Cross-LLM Consistency in Inference: Evidence from Shared Interactions

Large language models develop similar internal inference patterns despite differences in architecture and training data, according to a paper submitted to arXiv on 6 Jun 2026 [1]. The study finds that advanced models in particular share interaction patterns when predicting the same token from the same prompt [1]. The research, posted on the open-access repository arXiv, examines the hypothesis using interaction-based explanations [1]. Large language models, or LLMs, are neural networks trained on vast amounts of text for tasks such as generation and translation [10]. They typically rely on transformer architectures and are often pre-trained to predict the next word [10]. The paper notes that LLMs differ in architecture, training data, and optimization procedures, yet they may still converge on common inference patterns [1]. The authors report that LLMs often share interaction patterns when predicting the same target token from the same prompt [1]. This consistency is more pronounced among advanced LLMs [1]. Shared interactions also tend to be lower-order and show weaker positive-negative cancellation than non-shared interactions [1]. The findings suggest that advanced LLMs may be implicitly optimized toward common inference patterns, though the mechanisms that give rise to such cross-model consistency remain open [1]. Analogical reasoning, a cognitive process of transferring meaning from one subject to another, has been described as lying at "the core of cognition" [3]. The study's focus on shared inference patterns echoes broader questions about how models form internal representations that resemble analogical structures. Attribution theory in psychology examines how individuals perceive the causes of events as either internal or external [4]. Researchers have identified biases such as the fundamental attribution error, where people overemphasize dispositional factors over situational ones [4]. While the arXiv paper does not invoke psychological frameworks directly, its investigation into shared internal inference patterns raises parallel questions about whether models develop consistent internal "attributions" when processing language. arXiv itself, pronounced "archive," is an open-access repository of electronic preprints and postprints that are moderated but not peer reviewed [8]. It was begun on August 14, 1991, and by the end of 2021 had surpassed two million articles [8]. As of November 2024, the submission rate was about 24,000 articles per month [8]. The platform also hosts arXivLabs, a framework for community collaborators to develop experimental tools that appear on article record pages [7]. These tools include bibliographic explorers and recommenders, and partners must adhere to arXiv's values of openness, community, excellence, and user data privacy [7].