arXiv subjects

Hillary Dawkins

Publications and source records attributed to Hillary Dawkins.

17 recordsLinked to original sources

arXiv2026

Safety Measurements for Fine-tuned LLMs Should be Grounded in Capability

Adapting foundation large language models to a user's task or preferred style through fine-tuning can result in compromising the model's safety. Previous works examined the effects of fine-tuning on model safety in limited and seemingly random experimental settings. We argue that anchoring fine-tuning to a specific capability goal is essential for avoiding arbitrary empirical choices, allowing us to draw meaningful conclusions about safety impacts, and to compare mitigation methods on a consistent basis. We conduct a multi-dimensional evaluation of the effects of fine-tuning on model behavior by focusing on capability as well as safety. Our results surface important issues that (1) fine-tuned models can produce incoherent generations in response to safety prompts, (2) automated safety judgments are unreliable for such incoherent outputs, and (3) the conclusions about the effects of fine-tuning can change depending on the choice of safety benchmark as well as the safety evaluator.

cs.CL↗

arXiv2025

Fine-Tuning Lowers Safety and Disrupts Evaluation Consistency

Fine-tuning a general-purpose large language model (LLM) for a specific domain or task has become a routine procedure for ordinary users. However, fine-tuning is known to remove the safety alignment features of the model, even when the fine-tuning data does not contain any harmful content. We consider this to be a critical failure mode of LLMs due to the widespread uptake of fine-tuning, combined with the benign nature of the "attack". Most well-intentioned developers are likely unaware that they are deploying an LLM with reduced safety. On the other hand, this known vulnerability can be easily exploited by malicious actors intending to bypass safety guardrails. To make any meaningful progress in mitigating this issue, we first need reliable and reproducible safety evaluations. In this work, we investigate how robust a safety benchmark is to trivial variations in the experimental procedure, and the stochastic nature of LLMs. Our initial experiments expose surprising variance in the results of the safety evaluation, even when seemingly inconsequential changes are made to the fine-tuning setup. Our observations have serious implications for how researchers in this field should report results to enable meaningful comparisons in the future.

cs.CL↗

arXiv2025

Gender-Neutral Machine Translation Strategies in Practice

Gender-inclusive machine translation (MT) should preserve gender ambiguity in the source to avoid misgendering and representational harms. While gender ambiguity often occurs naturally in notional gender languages such as English, maintaining that gender neutrality in grammatical gender languages is a challenge. Here we assess the sensitivity of 21 MT systems to the need for gender neutrality in response to gender ambiguity in three translation directions of varying difficulty. The specific gender-neutral strategies that are observed in practice are categorized and discussed. Additionally, we examine the effect of binary gender stereotypes on the use of gender-neutral translation. In general, we report a disappointing absence of gender-neutral translations in response to gender ambiguity. However, we observe a small handful of MT systems that switch to gender neutral translation using specific strategies, depending on the target language.

cs.CL↗

arXiv2025

When Detection Fails: The Power of Fine-Tuned Models to Generate Human-Like Social Media Text

Detecting AI-generated text is a difficult problem to begin with; detecting AI-generated text on social media is made even more difficult due to the short text length and informal, idiosyncratic language of the internet. It is nonetheless important to tackle this problem, as social media represents a significant attack vector in online influence campaigns, which may be bolstered through the use of mass-produced AI-generated posts supporting (or opposing) particular policies, decisions, or events. We approach this problem with the mindset and resources of a reasonably sophisticated threat actor, and create a dataset of 505,159 AI-generated social media posts from a combination of open-source, closed-source, and fine-tuned LLMs, covering 11 different controversial topics. We show that while the posts can be detected under typical research assumptions about knowledge of and access to the generating models, under the more realistic assumption that an attacker will not release their fine-tuned model to the public, detectability drops dramatically. This result is confirmed with a human study. Ablation experiments highlight the vulnerability of various detection algorithms to fine-tuned LLMs. This result has implications across all detection domains, since fine-tuning is a generally applicable and realistic LLM use case.

cs.CL↗

arXiv2024

WMT24 Test Suite: Gender Resolution in Speaker-Listener Dialogue Roles

We assess the difficulty of gender resolution in literary-style dialogue settings and the influence of gender stereotypes. Instances of the test suite contain spoken dialogue interleaved with external meta-context about the characters and the manner of speaking. We find that character and manner stereotypes outside of the dialogue significantly impact the gender agreement of referents within the dialogue.

cs.CL↗

arXiv2024

Adaptable Moral Stances of Large Language Models on Sexist Content: Implications for Society and Gender Discourse

This work provides an explanatory view of how LLMs can apply moral reasoning to both criticize and defend sexist language. We assessed eight large language models, all of which demonstrated the capability to provide explanations grounded in varying moral perspectives for both critiquing and endorsing views that reflect sexist assumptions. With both human and automatic evaluation, we show that all eight models produce comprehensible and contextually relevant text, which is helpful in understanding diverse views on how sexism is perceived. Also, through analysis of moral foundations cited by LLMs in their arguments, we uncover the diverse ideological perspectives in models' outputs, with some models aligning more with progressive or conservative views on gender roles and sexism. Based on our observations, we caution against the potential misuse of LLMs to justify sexist language. We also highlight that LLMs can serve as tools for understanding the roots of sexist beliefs and designing well-informed interventions. Given this dual capacity, it is crucial to monitor LLMs and design safety mechanisms for their use in applications that involve sensitive societal topics, such as sexism.

cs.CL↗

arXiv2024

Detecting AI-Generated Text: Factors Influencing Detectability with Current Methods

Large language models (LLMs) have advanced to a point that even humans have difficulty discerning whether a text was generated by another human, or by a computer. However, knowing whether a text was produced by human or artificial intelligence (AI) is important to determining its trustworthiness, and has applications in many domains including detecting fraud and academic dishonesty, as well as combating the spread of misinformation and political propaganda. The task of AI-generated text (AIGT) detection is therefore both very challenging, and highly critical. In this survey, we summarize state-of-the art approaches to AIGT detection, including watermarking, statistical and stylistic analysis, and machine learning classification. We also provide information about existing datasets for this task. Synthesizing the research findings, we aim to provide insight into the salient factors that combine to determine how "detectable" AIGT text is under different scenarios, and to make practical recommendations for future work towards this significant technical and societal challenge.

cs.CL↗

arXiv2024

Projective Methods for Mitigating Gender Bias in Pre-trained Language Models

Mitigation of gender bias in NLP has a long history tied to debiasing static word embeddings. More recently, attention has shifted to debiasing pre-trained language models. We study to what extent the simplest projective debiasing methods, developed for word embeddings, can help when applied to BERT's internal representations. Projective methods are fast to implement, use a small number of saved parameters, and make no updates to the existing model parameters. We evaluate the efficacy of the methods in reducing both intrinsic bias, as measured by BERT's next sentence prediction task, and in mitigating observed bias in a downstream setting when fine-tuned. To this end, we also provide a critical analysis of a popular gender-bias assessment test for quantifying intrinsic bias, resulting in an enhanced test set and new bias measures. We find that projective methods can be effective at both intrinsic bias and downstream bias mitigation, but that the two outcomes are not necessarily correlated. This finding serves as a warning that intrinsic bias test sets, based either on language modeling tasks or next sentence prediction, should not be the only benchmark in developing a debiased language model.

cs.CL↗

arXiv2021

Marked Attribute Bias in Natural Language Inference

Reporting and providing test sets for harmful bias in NLP applications is essential for building a robust understanding of the current problem. We present a new observation of gender bias in a downstream NLP application: marked attribute bias in natural language inference. Bias in downstream applications can stem from training data, word embeddings, or be amplified by the model in use. However, focusing on biased word embeddings is potentially the most impactful first step due to their universal nature. Here we seek to understand how the intrinsic properties of word embeddings contribute to this observed marked attribute effect, and whether current post-processing methods address the bias successfully. An investigation of the current debiasing landscape reveals two open problems: none of the current debiased embeddings mitigate the marked attribute error, and none of the intrinsic bias measures are predictive of the marked attribute effect. By noticing that a new type of intrinsic bias measure correlates meaningfully with the marked attribute effect, we propose a new postprocessing debiasing scheme for static word embeddings. The proposed method applied to existing embeddings achieves new best results on the marked attribute bias test set. See https://github.com/hillary-dawkins/MAB.

cs.CL↗

arXiv2021

Second Order WinoBias (SoWinoBias) Test Set for Latent Gender Bias Detection in Coreference Resolution

We observe an instance of gender-induced bias in a downstream application, despite the absence of explicit gender words in the test cases. We provide a test set, SoWinoBias, for the purpose of measuring such latent gender bias in coreference resolution systems. We evaluate the performance of current debiasing methods on the SoWinoBias test set, especially in reference to the method's design and altered embedding space properties. See https://github.com/hillarydawkins/SoWinoBias.

cs.CL↗

arXiv2020

Combining $T_1$ and $T_2$ estimation with randomized benchmarking and bounding the diamond distance

The characterization of errors in a quantum system is a fundamental step for two important goals. First, learning about specific sources of error is essential for optimizing experimental design and error correction methods. Second, verifying that the error is below some threshold value is required to meet the criteria of threshold theorems. We consider the case where errors are dominated by the generalized damping channel (encompassing the common intrinsic processes of amplitude damping and dephasing) but may also contain additional unknown error sources. We demonstrate the robustness of standard $T_1$ and $T_2$ estimation methods and provide expressions for the expected error in these estimates under the additional error sources. We then derive expressions that allow a comparison of the actual and expected results of fine-grained randomized benchmarking experiments based on the damping parameters. Given the results of this comparison, we provide bounds that allow robust estimation of the thresholds for fault-tolerance.

quant-ph↗

arXiv2017

Data science for urban equity: Making gentrification an accessible topic for data scientists, policymakers, and the community

The University of Washington eScience Institute runs an annual Data Science for Social Good (DSSG) program that selects four projects each year to train students from a wide range of disciplines while helping community members execute social good projects, often with an urban focus. We present observations and deliberations of one such project, the DSSG 2017 'Equitable Futures' project, which investigates the ongoing gentrification process and the increasingly inequitable access to opportunities in Seattle. Similar processes can be observed in many major cities. The project connects issues usually analyzed in the disciplines of the built environment, geography, sociology, economics, social work and city governments with data science methodologies and visualizations.

cs.CY↗

arXiv2017

The impact of scaffolding and question structure on the gender gap

We address previous hypotheses about possible factors influencing the gender gap in attainment in physics. Specifically, previous studies claim that male advantage may arise from multiple-choice style questions, and that scaffolding may preferentially benefit female students. We claim that female students are not disadvantaged by multiple-choice style questions, and also present some alternative conclusions surrounding the scaffolding hypothesis. By taking both student attainment level and the degree of question scaffolding into account, we identify questions which exhibit real bias in favour of male students. We find that both multi-dimensional context and use of diagrams are common elements of such questions.

physics.ed-ph↗

arXiv2016

Small Codes for Magic State Distillation

Magic state distillation is a critical component in leading proposals for fault-tolerant quantum computation. Relatively little is known, however, about how to construct a magic state distillation routine or, more specifically, which stabilizer codes are suitable for the task. While transversality of a non-Clifford gate within a code often leads to efficient distillation routines, it appears to not be a necessary condition. Here we have examined a number of small stabilizer codes and highlight a handful of which displaying interesting, albeit inefficient, distillation behaviour. Many of these distill noisy states right up to the boundary of the known undististillable region, while some distill toward non-stabilizer states that have not previously been considered.

quant-ph↗

arXiv2016

Diffusion Monte Carlo study of strongly interacting two-dimensional Fermi gases

Ultracold atomic Fermi gases have been a popular topic of research, with attention being paid recently to two-dimensional (2D) gases. In this work, we perform T=0 ab initio diffusion Monte Carlo calculations for a strongly interacting two-component Fermi gas confined to two dimensions. We first go over finite-size systems and the connection to the thermodynamic limit. After that, we illustrate pertinent 2D scattering physics and properties of the wave function. We then show energy results for the strong-coupling crossover, in between the Bose-Einstein Condensation (BEC) and Bardeen-Cooper-Schrieffer (BCS) regimes. Our energy results for the BEC-BCS crossover are parametrized to produce an equation of state, which is used to determine Tan's contact. We carry out a detailed comparison with other microscopic results. Finally, we calculate the pairing gap for a range of interaction strengths in the strong coupling regime, following from variationally optimized many-body wave functions.

cond-mat.quant-gas↗

arXiv2015

Qutrit Magic State Distillation Tight in Some Directions

Magic state distillation is a crucial component in the leading approaches to implementing universal fault tolerant quantum computation, with existing protocols for both qubit and higher dimensional systems. Early work focused on determining the region of distillable states for qubit protocols, yet comparatively little is known about which states can be distilled and with what distillable region for d>2. Here we focus on d=3 and present new four-qutrit distillation schemes that improve upon the known distillable region, and achieve distillation tight to the boundary of undistillable states for some classes of state. As a consequence of recent results, this implies that there is a family of quantum states that enable universality if and only if they exhibit contextuality with respect to stabilizer measurements. We also identify a new routine whose fixed point is a magic state with maximal sum-negativity i.e., it is maximally non-stabilizer in a specific sense.

quant-ph↗

arXiv2013

Uniqueness of Quantum States Compatible with Given Measurement Results

We discuss the uniqueness of quantum states compatible with given results for measuring a set of observables. For a given pure state, we consider two different types of uniqueness: (1) no other pure state is compatible with the same measurement results and (2) no other state, pure or mixed, is compatible with the same measurement results. For case (1), it is known that for a d-dimensional Hilbert space, there exists a set of 4d-5 observables that uniquely determines any pure state. We show that for case (2), 5d-7 observables suffice to uniquely determine any pure state. Thus there is a gap between the results for (1) and (2), and we give some examples to illustrate this. The case of observables corresponding to reduced density matrices (RDMs) of a multipartite system is also discussed, where we improve known bounds on local dimensions for case (2) in which almost all pure states are uniquely determined by their RDMs. We further discuss circumstances where (1) can imply (2). We use convexity of the numerical range of operators to show that when only two observables are measured, (1) always implies (2). More generally, if there is a compact group of symmetries of the state space which has the span of the observables measured as the set of fixed points, then (1) implies (2). We analyze the possible dimensions for the span of such observables. Our results extend naturally to the case of low rank quantum states.

quant-ph↗