Lexical relations are not only labels on pairs of concepts: they are directed typed movements that can form paths. We study whether sentence embedding spaces preserve this structure at the level of individual relation instances. By representing two-hop lexical chains as composed movements on the normalised embedding sphere, we test whether true chains form more coherent routes than matched false-middle alternatives. Across controlled lexical path experiments, real chains exhibit a robust path signal, including for mixed relation compositions, and the composed movements retain information about ordered relation type. Beyond diagnosis, the same path geometry aligns with independent distributional context shifts and improves graded lexical-entailment transfer, indicating that it captures reusable lexical structure rather than a task-specific artefact.

Sample diversity depends on the task; within mathematics, precision and determinism are paramount, while storytelling thrives on creativity and surprise. This paper presents a simple self-regulating approach where we adjust sample diversity inference parameters dynamically based on the input prompt—in contrast to existing methods that require expensive and inflexible setups, or maintain static values during inference. Capturing a broad spectrum of sample diversities can be formulated as a straightforward self-supervised inference task, which we find significantly improves the quality of responses generically without model retraining or fine-tuning. In particular, our method demonstrates significant improvement in all supercategories of the MMLU multitask benchmark (GPT-3.5: +4.4%, GPT-4: +1.5%), which captures a large variety of difficult tasks covering STEM, the humanities and social sciences.

With the identification of the inequality encoded in information acquisition among social classes, we propose to leverage a powerful concept that has never been studied as a linguistic construct, “fun”, to deconstruct the inequality. Inspired by theories in sociology, we draw connection between social class and information cocoon, through the lens of fun, and hypothesize the measurement of “how fun one’s dominating social cocoon is” to be an indicator of the social class of an individual. Following this, we propose an NLP framework to combat the issue by measuring how fun one’s information cocoon is, and empower individuals to emancipate from their trapped cocoons. We position our work to be a domain-agnostic framework that can be deployed in a lot of downstream cases, and is one that aims to deconstruct, as opposed to reinforcing, the traditional social structure of beneficiaries.