Language-Conditioned Latent Affordance Operators for Culturally Robust Physical Commonsense Reasoning

• Rizky Ananta

  Universitas Negeri Makassar, Jalan Daeng Tata Raya, Makassar 90222, Indonesia

  Author

• Bagas Wicaksono

  Universitas Jenderal Achmad Yani Yogyakarta, Jalan Ringroad Barat, Sleman 55292, Indonesia

  Author

Large-scale language models increasingly mediate everyday decision-making, yet many deployed settings require physical commonsense judgments that are both linguistically diverse and culturally situated. Physical commonsense is not only about abstract plausibility; it also encodes community-specific conventions over tools, materials, foods, and routines, and such conventions interact with morphology, script, and discourse style. This paper proposes a technical framework for culturally robust physical commonsense reasoning that does not rely on translating all inputs into a single pivot language. The core contribution is a latent affordance operator model that maps utterances to a canonical action-object state space, then performs constrained probabilistic inference over feasible physical transformations. Language enters through a conditional operator generator that produces distributions over affordance parameters, while an invariance regularizer drives cross-lingual agreement at the level of predicted physical effects rather than surface form. The resulting system separates language-specific expression from shared physical structure and admits principled uncertainty estimates when cultural priors differ. We formalize the inference problem as energy-based posterior decoding with compositional operators, derive training objectives that couple contrastive alignment with calibration constraints, and analyze robustness under distribution shift in the language channel. We also outline an evaluation protocol spanning text-only plausibility, tool-use planning, and counterfactual intervention queries, emphasizing measurement that is sensitive to cultural specificity while remaining physically grounded.

• PDF