# Dynamic Action-Space Adaptation via Contrastive Transition Adaptation in Active Inference Presenter: Tran Hoang Linh Contribution: Pre-recorded Video presentation Public ID: talk-d0717875c7768add ## Submitted abstract Active Inference commonly assumes a fixed, pre-specified action space and sufficiently informative preferences for planning. In realistic settings, however, an agent may encounter previously unavailable actions while receiving only sparse goal information and operating with a short planning horizon. Existing Active Inference formulations generally assume transition models for a predefined action set, leaving the initialization of newly introduced actions without direct transition evidence underexplored. Moreover, while Dirichlet transition learning provides a principled mechanism for accumulating observed evidence, it primarily reinforces supported transitions and does not explicitly exploit counter-evidence indicating that candidate transitions are inconsistent with observed outcomes. We propose Dynamic Action-Space Adaptation (DASA), a framework that enables Active Inference agents to incorporate new actions during interaction, initialize each new action’s transition model using the average transition slice of existing actions, and progressively refine it through Contrastive Transition Adaptation (CTA) as evidence accumulates. At the core of DASA, CTA combines complementary positive and negative updates for newly introduced actions. Positive updates adapt the reinforcement magnitude according to the current transition belief, assigning stronger updates to low-probability transitions supported by evidence while limiting excessive reinforcement of already confident transitions. Negative updates explicitly reduce support for candidate transitions contradicted by observed evidence, enabling the transition model to adapt in both supportive and corrective directions. Across eight random seeds and two target directions, CTA achieved lower MAE than Dirichlet transition learning in all runs, reducing the mean error from 0.1581 to 0.1370 (13.4%), while also achieving a slightly lower mean RMSE (0.3006 vs 0.3030). This indicates more consistent overall recovery of the ground-truth transition dynamics, despite seed-dependent variation in larger localized errors. Overall, DASA provides a practical mechanism for incorporating and adapting newly introduced actions in Active Inference, allowing their transition models to become progressively better aligned with the underlying environment dynamics. Source: [submitted source]() Abstract SHA256: 80fed90d44032b793ea00070fa7f128684abf0ea38547d46bccbb86594735991 ## Organizer review checklist - [ ] Confirm the submitted title, presenter, and abstract with the presenter. - [ ] Confirm contribution format and scheduling needs. - [ ] Confirm public permissions for any additional material before publication. - [ ] Prepare discussion questions grounded in the submitted abstract. ## Public learning resources These approved event resources are starting points; topic relevance and author identity have not been inferred. - Active Inference Ontology: https://github.com/ActiveInferenceInstitute/Active_Inference_Ontology - Cognitive Agent Modeling: https://github.com/ActiveInferenceInstitute/ActiveInferenceImplementations - GEO-INFER: https://github.com/ActiveInferenceInstitute/GEO_INFER - Generalized Notation Notation: https://github.com/ActiveInferenceInstitute/Generalized_Notation_Notation - pymdp: https://github.com/infer-actively/pymdp - Reproducible publication template: https://github.com/docxology/template OpenAlex discovery-only search (not a verified identity or literature match): https://openalex.org/works?search=Dynamic%20Action-Space%20Adaptation%20via%20Contrastive%20Transition%20Adaptation%20in%20Active%20Inference