A Mechanistic Understanding of Pronoun Fidelity in LLMs
Authors: Katharina Trinley, Jesujoba O. Alabi, Dietrich Klakow, Vagrant Gautam
Organizations: Saarland University, Germany · Heidelberg Institute for Theoretical Studies, Germany
Abstract
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.
Third-person singular pronouns have long been used to study stereotypical biases in language models and to test their abilities to reason about reference. More recently, the interplay between reasoning and bias has been investigated with the task of pronoun fidelity, which assesses models' abilities to correctly reuse a previously-specified pronoun for a discourse entity, independent of other potentially distracting discourse entities mentioned in between. However, such research focuses on English, which is a language with limited grammatical gender and almost no gender agreement. In this paper we contribute a novel, large-scale dataset, GRUFF, to measure pronoun fidelity in German, covering four different gender agreement systems in nouns, and four sets of pronouns. With this dataset, we show that LLMs show strong grammatical agreement for masculine and feminine entities in the absence of explicit context, but not for neopronouns xier and en. Models are generally not robust to distractors, but encoder-only models are more robust in German than in English, reflecting the importance of grammatical gender. Finally, we show that occupational stereotypes in this context are poorly correlated across grammatical cases, and across most models, except ones with closely related architectures. We release all code and data to encourage further work on gender-inclusive language and referential reasoning in German.
Large language models (LLMs) are increasingly used as automatic judges for summarization and dialogue evaluation. Prior work has documented biases such as position, verbosity, and style preferences, but largely focuses on outcomes, leaving judge explanations underexplored. We instead ask whether LLM judges are cue-invariant, i.e., whether their rankings and explanations remain stable when non-evidential cues are perturbed while holding the underlying texts fixed. We introduce a suite of cue interventions (Blind, Truth, Flip, Placebo, Reveal-After) and tie-aware metrics that quantify outcome anchoring and rationale anchoring, including label-aligned rhetoric and explanation drift, alongside consistency and stereotype-intrusion checks. We design anchoring attacks using verbosity and confidence cues, and compare two mitigations: structured chain-of-thought prompting and PROOF-BEFORE-PREFERENCE (evidence lock, score, rank). Using a new dataset of 1,000 summaries from traditional extractive models and LLMs, we find substantial cue-anchored rationalization under label and placebo perturbations, while PROOF-BEFORE-PREFERENCE markedly improves cue invariance over baselines.
When a language model agrees with a user's false belief, is it failing to detect the error, or noticing and agreeing anyway? We show the latter. Across twelve open-weight models from five labs, spanning small to frontier scale, the same small set of attention heads carries a "this statement is wrong" signal, whether the model is evaluating a claim on its own or being pressured to agree with a user. Silencing these heads flips sycophantic behavior sharply while leaving factual accuracy intact, so the circuit controls deference rather than knowledge. Edge-level path patching confirms that the same head-to-head connections drive sycophancy, factual lying, and instructed lying. Opinion-agreement, where no factual ground truth exists, reuses these head positions but writes into an orthogonal direction, ruling out a simple "truth-direction" reading of the substrate. Alignment training leaves this circuit in place: an RLHF refresh cuts sycophantic behavior roughly tenfold while the shared heads persist or grow, a pattern that replicates on an independent model family and under targeted anti-sycophancy DPO. When these models sycophant, they register that the user is wrong and agree anyway.