR and FP-AAT were administered in a counterbalanced order (see Fig. 3). Graphical overview of the experimental design and procedure. (1) Timeline of the three sessions: Baseline (Session 1: consent, anthropometric measurements, BR, FP-AAT), FBI Session 2 (Slim or Large avatar: BSET before and after FBI, BOQ, BR, FP-AAT), and FBI Session 3 (other avatar size: BSET after FBI, BOQ, BR, FP-AAT). (2) Example screenshots of the full-body illusion (FBI) with Slim and Large avatars. (3) Schematic trial structure of the binocular rivalry (BR) and Food Preferences Approach-Avoidance Test (FP-AAT) tasks. Participants were immersed in the virtual environment using a head-mounted display (Oculus Rift S, equipped with two PenTile OLED displays with a 1080 × 1200 pixels resolution, a 90 Hz refresh rate, a 110-degree field of view, and 6 degrees of freedom). The scenario was created and implemented in Unity3D (http://unity.com). It included a room with a chair, a table, a window, and a mirror, resembling the actual room where the study took place. Once the scene was loaded, a dressed female avatar was placed standing in front of a mirror. The avatars were created individually for each participant using MakeHuman software (www.makehumancommunity.org), and their body size was manipulated by adding 15% (Large condition) or subtracting 15% (Slim condition) from the participant’s actual body size. At the beginning of the FBI procedure, the participants had to put on the HMD, and they could only see a blue background surrounding them. During this phase, the pre-embodiment BSET was administered. Once the initial BSET was completed, the virtual scenario was activated, and one of the avatars (slim or large, depending on the condition) became visible. At this point, participants were asked whether the virtual body overlapped with their own body, and, eventually, make any necessary positional adjustments until they perceived complete alignment from a 1PP. Once the alignment was confirmed, the embodiment phase began. Participants were asked to move their heads to look around and towards themselves (i.e., avatars) for 30 s to familiarize themselves with the virtual environment and body. To induce the FBI, a synchronous visuo-tactile stimulation procedure was followed: a controlled vibrotactile stimulator called Arduino (diameter × height = 8 mm × 3 mm, 3 V, rated speed – 12,000 RPM; www.arduino.cc) placed on the participants’ abdomen delivered tactile pulses. Simultaneously, participants observed a green ball rolling toward and away from the avatar’s abdomen, with each tactile pulse synchronized to the ball’s virtual contact (5 Hz, six repetitions). During synchronous stimulation, participants were instructed to constantly look down at their body, now replaced by the virtual body, or to alternate their gaze between the mirror in front of them (reflecting the avatar) and the green ball. After 2 min, the ball stopped moving and remained suspended in midair. The virtual scene then faded to the original blue background, and participants completed the BSET and BOQ. Finally, they removed the HMD and continued with the remaining tasks. Body ownership questionnaire (BOQ) The BOQ was adapted from questionnaires used in previous studies (Lenggenhager et al.47; Tambone et al.21) and included 10 questions (5 illusion and 5 control) used to assess various aspects of the subjective experience of the FBI. The questions were administered in a randomized order and were related to the location (Q1), ownership (Q2), ownership through the mirror-that is, the identification with the avatar in the mirror (Q3), appearance (Q4), and touch coming from the stimulation (Q5). The answers were rated on a Likert scale varying from −3 (total disagreement) to +3 (total agreement), with 0 indicating “I don’t know” (see Table 2) Table 2Body Ownership Questionnaire (BOQ): Q1-5 illusion and control questionsTYPENo.IllusionControlLocation1At times, it felt as if my body was located where the virtual body was.At times, it felt as if I was outside of my body.Ownership2At times, it felt like the virtual body I saw when looking down was my body.At times, it felt like the virtual body belonged to someone else.Ownership (Mirror)3At times, it felt like the virtual body I saw reflected in the mirror was mine.At times, it felt as if I had two bodies.Appearance4At times, it felt like the virtual body resembled my own (e.g., shape, skin color, or other characteristics).At times, it felt as if my body had disappeared.Touch5At times, it felt as if the touch I perceived was in the exact location where the virtual body was touched.At times, it felt as if my body was being touched in a different place than where the virtual body was touched.Participants rated their level of agreement on a Likert scale ranging from -3 (total disagreement) to +3 (total agreement), with 0 indicating “I don’t know”. Body Ownership Questionnaire (BOQ): Q1-5 illusion and control questions Participants rated their level of agreement on a Likert scale ranging from -3 (total disagreement) to +3 (total agreement), with 0 indicating “I don’t know”. Body size estimation task (BSET) The BSET was used to measure the subjective body size estimation before and after the FBI21. This task was intended to assess subjective body representation and perception, which are thought to reflect predominantly top-down, cognitive-affective processes rather than unconscious perceptual mechanisms. While participants were standing and wearing the HMD, they were asked to extend their arms along their bodies (without touching them). Afterwards, they were instructed to flex their arms in front of them and to adapt the distance between their palms to the perceived size of their hips. We also collected measurements of the foot size as a control. A measuring tape was used to measure the distance between the palms, and the procedure was repeated three times. For each body part and estimation type (i.e., pre and post), we calculated the percentage of estimation using the formula: percentage of estimation = [(estimated size - actual size)/actual size] × 100. A negative value indicated an underestimation, while a positive value indicated an overestimation. This measure is mathematically equivalent to the commonly used Body Perception Index (estimated/actual × 100) but is centered around zero to facilitate the interpretation of over- and underestimation. Statistical analysis To assess the efficacy of the FBI, we performed nonparametric tests on questionnaire data (e.g., BOQ), specifically, Wilcoxon signed-rank tests were used for within-group comparisons, and Mann–Whitney U tests for between-group comparisons. Regarding the BSET, we performed a repeated-measure ANOVA with Body Part (e.g., hips, foot), Condition (Baseline, Slim, Large), and Group (AN-R, HC) as factors. Concerning the experimental tasks, D-scores (FP-AAT) or reaction times (BR) were entered into repeated-measures ANOVAs with Calorie (High, Low) and Condition (Baseline, Slim, Large) as within-subject factors and Group (AN-R, HC) as a between-subject factor. All ANOVA models were specified as full factorial models including all interaction terms. When the highest-order interaction was not significant, we checked whether the relevant lower-order interactions remained stable after removing the non-significant highest-order interaction. Spearman correlations were performed between the dependent variables of the two experimental tasks and the clinical assessment questionnaires. These analyzes were exploratory and did not allow causal or mechanistic inferences. P-values were corrected for multiple comparisons using the false discovery rate (FDR; Benjamini-Hochberg procedure). Statistical analyzes were conducted using Statistica 14.0.1.25 (TIBCO) with an alpha level set at 0.05. Age and education differed between groups. Although they were not included as covariates in the primary models because all participants were within a restricted early-adulthood age range, education was partly confounded with age and clinical trajectory, and the sample size was limited for additional covariate-adjusted models. Results Full body illusion Body ownership questionnaire (BOQ) In the within-group analyzes, the Wilcoxon matched-pairs tests showed that the ratings for Q1 (perceived self-location), Q4 (perceived resemblance between the virtual and the real body), and Q5 (visuotactile congruence between seen and felt touch) were significantly higher (p < 0.05) for the Illusion than for the corresponding Control questions, regardless of group or condition. For the between-groups analysis, the Mann-Whitney U tests revealed that ratings for each illusion response did not differ between the two groups (p > 0.05). These results indicate that three (out of five) illusion statements were significantly higher than the corresponding control questions, regardless of group or condition (see Fig. 4). Mean scores (±Standard Error; SE) for ILLUSION and CONTROL responses, collapsed across conditions, for the three illusion questions indexing perceived self-location (Q1), perceived resemblance between the virtual and the real body (Q4), and visuotactile congruence (Q5). Bars represent the mean rating. Scores ranged from −3 to +3. Across all questions, participants reported significantly stronger experiences (higher ratings) in the ILLUSION condition than in the CONTROL condition, with the largest difference observed for Q5. BOQ responses were collected immediately after each full-body illusion session and collapsed across embodiment conditions, as this figure is intended to demonstrate successful induction of the full-body illusion rather than condition-specific effects of avatar size. Error bars represent the standard error of the Mean (SE). Asterisks * indicate significant differences between ILLUSION and CONTROL responses (p < 0.05). Body size estimation task We conducted a repeated-measures ANOVA with Condition (Baseline, Large, Slim), Group (AN-R, HC), and Body Parts (Hips, Foot) as independent variables and the percentage estimation error as the dependent variable. The analysis revealed a significant main effect of group (F(1, 57) = 10.367, p = 0.002, ηp² = 0.01), as well as two significant interactions, namely Group x Body (F(1, 57) = 40.43, p < 0.001) and Condition × Body F(2, 114) = 6.39, p = 0.002). Concerning the Condition x Body interaction, post-hoc indicated that at baseline, the Hips had a significantly (p = .001) higher estimation percentage than the Foot. Crucially, post-hoc tests for the Group x Body interaction, post-hoc tests indicated that in the AN-R group, the estimation percentage for the Hips (Mean = 5.14, SE = 2.65) was significantly (p < .05) higher than for the Foot (Mean = −1.59, SE = 1.15), whereas in the HC group, the estimation percentage for the Foot (Mean = −0.35, SE = 1.38) was significantly (p < 0.01) higher than for the Hips (Mean = −5.53, SE = 2.52). Overall, these results indicate that, regardless of the condition, the AN-R group overestimated the hips relative to the feet. In contrast, the HC group overestimated the foot compared to the hips (see Fig. 5). Dot plot showing body size estimation (% overestimation) for hips and feet in AN-R and HC groups. Each dot represents an individual participant. Group means (black circles) and standard errors (vertical lines) are shown. Within-group comparisons revealed significant differences between hips and feet in both groups: AN-R (p < 0.001; ***) and HC (p < 0.01; **). Dot colors indicate body part (pink = hips, blue = foot). Condition effects are not depicted, as no significant Condition × Group interaction emerged for body size estimation. Binocular rivalry We performed a repeated-measures ANOVA with Condition (Baseline, Large, Slim), Calorie (High, Low), and Group (HC, AN-R) as independent variables and the reaction time (RT) as the dependent variable. The analysis revealed a significant main effect of Calorie (F(1, 40) = 6.58, p = 0.014, ηp² = 0.14) and Condition (F(2, 80) = 18.52, p < 0.001. ηp² = 0.32), and a Group × Calorie interaction (F(1, 40) = 4.56, p = 0.038). Crucially, Post hoc analyses of the significant Group × Calorie interaction revealed that in the AN-R group, reaction times were longer for high- than low-calorie food (13.75 ± 1.45 vs. 10.29 ± 1.44; p < 0.05). Reaction times were also overall longer in the AN-R group than in the HC group (p < 0.001). These results indicate that the AN-R group showed greater perceptual dominance for high-calorie foods across conditions. However, this effect should be interpreted with caution, as it did not reach conventional statistical significance when the non-significant higher-order interaction was removed from the model (p = .065). In contrast, the HC group showed no differential dominance (see Fig. 6). Line plot showing BR reaction times (RTs) for high- and low-calorie food stimuli across AN-R and HC groups. Colored lines represent group means, with error bars indicating standard errors (SE). Individual dots represent participant-level data. Asterisks (p < 0.05; *) indicate significant differences (p < 0.001; ***). Food preferences approach-avoidance test A D-score was calculated by subtracting the mean push RT from the mean pull RT for each stimulus group, then dividing the difference by the pooled Standard Deviation (SD). A positive score indicates a preference for pushing (i.e., avoidance tendencies). We ran a repeated-measures ANOVA with Condition (Baseline, Large, Slim), Calorie (High-calorie, Low-calorie), and Group (HC, AN-R) as independent variables and the D-Score as the dependent variable. The analysis revealed a significant main effect of Condition (F(2, 112) = 4.35, p = 0.015, ηp²= 0.07), and of Condition × Calorie × Group (F(2, 112) = 6.83, p = 0.001) interaction. For the interaction, post-hoc test showed that in the AN-R group, D-Score for high-calorie food was significantly (p < 0.05) lower in the Large (Mean = −0.27, SE = 0.31) than in the other two conditions (Baseline Mean = 0.52, SE = 0.29; Slim Mean = 0.45, SE = 0.31), whereas the D-Score for Low-calorie food was significantly (p = 0.026) higher in the Slim condition (Mean = 0.49, SE = 0.27) than in the other two conditions (Large Mean = −0.33, SE = 0.30; Baseline Mean = −0.43, SE = 0.29). In the HC group, the D-score for Low-calorie food was significantly (p < 0.05) lower in the Large (Mean = −0.38, SE = 0.28) than in the other two conditions (Baseline Mean = 0.49, SE = 0.27; Slim Mean = 0.35, SE = 0.25), whereas the D-Score for high-calorie food was significantly (p = 0.015) lower in the Baseline (Mean = −0.81, SE = 0.27) than in the other two conditions (Large Mean = −0.07, SE = 0.28; Slim Mean = 0.17, SE = 0.29). Spearman correlations showed that in the Slim condition, the D-score of High-calorie food in the AN-R group was negatively correlated with a variety subscales of the clinical assessment, namely the EDI-2 (perfectionism scale r = −0.54, p = 0.01; asceticism r = −0.52, p = 0.016), and the BDI-II (cognitive r = −0.58, p = 0.006; somatic r = −0.56, p = .008, total score r = −0.59, p = 0.005). In the same condition, the D-score of Low-calorie food of the HC group was positively correlated with the BDI-II (cognitive r = 0.57, p = 0.002; somatic r = 0.72, p < 0.001, total score r = 0.66, p < 0.001), the EDI-2 interoceptive awareness subscale (r = 0.48, p = 0.008), and body listening (r = 0.54, p = 0.004) as measured by the MAIA-2. These results indicate that in the Baseline, the AN-R group showed a clear behavioral pattern of avoiding High-calorie foods and approaching Low-calorie foods. In contrast, the HC group displayed the opposite pattern: participants tended to approach High-calorie foods and avoid Low-calorie foods. In the Large condition, both groups shifted their approach tendencies: participants moved toward the food types they typically avoided at Baseline. However, in the Slim condition, an avoidance pattern emerged across all food categories (Fig. 7 about here). Mean FP-AAT D-scores ( ± SE) across experimental conditions (Baseline, Large, Slim) for high- and low-calorie food stimuli in AN-R (left) and HC (right) groups. Positive scores reflect avoidance behavior. Orange lines indicate high-calorie stimuli and blue lines indicate low-calorie stimuli. Asterisks (*) and black bars denote significant within-group differences (p < 0.05). This figure illustrates the significant Condition × Calorie × Group interaction observed in the FP-AAT. Discussion The present study investigated how illusory body ownership, induced through virtual embodiment of slimmer or larger avatars, affects food-related automatic and perceptual processing as well as body-size representation in women with AN-R and HC. We combined a FBI paradigm with a body size estimation task, a binocular rivalry task indexing visual awareness of high- and low-calorie foods, and a food-specific approach-avoidance task. We related these measures to eating-disorder symptoms and other clinical characteristics, and three main findings emerged. First, individuals with AN-R consistently showed greater perceptual dominance for high-calorie foods than HC, regardless of the embodiment condition. Second, embodying a larger avatar transiently reduced avoidance of high-calorie foods in AN-R on the approach-avoidance task, indicating that rigid food-related biases can be modulated in the short term. Third, healthy participants tended to show the opposite pattern, shifting toward greater restraint for high-calorie foods after embodying a larger body. These results suggest that changes in embodied self-representation can influence implicit behavioral tendencies toward food, although the broader clinical implications of such effects require careful evaluation. Regarding the illusion itself, both AN-R and healthy participants reported a strong sense of ownership over their avatars, with no differences between groups. This was reflected in the experimental questions, which assessed participants’ subjective experience of embodiment by focusing on their identification with the avatar in terms of location, appearance, and tactile feedback. These findings suggest that the basic multisensory integration processes supporting the sense of body ownership remain intact in individuals with AN-R, aligning with previous research showing that body illusions are not necessarily disrupted in eating disorders24–26. However, contrary to our initial hypothesis and some prior findings17,19–21,24,25,48, the experimental manipulation in either group did not modulate the perception of one’s body size. AN-R participants persistently overestimated the size of their hips relative to their feet, while healthy controls tended to underestimate. This aligns with prior evidence49–52 indicating that body-size overestimation in AN is highly resistant to change, even after exposure to a healthier-weight avatar. One explanation may lie in the “expertise effect”53: individuals with AN, who are highly attuned to low BMI ranges, show greater perceptual uncertainty when evaluating larger bodies, thereby reinforcing overestimation. Methodological differences may also explain discrepancies with prior work. Unlike earlier studies that used pre-defined avatars or manipulated only one body region19,21,24, we personalized avatars by adjusting multiple body dimensions (e.g., waist, hips, thighs) by ±15% of each participant’s real measurements. While this likely enhanced realism and body ownership, it may have also reinforced participants’ existing body image, limiting recalibration, particularly in AN-R, where body schema distortions are deeply entrenched. These findings support the notion that body misperception in AN is shaped by enduring cognitive-affective mechanisms and is less amenable to short-term perceptual recalibration. However, repeated and emotionally targeted embodiment-based interventions may still offer therapeutic potential26. Regarding implicit visual food processing, AN-R participants consistently demonstrated a perceptual dominance for high-calorie foods in the BR task, replicating prior findings6 and aligning with evidence that emotionally salient stimuli dominate visual awareness54–56. This dominance likely reflects heightened affective responses to food in AN57,58, which persisted regardless of the avatar embodiment. The lack of modulation by the FBI may stem from the task’s sensitivity to early, perceptual processing, which may be less responsive to transient changes in body ownership, and from the absence of explicit emotional reappraisal often targeted by VR-based exposure therapies26. Thus, perceptual salience reflects more stable emotional biases that require deeper therapeutic intervention to modify. The other main result regarded the automatic preference for food stimuli, as measured by the FP-AAT. Here, behavioral patterns sharply diverged between groups: at baseline, AN-R participants avoided high-calorie foods and approached low-calorie foods, while HC participants exhibited the opposite pattern. These observations in the HC group should be interpreted with caution, as they were not the primary focus of the study and remain inferential in nature. Beyond replicating previous findings, that high-calorie foods in AN function as emotionally aversive stimuli, triggering automatic avoidance responses6, the embodiment of a Large avatar temporarily reduced repulsion of the initially avoided food category, suggesting that experiencing ownership over a larger body can momentarily modulate maladaptive behavioral biases. This supports the idea that the FBI elicits behavioral responses aligned with stereotyped expectations linked to the embodied body size22, even in clinical populations. However, this modulation was not maintained after exposure to a Slim avatar, where both groups exhibited generalized food avoidance. While our primary hypotheses concerned the effects of the Large condition, the pattern observed in the Slim condition should be considered exploratory. These findings point to the influence of internalized body ideals where embodying a Slim avatar, likely congruent with AN-R participants’ ideal self-image, may have reactivated restrictive eating schemas, reinstating baseline avoidance. In contrast, Large avatars may have momentarily disrupted these schemas. Notably, avatar size may have produced differential perceptual effects across groups. For AN-R individuals, a 15% size increase may still result in a relatively thin avatar, whereas for HC participants near the upper end of the BMI range, it may appear overweight, potentially eliciting restraint due to normative weight concerns. This may explain why, in HC, Large embodiment led to increased avoidance of high-calorie food. Our findings partially contrast with Tambone and coworkers21, who reported no effect of large avatars on food behavior, and align more with Verhulst and colleagues23, who found that embodying an obese avatar did not promote unhealthy eating. Notably, recent evidence suggests that embodiment of a larger body can increase implicit approach tendencies toward high-calorie food59. However, our results indicate that such effects are not unidirectional, as in our HC sample, the same manipulation was associated with increased avoidance. Collectively, these results challenge assumptions that larger body representations drive impulsive food intake, and instead suggest that they may evoke compensatory restraint, potentially linked to internalized social norms or fear of weight stigma60. Rather than promoting impulsive/hedonistic eating, the illusory experience may trigger compensatory restraint strategies to maintain conformity to societal ideals around thinness and health. Taken together, the dissociation between perceptual dominance in the BR task and behavioral avoidance in the FP-AAT suggests that food-related biases in AN may operate at different levels of processing. While BR reflects early perceptual salience, FP-AAT captures later motivational and action-related tendencies. This pattern is consistent with frameworks that distinguish between the perceptual and behavioral components of cognitive-emotional processing. Correlation analyzes further linked these behavioral patterns to symptom severity. In the Slim condition, lower avoidance of high-calorie foods in AN-R was associated with higher perfectionism and depressive symptom severity. In HC, stronger avoidance of low-calorie foods was similarly associated with depressive symptoms, interoceptive awareness, and body listening. These findings indicate that automatic biases toward food are not only sensitive to embodied self-perception but are also systematically related to the severity of eating disorder symptoms and emotional dysregulation in both clinical and non-clinical populations. Overall, while embodiment can transiently modulate behavioral tendencies toward food in AN-R, deeply entrenched emotional and body-related factors continue to sustain rigid avoidance behaviors. Before concluding, some limitations must be acknowledged. First, although the FBI successfully induced ownership, it may not have been strong enough to provoke deeper emotional or behavioral restructuring. Future interventions should combine embodiment with explicit emotional retraining to better target cognitive-affective biases. Second, the ±15% body size manipulation might have been too subtle to disrupt entrenched body schemas; larger or more ecologically valid modifications should be explored. Third, while a separate baseline provides an uncontaminated reference condition, it did not fully control for possible fluctuations in participants’ state across sessions; adding a baseline assessment immediately before each embodiment session would have improved within-session control but would also have increased session duration and participant burden, particularly for AN-R patients. Fourth, the relatively short intervention period may have limited transient effects, suggesting that longer, repeated exposure sessions could yield stronger outcomes. Fifth, while the BR and FP-AAT paradigms captured perceptual and automatic processing, additional tasks targeting higher-order regulation and body schema updating are needed to fully assess embodiment effects. Sixth, the sample was relatively small and included only females with AN-R recruited from a specialized eating-disorder center, limiting generalizability. Although this makes the sample fairly typical of treatment-seeking restrictive AN, here the two groups also differed in age and education. Because all participants were within a restricted early-adulthood age range, these differences are unlikely to fully account for the observed effects; however, they may have influenced reaction times, attentional flexibility, or task performance and should therefore be considered when interpreting the findings. Future studies should replicate findings in larger samples more closely matched for age and education, as well as gender-diverse populations, individuals with anorexia binge-purging subtype, and weight-restored or non-treatment-seeking individuals to assess gender-related and subtype-specific differences in embodiment and food processing. Finally, we did not measure individual liking/healthiness ratings for the food stimuli, so the caloric category may not fully capture subjective value61. Additionally, external factors such as meal context, hunger/satiety, emotional state, and social environment may influence automatic food responses and warrant further investigation. In conclusion, by combining repeated embodiment sessions with both implicit and explicit measures of food-related processing and body image perception, this study suggests that altering body ownership through immersive VR can transiently modulate implicit mechanisms underlying restrictive eating in individuals with AN-R. Importantly, the observed effects were short-term and should not be interpreted as evidence of direct clinical efficacy. These findings challenge the long-held view that disordered eating patterns in AN are impervious to short-term intervention, revealing instead that even deeply entrenched behavioral and perceptual biases can be temporarily reshaped through full-body illusions. Notably, these findings provide novel evidence that eating-related responses in AN-R may not be entirely fixed, but can show some degree of short-term flexibility, which may inform future interventions. Future research should capitalize on this insight by including embodiment-based interventions to access and recalibrate the unconscious, sensory-affective mechanisms that sustain the disorder. Using size-specific avatars may offer an effective, targeted method to disrupt pathological body ideals and foster more adaptive responses to food, potentially marking a paradigm shift in the treatment of AN and consequently other eating disorders. These ideas are consistent with recent VR body-exposure protocols in AN, where repeated, carefully framed embodiment of larger or healthy-weight avatars is used as an exposure-based intervention and has been shown to reduce fear of weight gain and body-related anxiety26,27. At the same time, the emotional sensitivity of embodying larger or smaller bodies in AN means that any clinical application of similar procedures requires gradual implementation, close monitoring, and integration into psychotherapy. Given AN’s high mortality and treatment resistance62–64, these findings open promising therapeutic avenues. Acknowledgements This work was supported by a PRIN 2022 grant awarded to L.P. by the Italian Ministry of University and Research (MUR) (Project No. 2022REJEB4, SENSIBLE—‘SENSing Inner BodiLy state: understanding the role of interoception in obesity’), funded by the European Union – NextGenerationEU, Mission 4, Component 2, Investment 1.1 (CUP D53D23009830006). Author contributions M.D. designed the study, conducted the experiments, analyzed the results, and wrote the manuscript. G.A. designed the study, and L.P. designed the study, contributed to statistical analysis, and data interpretation. All authors reviewed and approved the final manuscript. Data availability The datasets generated and/or analyzed during the current study are not publicly available due to the sensitive and potentially identifiable nature of the qualitative data collected, but are available from the corresponding author on reasonable request. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets generated and/or analyzed during the current study are not publicly available due to the sensitive and potentially identifiable nature of the qualitative data collected, but are available from the corresponding author on reasonable request.