Relation-Oriented: Toward Causal Knowledge-Aligned AGI

The current relationship modeling paradigm, grounded in the observational i.i.d assumption, inherently misaligns with our causal knowledge comprehension due to two vital oversights: 1) the unobservable relations, which lead to undetectable hierarchical levels of knowledge, driving the need for model generalizability; 2) the counterfactual relative timings to support our structural causal reasoning, which lead to inherent biases in models under the current Observation-Oriented paradigm. This paper proposes a novel Relation-Oriented framework, to reconsider these fundamental questions and unify various confusions surrounding AI-based causal learning, ranging from traditional causal inference to modern language models. Also, relation-indexed representation learning (RIRL) is raised as a baseline implementation method of the proposed new paradigm, alongside comprehensive experiments demonstrating its efficacy in autonomously identifying dynamical effects in relationship modeling.