REACT 2026: The Fourth Multiple Appropriate Facial Reaction Generation Challenge: Personalised MAFRG and Appropriate EEG Reaction Prediction

Organizers have opened the REACT 2026 challenge, the fourth iteration of a competition focused on building machine learning models that generate personalized facial reactions to a speaker’s behavior during one-on-one conversations [1]. The challenge targets the problem of multiple appropriate facial reaction generation, or MAFRG. In dyadic interactions, a range of facial responses can be suitable for any given speaker behavior, and the REACT series has driven the development of generative deep learning models to address this variability since its 2023 launch [1][2]. This year’s edition pushes further into personalization by supplying participants with individual-level Big-Five personality labels and electroencephalography recordings alongside the existing MARS dataset [1][3]. The MARS dataset was first introduced for REACT 2025 and is described as the first natural, large-scale audio-visual corpus built specifically for MAFRG tasks. It captures 137 human-human dyadic interactions across 3,105 sessions covering five distinct topics [5][6]. For REACT 2026, the addition of personality and neurophysiological data creates what organizers call a one-to-many personalized reaction generation setting that combines behavioral, affective, and brain signals — an area they note remains largely unexplored in dyadic interaction modeling [1][6]. The competition is structured around four sub-challenges: offline generic MAFRG, offline personalized MAFRG, online generic MAFRG, and online personalized MAFRG [1][2]. Baseline models and evaluation guidelines have been published on GitHub [1]. Beyond facial reaction generation, the challenge also encourages participants to predict appropriate responsive brain activity, or EEG signals, from audio-visual speaker behaviors. Early baseline results provided what organizers describe as the first evidence that a listener’s brain activity can be partially predicted from a speaker’s expressive behavior alone, though the associations remained weak [3][4]. Submissions will be benchmarked on the appropriateness, diversity, realism, and synchronization of the generated facial reactions [5][6]. The challenge paper was submitted to arXiv on June 6, 2026 [1].