Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Context-Dependent Relational Warping Underlies Triangle Inequality Violations in a Large Language Model

Sunsikham¹, selin samra², Taylor Whittington Webb³, Seng Bum Michael Yoo¹; ¹Sung Kyun Kwan University, ²Sungkyunkwan University, Institute for Basic Science, ³Université de Montréal

Presenter: Sunsikham

Relational similarity has often been explained by comparing relations within spatial or vector representations. These accounts infer similarities among relations from their positions or directions within a common representational space. Such comparisons are constrained by the geometry of that space. However, it remains unclear how spatial accounts explain judgments when a pair's inferred relation changes with context. Here we show that context-dependent relational judgments can arise within vector-based systems through shifts in the representation of the relation being compared. Human judgments showed selective violations of this geometric constraint: relational similarity followed the expected structure in consistent comparisons but broke down when the same item pair supported different relations across contexts. The LLM reproduced this asymmetry, showing larger violations in mixed than in consistent comparisons. Analyses of the model's hidden states showed that the ambiguous pair did not occupy a single stable representational position, but shifted toward different relation-specific directions across contexts. These findings suggest that apparent violations of geometric constraints need not imply the absence of vector-space representations; instead, they can arise when the relevant representations are dynamically reconfigured by context.

Topic Area: Memory, Learning & Knowledge Structures