ACL Findings2024
Fantastic Semantics and Where to Find Them: Investigating Which Layers of Generative LLMs Reflect Lexical Semantics
Zhu Liu, Cunliang Kong, Ying Liu, Maosong Sun
Abstract
Large language models have achieved remarkable success in general language understanding tasks. However, as a family of generative methods with the objective of next token prediction, the semantic evolution with the depth of these models are not fully explored, unlike their predecessors, such as BERT-like architectures. In this paper, we specifically investigate the bottom-up evolution of lexical semantics for a popular LLM, namely Llama2, by probing its hidden states at the end of each layer using a contextualized word identification task. Our experiments show that the representations in lower layers encode lexical semantics, while the higher layers, with weaker semantic induction, are responsible for prediction. This is in contrast to models with discriminative objectives, such as mask language modeling, where the higher layers obtain better lexical semantics. The conclusion is further supported by the monotonic increase in performance via the hidden states for the last meaningless symbols, such as punctuation, in the prompting strategy. Our codes are available at https: //github.com/RyanLiut/LLM_LexSem.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
16 top-tier papers have cited it since. Know what they found before your reviewers do.
Lune can draw on
the full paper · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
- Hierarchical Token Prepending: Enhancing Information Flow in Decoder-based LLM Embeddings cites Scaling Sentence Embeddings with Large Language Models
- Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 Small cites Causal Abstractions of Neural Networks
- Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 Small cites Language Models are Few-Shot Learners
Cited by top-tier papers16
- GMSA: Enhancing Context Compression via Group Merging and Layer Semantic AlignmentJiwei Tang, Zhicheng Zhang, Shunlong Wu, Jingheng Ye et al.ACL 2026 · 24 citations
- Remarkable Robustness of LLMs: Stages of Inference?Vedang Lad, Jin Hwa Lee, Wes Gurnee, Max TegmarkNeurIPS 2025 · 5 citations
- Hierarchical Token Prepending: Enhancing Information Flow in Decoder-based LLM EmbeddingsXueying Ding, Xingyue Huang, Mingxuan Ju, Liam Collins et al.ACL 2026 · 3 citations
- From Associations to Activations: Comparing Behavioral and Hidden-State Semantic Geometry in LLMsLouis Schiekiera, Max Zimmer, Christophe Roux, Sebastian Pokutta et al.ICML 2026 · 2 citations
- TRIM: Token-wise Attention-Derived Saliency for Data-Efficient Instruction TuningManish Nagaraj, Sakshi Choudhary, Utkarsh Saxena, Deepak Ravikumar et al.ICML 2026 · 2 citations
Builds on6
- Language Models are Few-Shot LearnersTom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah et al.NeurIPS 2020 · 64,255 citations
- Causal Abstractions of Neural NetworksAtticus Geiger, Hanson Lu, Thomas Icard, Christopher PottsNeurIPS 2021 · 516 citations
- Scaling Sentence Embeddings with Large Language ModelsTing Jiang, Shaohan Huang, Zhongzhi Luan, Deqing Wang et al.EMNLP Findings 2024 · 61 citations
- ConSeC: Word Sense Disambiguation as Continuous Sense ComprehensionEdoardo Barba, Luigi Procopio, Roberto NavigliEMNLP 2021 · 60 citations
- Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 SmallKevin Ro Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris et al.ICLR 2023 · 50 citations
Related papers
- On the Origins of Linear Representations in Large Language ModelsYibo Jiang, Goutham Rajendran, Pradeep Kumar Ravikumar, Bryon Aragam et al.ICML 2024 · 68 citations
- LLaMAs Have Feelings Too: Unveiling Sentiment and Emotion Representations in LLaMA Models Through ProbingDario Di Palma, Alessandro De Bellis, Giovanni Servedio, Vito Walter Anelli et al.ACL 2025 · 11 citations
- Automatically Generated Definitions and their utility for Modeling Word MeaningFrancesco Periti, David Alfter, Nina TahmasebiEMNLP 2024 · 2 citations
- Measuring Intent Comprehension in LLMsNadav Kunievsky, James EvansICML 2026 · 1 citation
- Differential syntactic and semantic encoding in LLMsSantiago Acevedo, Alessandro Laio, Marco BaroniICML 2026 · 7 citations
