ICL separability rises first, then alignment with the unembedding

measured in 1 paper

Yang et al. bound classification accuracy by hidden-state linear separability, requiring both output alignment and directional alignment with the label unembedding-difference vector [yang-etal-2025] Across 7 models on 6 datasets, ICL proceeds in two stages: separability rises rapidly in early layers, then middle-to-late layers spike four alignment measures together [yang-etal-2025] The ICL-versus-zero-shot separability gap is small despite an ~80-point accuracy gap, so ICL's gains come almost entirely from improved alignment, not separability [yang-etal-2025] Ablating Previous-Token Heads collapses separability while sparing alignment, and ablating Induction Heads collapses alignment and accuracy (to 2.5%), giving a geometric account of function vectors [yang-etal-2025]

Structure

Linear Direction, Linear Separability

Context

in-context learning, separability, alignment, phase transition, attention heads, task vectors

Confirmed in models

Llama-2-7B · Llama, Llama-2-13B · Llama, Llama-2-70B · Llama, Llama-3-8B · Llama, Llama-3-70B · Llama, Gemma-2B · Gemma, Gemma-7B · Gemma

Method

Linear probing, PCA, Causal interventions (steering), Function vector extraction (causal-mediation head selection + sum), Separability/alignment decomposition

Papers

Unifying Attention Heads and Task Vectors via Hidden State Geometry in In-Context Learning — Yang, Haolin, Cho, Hakaze, Zhong, Yiqiao, Inoue, Naoya2025 · arXiv:2505.18752