A Mechanistic Understanding of Pronoun Fidelity in LLMs

A new study offers the first mechanistic account of why large language models stumble over pronoun use, identifying three competing internal processes that together explain nearly all model behavior on the task [1]. The research, posted to arXiv on 15 June 2026, moves beyond behavioral testing to examine the internal causal pathways that govern pronoun fidelity — the ability to correctly use pronouns when multiple referents in a text use different pronouns [1]. Prior work on the topic relied exclusively on behavioral approaches, which the authors note “may not reflect a model's internal workings” [2]. The team instead applied Boundless Distributed Alignment Search to probe several state-of-the-art language models [1]. Large language models are machine learning systems trained on vast text corpora for natural language generation tasks [8]. The researchers tested three hypothesized mechanisms: group entity binding, recency bias, and stereotypical bias [1]. Group entity binding refers to linking an entity with its associated pronoun as a conceptual unit. Recency bias describes a tendency to repeat surface forms that appeared most recently in the text. Stereotypical bias captures the influence of learned associations — for example, linking certain occupations to specific pronouns [2]. The analysis found that all three mechanisms coexist as causal subspaces distributed across network depth, and no single mechanism fully explains model behavior [1]. A combination of the three, however, consistently accounts for 91-99.5% of model behavior across the tested systems [2]. The paper further identifies two competing copying routes within attention heads: group binding and stereotype share a localized concept-level route that retrieves a bound occupation-pronoun unit, while recency uses a distributed token-level route that repeats surface forms [1]. The findings arrive amid broader scrutiny of how large language models handle fairness and coherence. Models such as DeepSeek, developed by the Hangzhou-based firm founded in July 2023, and Alibaba Cloud’s Qwen family have drawn attention for their open-weight releases and competitive performance [7][9]. DeepSeek’s R1 model, launched in January 2025, provided responses comparable to OpenAI’s GPT-4 while reportedly costing far less to train [7]. The new pronoun fidelity study adds a layer of internal interpretability to such systems, showing that pronoun errors are not random failures but the product of competition between simultaneously active causal subspaces [1][2].