Where meaning lives: Layer-wise accessibility of psycholinguistic features in encoder and decoder language models

TL;DR

The paper investigates where meaning lives in transformer representations by conducting a comprehensive layer-wise probing study across 10 models (encoder and decoder) for 58 psycholinguistic features. It compares three embedding extraction methods and uses ridge-linear probes to quantify feature selectivity per layer, revealing that contextualized embeddings yield higher recoverability and that final layers often undercapture psycholinguistic information. A robust depth ordering emerges, with lexical features peaking earlier and experiential/affective features later, a pattern that generalizes across architectures albeit realized differently. These findings highlight that both embedding extraction choices and architectural constraints shape how meaning is distributed in neural language models, with important implications for interpretability, controllability, and deployment in meaning-sensitive applications.

Abstract

Understanding where transformer language models encode psychologically meaningful aspects of meaning is essential for both theory and practice. We conduct a systematic layer-wise probing study of 58 psycholinguistic features across 10 transformer models, spanning encoder-only and decoder-only architectures, and compare three embedding extraction methods. We find that apparent localization of meaning is strongly method-dependent: contextualized embeddings yield higher feature-specific selectivity and different layer-wise profiles than isolated embeddings. Across models and methods, final-layer representations are rarely optimal for recovering psycholinguistic information with linear probes. Despite these differences, models exhibit a shared depth ordering of meaning dimensions, with lexical properties peaking earlier and experiential and affective dimensions peaking later. Together, these results show that where meaning "lives" in transformer models reflects an interaction between methodological choices and architectural constraints.

Figures (18)

• Figure 1:
• Figure 2: The heatmaps visualize the average performance drop ($\Delta$) relative to the best-performing layer for each model across three embedding extraction methods: isolated, template, and averaged. The top row displays results for selectivity, while the bottom row shows raw $R^2$ scores. Within each panel, decoders are grouped at the top and encoders at the bottom, with the x-axis representing the normalized layer index ($0 = \text{first}$, $1 = \text{last}$). Red dots mark the single best layer (argmax) for each individual model-feature pair.
• Figure 3: Selectivity-weighted layer positions of psycholinguistic feature categories for the averaged embedding extraction method. The heatmap depicts the mean $\Delta$-from-best-layer across features comprising each category (X-axis: normalized layer index from first to last; Y-axis: psycholinguistic feature categories) based on selectivity score. Black dots indicate the selectivity-weighted center of mass (COM) of each feature’s layer profile, while red dots mark the single best-performing layer (argmax) for each feature. The top panel corresponds to the mean over all models, and the two bottom panels correspond to language models (decoders, top row, encoders, bottom row).
• Figure 4: Each panel shows the pairwise Spearman correlation (p) between models, computed from vectors of feature-specific center-of-mass (COM) layer positions within a given embedding extraction method (isolated, template, averaged). For each model and feature, the COM was calculated (using selectivity scores), summarizing where in the network a feature is most strongly represented. Correlations are computed across feature orders, yielding a similarity matrix that reflects how similarly different models localize psycholinguistic information across layers.
• Figure 6: Selectivity-weighted layer positions of psycholinguistic features in the isolated context. Each panel shows a language model (decoders top row, encoders bottom row). The heatmap depicts the $\Delta$-from-best-layer across layers for each feature (x-axis: normalized layer index from first to last; y-axis: psycholinguistic features) based on Selectivity score. Black dots indicate the Selectivity-weighted center of mass (COM) of each feature’s layer profile, while red dots mark the single best-performing layer (argmax). Lower (yellow) heatmap values indicate layers closer to the optimal representation of a feature.