# Divergent large language model predictions from convergent representations in ambiguous word pairs

arXiv: https://arxiv.org/abs/2608.01816 · HTML: https://mlsanity.com/paper/2608.01816/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 11 of its 25 references that are in the map.
- First posted on arXiv 3 Aug 2026.

## Builds on (most cited first, 11 of 11)

- [SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing](https://mlsanity.com/paper/1808.06226/) (2018)
- [How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddings](https://mlsanity.com/paper/1909.00512/) (2019)
- [Eliciting Latent Predictions from Transformers with the Tuned Lens](https://mlsanity.com/paper/2303.08112/) (2023)
- [Layer by Layer: Uncovering Hidden Representations in Language Models](https://mlsanity.com/paper/2502.02013/) (2025)
- [How to use and interpret activation patching](https://mlsanity.com/paper/2404.15255/) (2024)
- [All Bark and No Bite: Rogue Dimensions in Transformer Language Models Obscure Representational Quality](https://mlsanity.com/paper/2109.04404/) (2021)
- [The Remarkable Robustness of LLMs: Stages of Inference?](https://mlsanity.com/paper/2406.19384/) (2024)
- [Is Cosine-Similarity of Embeddings Really About Similarity?](https://mlsanity.com/paper/2403.05440/) (2024)
- [Verbalizable Representations Form a Global Workspace in Language Models](https://mlsanity.com/paper/2607.15495/) (2026)
- [Attention Sinks and Compression Valleys in LLMs are Two Sides of the Same Coin](https://mlsanity.com/paper/2510.06477/) (2025)
- [How Do Multilingual Language Models Remember Facts?](https://mlsanity.com/paper/2410.14387/) (2024)

## Built on by (most cited first, 1 of 1)

- [Technical Manual for a Toolkit for Measuring Contextual Individuation in Transformer Language Models](https://mlsanity.com/paper/2609.05333/) (2026)

## Abstract (arXiv)

In this work we investigate how decoder-only transformers resolve lexical ambiguity through layer-by-layer analysis of three models spanning three parameter sizes (GPT-2-Small-117M, Llama-3.2-3B, Qwen2.5-32B). For both homonyms and polysemes, we find that representations become maximally distinct in middle layers, then partially reconverge in late layers, while the KL divergence between their next-token predictions reaches its maximum in the final layers. The activation patching experiment provides causal evidence that late-layer representational differences directly determine outputs despite apparent increased similarity in embedding space. Our single-layer ablation experiment indicates that models achieve equivalent disambiguation despite qualitatively different layer-wise vulnerabilities. These findings offer a mechanism for recent observations where models' internal embedding similarities show low correlation with their behavioural outputs despite strong performance. The semantic distinctions therefore remain present but become increasingly invisible to similarity measures over the embeddings, with implications for embedding-based methods such as semantic search, retrieval, and clustering that rely on late-layer cosine similarity.