arXiv subjects

Keren Fuentes

Publications and source records attributed to Keren Fuentes.

3 recordsLinked to original sources

arXiv2026

Who Do Language Models Think Is Competent? A Mechanistic Analysis of Occupational Bias

Language models (LMs) often pass behavioral bias evaluations, but it remains unclear whether they no longer represent the underlying associations that give rise to biases, or have merely learned not to express them. In this study, we show that representational biases are often detectable, even when behavioral biases are not visible. We introduce a causal framework that decomposes occupational bias into two measurement points: a model's internal representation of a user's competence, and its observable outputs. We derive steering vectors for representations of user expertise, and verify that they causally mediate model behavior in both a question-answering task and a hiring task. Applying this framework to several open-weight models, we find that demographic attributes, such as gender, race, and socioeconomic status, influence a model's representation of user expertise, even in cases where behavioral metrics detect no disparity between demographics. We show that these model representations can influence downstream behavior under intervention, suggesting failure modes that behavioral metrics alone may not detect.

cs.CL↗

arXiv2025

Privacy-Preserving Dataset Combination

Access to diverse, high-quality datasets is crucial for machine learning model performance, yet data sharing remains limited by privacy concerns and competitive interests, particularly in regulated domains like healthcare. This dynamic especially disadvantages smaller organizations that lack resources to purchase data or negotiate favorable sharing agreements, due to the inability to \emph{privately} assess external data's utility. To resolve privacy and uncertainty tensions simultaneously, we introduce {\SecureKL}, the first secure protocol for dataset-to-dataset evaluations with zero privacy leakage, designed to be applied preceding data sharing. {\SecureKL} evaluates a source dataset against candidates, performing dataset divergence metrics internally with private computations, all without assuming downstream models. On real-world data, {\SecureKL} achieves high consistency ($>90\%$ correlation with non-private counterparts) and successfully identifies beneficial data collaborations in highly-heterogeneous domains (ICU mortality prediction across hospitals and income prediction across states). Our results highlight that secure computation maximizes data utilization, outperforming privacy-agnostic utility assessments that leak information.

cs.LG↗

arXiv2022

Language model acceptability judgements are not always robust to context

Targeted syntactic evaluations of language models ask whether models show stable preferences for syntactically acceptable content over minimal-pair unacceptable inputs. Most targeted syntactic evaluation datasets ask models to make these judgements with just a single context-free sentence as input. This does not match language models' training regime, in which input sentences are always highly contextualized by the surrounding corpus. This mismatch raises an important question: how robust are models' syntactic judgements in different contexts? In this paper, we investigate the stability of language models' performance on targeted syntactic evaluations as we vary properties of the input context: the length of the context, the types of syntactic phenomena it contains, and whether or not there are violations of grammaticality. We find that model judgements are generally robust when placed in randomly sampled linguistic contexts. However, they are substantially unstable for contexts containing syntactic structures matching those in the critical test content. Among all tested models (GPT-2 and five variants of OPT), we significantly improve models' judgements by providing contexts with matching syntactic structures, and conversely significantly worsen them using unacceptable contexts with matching but violated syntactic structures. This effect is amplified by the length of the context, except for unrelated inputs. We show that these changes in model performance are not explainable by simple features matching the context and the test inputs, such as lexical overlap and dependency overlap. This sensitivity to highly specific syntactic features of the context can only be explained by the models' implicit in-context learning abilities.

cs.CL↗