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

• URL: http://arxiv.org/abs/2608.01816v1
• Date: Mon, 03 Aug 2026 07:26:45 GMT
• Title: Divergent large language model predictions from convergent representations in ambiguous word pairs
• Abstract summary: We investigate how decoder-only transformers resolve lexical ambiguity through layer-by-layer analysis.<n>For both homonyms and polysemes, we find that representations become maximally distinct in middle layers, then partially reconverge in late layers.
• Score: 2.4087148947930634
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: 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.

Related papers

• Mapping and Measuring the Behavioral Evolution of Large Language Models [28.603443859513153]
  Benchmark leaderboards summarize how well a language model performs, but not how its behavior relates to that of other models or changes across generations.<n>We characterize the output behavior of 32 models from six families using their responses to a shared bank of 10,000 prompts.
  arXiv  Detail & Related papers  (2026-08-11T15:07:28Z)
• Investigating Anisotropy in Visual Grounding under Controlled Counterfactual Perturbations [1.423280626666929]
  Grounding models are rarely evaluated under semantically mismatched captions.<n>This work examines whether embedding anisotropy contributes to counterfactual failures.
  arXiv  Detail & Related papers  (2026-05-09T17:54:31Z)
• On the Structural Non-Preservation of Epistemic Behaviour under Policy Transformation [51.56484100374058]
  We formalise such information-conditioned interaction patterns as behavioural dependency.<n>This induces a probe-relative notion of $$-behavioural equivalence and a within-policy behavioural distance.<n>Results identify structural conditions under which probe-conditioned behavioural separation is not preserved under common policy transformations.
  arXiv  Detail & Related papers  (2026-02-24T22:55:21Z)
• Convergence and Divergence of Language Models under Different Random Seeds [26.31262670498101]
  We measure convergence as the expected per-token Kullback--Leibler divergence across seeds.<n>We observe that larger models reconverge faster in later training stages, while smaller models never actually reconverge.
  arXiv  Detail & Related papers  (2025-09-30T17:59:50Z)
• Multi-Level Collaboration in Model Merging [56.31088116526825]
  This paper explores the intrinsic connections between model merging and model ensembling.<n>We find that even when previous restrictions are not met, there is still a way for model merging to attain a near-identical and superior performance similar to that of ensembling.
  arXiv  Detail & Related papers  (2025-03-03T07:45:04Z)
• Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity [20.17288970927518]
  We study the similarity of representations between the hidden layers of individual transformers.<n>We show that representations across layers are positively correlated, with similarity increasing when layers get closer.<n>We propose an aligned training method to improve the effectiveness of shallow layer.
  arXiv  Detail & Related papers  (2024-06-20T16:41:09Z)
• Exploring Activation Patterns of Parameters in Language Models [27.454051736471374]
  We propose a gradient-based metric to assess the activation level of model parameters. We find that parameters in the shallow layers will be activated densely, which means a larger portion of parameters will have great impacts on the outputs. In deep layers, the similarity of the distributions of activated parameters is positively correlated to the empirical data relevance.
  arXiv  Detail & Related papers  (2024-05-28T03:49:54Z)
• Duality-Induced Regularizer for Semantic Matching Knowledge Graph Embeddings [70.390286614242]
  We propose a novel regularizer -- namely, DUality-induced RegulArizer (DURA) -- which effectively encourages the entities with similar semantics to have similar embeddings. Experiments demonstrate that DURA consistently and significantly improves the performance of state-of-the-art semantic matching models.
  arXiv  Detail & Related papers  (2022-03-24T09:24:39Z)
• Towards Robust and Adaptive Motion Forecasting: A Causal Representation Perspective [72.55093886515824]
  We introduce a causal formalism of motion forecasting, which casts the problem as a dynamic process with three groups of latent variables. We devise a modular architecture that factorizes the representations of invariant mechanisms and style confounders to approximate a causal graph. Experiment results on synthetic and real datasets show that our three proposed components significantly improve the robustness and reusability of the learned motion representations.
  arXiv  Detail & Related papers  (2021-11-29T18:59:09Z)
• Learning Disentangled Representations with Latent Variation Predictability [102.4163768995288]
  This paper defines the variation predictability of latent disentangled representations. Within an adversarial generation process, we encourage variation predictability by maximizing the mutual information between latent variations and corresponding image pairs. We develop an evaluation metric that does not rely on the ground-truth generative factors to measure the disentanglement of latent representations.
  arXiv  Detail & Related papers  (2020-07-25T08:54:26Z)
• Learning from Aggregate Observations [82.44304647051243]
  We study the problem of learning from aggregate observations where supervision signals are given to sets of instances. We present a general probabilistic framework that accommodates a variety of aggregate observations. Simple maximum likelihood solutions can be applied to various differentiable models.
  arXiv  Detail & Related papers  (2020-04-14T06:18:50Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.