Search papers, labs, and topics across Lattice.
This study investigates the mechanisms underlying pronoun fidelity in large language models (LLMs) by examining group entity binding, recency bias, and stereotypical bias as causal factors. Using Boundless Distributed Alignment Search, the authors demonstrate that these three mechanisms coexist and collectively account for 91-99.5% of model behavior, revealing a complex interplay rather than a single explanatory factor. An analysis of attention heads further uncovers two competing routes for pronoun retrieval, highlighting the nuanced internal processes that govern pronoun use in LLMs.
Pronoun fidelity in LLMs is driven by a competition between group binding, recency, and stereotype biases, revealing a complex internal landscape that challenges simplistic behavioral interpretations.
Faithful and robust pronoun use is important for fair and coherent generations, yet large language models largely fail when multiple referents use different pronouns. To study the interplay of reasoning, repetition, and bias in this task, prior work relies exclusively on behavioural approaches, which may not reflect a model's internal workings. Therefore, we provide a mechanistic, model-internal perspective on pronoun fidelity, testing whether three mechanisms -- group entity binding (G), recency bias (R), and stereotypical bias (S) -- are causally implemented across several SOTA language models. Using Boundless Distributed Alignment Search, we find all three coexist as causal subspaces distributed across network depth. No single mechanism fully explains model behaviour, but a combination of the three consistently accounts for 91-99.5%. An attention head analysis further reveals two competing copying routes; 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. In sum, pronoun fidelity arises from competition between simultaneously active causal subspaces.
