IC-483When in-context graph structure conflicts with pretrained semantic priors, Llama-3.1-8B encodes the in-context structure in higher principal components while the semantic prior dominates the first two
Core Francisco Park, Andrew Lee, Ekdeep Singh Lubana, Yongyi Yang, Maya Okawa, Kento Nishi, Martin Wattenberg, Hidenori Tanaka
Using days of the week (which have a circular semantic representation from pretraining) as nodes on a non-semantic 7-node ring, the model's first two principal components still show the original semantic ring geometry, while the in-context ring structure appears in the third and fourth principal components. The Dirichlet energy minimization occurs specifically in the PCA 3,4 dimensions corresponding to the in-context structure. Despite the semantic prior dominating the leading PCs, the model transitions from making semantic-prior predictions to in-context-structure predictions as context length increases, eventually performing well on the task.
Evidence
correlational
Key metric
Energy minimization occurs in PCA 3,4 (in-context dimensions) rather than PCA 1,2 (semantic dimensions), as shown in Fig. 15. The model transitions from semantic-prior accuracy to in-context accuracy with increasing context (Fig. 14).
Caveat
The in-context structure is present but 'unable to dominate the pretrained structure.' The authors note this is a limitation of the representation reorganization under semantic conflict.