Abstract
BACKGROUND: Despite significant advances in understanding facial emotion recognition (FER) in autistic adults in recent decades, the mechanisms underlying FER difficulties in individuals with autism remain unclear, with inconsistent findings across studies. A key limitation may be the reliance on aggregate accuracy scores, which overlook item- and subject-level variability. Here, we investigated the effects of task adaptation and stimulus properties on FER performance in autistic and non-autistic adults using mixed-effects modelling. METHODS: A total of 120 autistic and 116 non-autistic participants completed the Berlin Emotion Recognition Test 2. Performance was analyzed on a trial-by-trial basis, considering trial number, stimulus properties—derived from automated facial analysis—and their interactions with diagnostic group. Response times were analyzed using mixed-effects linear regression models, while accuracy was analyzed using mixed-effects logistic regression models. RESULTS: Compared with non-autistic participants, autistic participants demonstrated lower overall accuracy and slower responses, accompanied by significantly reduced task adaptation. Contextual ambiguity of stimulus faces moderated group differences in FER accuracy, with non-autistic subjects showing greater use of contextual information. Social-cognitive traits further moderated the effect of contextual ambiguity in autistic subjects. LIMITATIONS: Our findings are specific to the design and stimulus material of the Berlin Emotion Recognition Test 2 and may not generalize to other FER tasks. Furthermore, our sample did not include individuals with intellectual disabilities, limiting generalizability across the autism spectrum. Lastly, the reliability of stimulus property estimates derived from automated facial analysis may require validation on a larger sample of stimulus faces. CONCLUSIONS: Our findings reveal that differences in task adaptation and contextual cue processing underlie FER performance differences in individuals with autism, emphasizing the importance of participant- and item-level analysis. These results may inform future study designs and highlight the advantages of integrating automated FER with mixed-effects modeling in autism research. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s13229-026-00711-6.