Word Class Representations Spontaneously Emerge from Successor Representations Trained on Natural Language

A neural network trained not on next-word prediction but on a reinforcement-learning concept called successor representations spontaneously organized words by part of speech, according to a paper submitted 23 May 2026 [1]. The finding suggests syntactic categories can emerge from predictive sequence learning without explicit linguistic supervision [1]. The researchers trained a deep residual neural network on the WikiText-103 dataset, which contains 103 million tokens and a 20,000-word vocabulary [1][2]. Instead of optimizing for the immediate next token, the network learned to predict future word distributions across multiple temporal horizons using Kullback-Leibler divergence as the loss function [1][2]. The successor representation framework, borrowed from reinforcement learning, models the expected discounted distribution of future states rather than the next state alone [2]. After training, the learned representation space exhibited a clear geometric organization aligned with part-of-speech categories. Nouns, verbs, and adjectives became separable and could be recovered through unsupervised clustering [1][2]. The strength of this syntactic structure depended systematically on the predictive horizon: short horizons produced the strongest part-of-speech clustering, while longer horizons integrated broader contextual and semantic information [1][2]. At finer resolutions, additional interpretable lexical substructure emerged, revealing coherent subclasses within major word categories [2]. The work establishes what the authors describe as the first systematic application of successor representations to natural language, creating a conceptual bridge between reinforcement learning, linguistics, and cognitive neuroscience [1][2]. The results imply that syntactic categories need not be hard-coded into language models but may arise as a natural consequence of learning to anticipate future linguistic states [1][2].