IC-043Five ~7B decoder-only LLMs develop a high-intrinsic-dimensionality phase in intermediate layers that marks the transition from surface-form to abstract linguistic processing, with earlier onset predicting better next-token prediction

Emily Cheng, Diego Doimo, Corentin Kervadec, Iuri Macocco, Lei Yu, Alessandro Laio, Marco Baroni

SourceEmergence of a High-Dimensional Abstraction Phase in Language Transformers

The paper estimates the intrinsic dimension (ID) of last-token hidden representations at every layer of five ~7B LMs (OPT-6.7B, Llama-3-8B, Pythia-6.9B, OLMo-7B, Mistral-7B) using the GRIDE estimator on three corpora. All models exhibit a distinct ID peak in intermediate layers (roughly layers 6–20, later for OPT at 17–23), which is absent when processing shuffled text and in untrained Pythia checkpoints. MLP probes show that surface-form information (sentence length, word content) is lost at the ID peak while syntactic and semantic tasks (bigram shift, coordination inversion, odd man out) reach asymptotic accuracy within the peak. Cross-model information imbalance (Δ) is lowest at the intersection of ID peaks, indicating shared representation geometry across architectures. The onset of the ID peak correlates with lower surprisal (ρ = 0.65, p = 0.01), and ID-peak layers are the first to support viable linear transfer to sentiment and toxicity classification.

Evidence
correlational
Key metric
ID peak at layers ~6–20 (most models), 17–23 (OPT); ID magnitude O(10) vs hidden dim 4096; surprisal vs max ID: Spearman ρ = −0.46, p = 0.09; surprisal vs relative ID peak onset: ρ = 0.65, p = 0.01; cross-model Δ at peak intersections in range [0, 0.03]–[0.05, 0.07]
Caveat
The correlation between maximum ID and surprisal is only marginal (p = 0.09). The authors note that low-Δ values can also occur outside peak intersections, suggesting high ID might be a sufficient rather than necessary condition for cross-model similarity. All models are in a narrow size range (6.5–8B parameters, 32 layers, hidden dim 4096), limiting generalisability to other scales.
Model
OPT 6.7B, Llama 3 8B, Pythia Pythia-6.9B, 2.8B, Pythia-12B, OLMo / OLMo base OLMo-7B, Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1
Concepts
Depth-dependent structure
Datasets
BookCorpus [eval], The Pile [eval], Wikitext-103 [eval], Conneau et al. Probing Tasks [eval], IMDB [eval]
Methods
GRIDE [primary], Information Imbalance [primary], Probing classifiers / MLP probing classifiers / Q16 classifier [primary], Linear CKA [validation]
Extraction
automatic-extraction