IC-482In Llama-3.1-8B, in-context graph structure is absent in early layers dominated by semantic priors and emerges clearly in deeper layers

Core Francisco Park, Andrew Lee, Ekdeep Singh Lubana, Yongyi Yang, Maya Okawa, Kento Nishi, Martin Wattenberg, Hidenori Tanaka

SourceICLR: In-Context Learning of Representations

When processing in-context graph traces, Llama-3.1-8B's earlier layers (e.g., layer 6) show token representations organized by pretrained semantic relationships, while deeper layers (e.g., layer 26) show representations organized by the in-context graph geometry. In the ring graph, semantically related concepts (apple, orange) are closer together in layer 6 but become essentially antipodal by layer 26, matching the graph's structure. Layer-by-layer PCA projections (Figs. 9, 10) show the grid and hexagonal structures becoming progressively clearer from layer 0 through layer 30.

Evidence

correlational

Key metric

In the ring graph, concepts apple and orange are closer to each other in layer 6 but become essentially antipodal around layer 26, as dictated by the graph.

Caveat

The depth-dependent observation is based on visual inspection of 2D PCA projections, which the authors acknowledge can be difficult to interpret or misleading.

Model

Llama 3.1 8B

Concepts

Depth-dependent structure

Methods

Principal component analysis [primary]

Related findings

IC-481, IC-483, IC-484

Extraction

automatic-extraction