IC-547Llama3.2 3B and Llama3.1 8B exhibit saturation events in which the top prediction, once it appears at a given layer, remains unchanged through all subsequent layers

Jiachen Jiang, Jinxin Zhou, Zhihui Zhu

SourceTracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity

The authors apply the logit lens (last-layer classifier) to intermediate hidden states of Llama3.2 3B (24 layers) and Llama3.1 8B (32 layers) using the same prompt and greedy decoding. They observe that the model's predicted token stabilises at a certain layer and does not change in any deeper layer, a phenomenon previously described by Geva et al. (2022) for smaller models. A specific instance in Llama3.2 3B with the prompt 'simply put, the theory of relativity states' is visualised across 24 decoding steps, showing the predicted word locking in at an intermediate layer. This confirms that the saturation-event phenomenon generalises to large, recently released LLMs.

Evidence
observational
Caveat
The observation is made on a single prompt per model and a small number of decoding steps; no quantitative frequency of saturation events across a corpus is reported.
Model
Llama-3.2-3B, Llama 3.1 8B
Concepts
Depth-dependent structure
Methods
Logit lens [primary]
Related findings
IC-548
Extraction
automatic-extraction