Searcharxiv⌕ Search

arXiv subjects

Samuel Dooley

Publications and source records attributed to Samuel Dooley.

At least 19 recordsLinked to original sources
arXiv2026

Embedding by Elicitation: Dynamic Representations for Bayesian Optimization of System Prompts

System prompts are a central control mechanism in modern AI systems, shaping behavior across conversations, tasks, and user populations. Yet they are difficult to tune when feedback is available only as aggregate metrics rather than per-example labels, failures, or critiques. We study this aggregate feedback setting as sample-constrained black-box optimization over discrete, variable-length text. We introduce ReElicit, a Bayesian optimization framework based on \emph{embedding by elicitation}. Given a task description, previously evaluated prompts, and scalar scores, an LLM elicits a compact, interpretable feature space and maps prompts into it. Leveraging a probabilistic Gaussian process surrogate, an acquisition function then selects target feature vectors, which the LLM realizes and refines into deployable system prompts. Re-eliciting the feature space as new evaluations arrive lets the representation adapt to the observed prompt-score history. We evaluate the setting using offline benchmark accuracy as a controlled aggregate proxy: the optimizer observes one scalar score per prompt and no per-example labels, errors, or critiques. Across ten system prompt optimization tasks with a 30 total evaluation budget, ReElicit achieves the strongest aggregate performance profile among representative aggregate-only prompt-optimization baselines. These results suggest that LLMs can serve as adaptive semantic representation builders, not only prompt generators, for Bayesian optimization over natural-language artifacts.

cs.AI↗
arXiv2025

Transformers Boost the Performance of Decision Trees on Tabular Data across Sample Sizes

Large language models (LLMs) perform remarkably well on tabular datasets in zero- and few-shot settings, since they can extract meaning from natural language column headers that describe features and labels. Similarly, TabPFN, a recent non-LLM transformer pretrained on numerous tables for in-context learning, has demonstrated excellent performance for dataset sizes up to a thousand samples. In contrast, gradient-boosted decision trees (GBDTs) are typically trained from scratch on each dataset without benefiting from pretraining data and must learn the relationships between columns from their entries alone since they lack natural language understanding. LLMs and TabPFN excel on small tabular datasets where a strong prior is essential, yet they are not competitive with GBDTs on medium or large datasets, since their context lengths are limited. In this paper, we propose a simple and lightweight approach for fusing large language models and TabPFN with gradient-boosted decision trees, which allows scalable GBDTs to benefit from the natural language capabilities and pretraining of transformers. We name our fusion methods LLM-Boost and PFN-Boost, respectively. While matching or surpassing the performance of the transformer at sufficiently small dataset sizes and GBDTs at sufficiently large sizes, LLM-Boost and PFN-Boost outperform both standalone components on a wide range of dataset sizes in between. We demonstrate state-of-the-art performance against numerous baselines and ensembling algorithms. We find that PFN-Boost achieves the best average performance among all methods we test for all but very small dataset sizes. We release our code at http://github.com/MayukaJ/LLM-Boost .

cs.CL↗
arXiv2024

Style Outweighs Substance: Failure Modes of LLM Judges in Alignment Benchmarking

The release of ChatGPT in November 2022 sparked an explosion of interest in post-training and an avalanche of new preference optimization (PO) methods. These methods claim superior alignment by virtue of better correspondence with human pairwise preferences, often measured by LLM-judges. In this work, we attempt to answer the following question -- do LLM-judge preferences translate to progress on other, more concrete metrics for alignment, and if not, why not? We define a concrete metric for alignment, and introduce SOS-Bench (Substance Outweighs Style Benchmark), which is to the best of our knowledge the largest standardized, reproducible LLM meta-benchmark to date. We find that (1) LLM-judge preferences do not correlate with concrete measures of safety, world knowledge, and instruction following; (2) LLM-judges have powerful implicit biases, prioritizing style over factuality and safety; and (3) the supervised fine-tuning (SFT) stage of post-training, and not the PO stage, has the greatest impact on alignment, with data scaling and prompt diversity as the driving factors. Our codebase and complete results can be found at https://github.com/penfever/sos-bench.

cs.LG↗
arXiv2024

LiveBench: A Challenging, Contamination-Limited LLM Benchmark

Test set contamination, wherein test data from a benchmark ends up in a newer model's training set, is a well-documented obstacle for fair LLM evaluation and can quickly render benchmarks obsolete. To mitigate this, many recent benchmarks crowdsource new prompts and evaluations from human or LLM judges; however, these can introduce significant biases, and break down when scoring hard questions. In this work, we introduce a new benchmark for LLMs designed to be resistant to both test set contamination and the pitfalls of LLM judging and human crowdsourcing. We release LiveBench, the first benchmark that (1) contains frequently-updated questions from recent information sources, (2) scores answers automatically according to objective ground-truth values, and (3) contains a wide variety of challenging tasks, spanning math, coding, reasoning, language, instruction following, and data analysis. To achieve this, LiveBench contains questions that are based on recently-released math competitions, arXiv papers, news articles, and datasets, and it contains harder, contamination-limited versions of tasks from previous benchmarks such as Big-Bench Hard, AMPS, and IFEval. We evaluate many prominent closed-source models, as well as dozens of open-source models ranging from 0.5B to 405B in size. LiveBench is difficult, with top models achieving below 70% accuracy. We release all questions, code, and model answers. Questions are added and updated on a monthly basis, and we release new tasks and harder versions of tasks over time so that LiveBench can distinguish between the capabilities of LLMs as they improve in the future. We welcome community engagement and collaboration for expanding the benchmark tasks and models.

cs.CL↗
arXiv2024

Large Language Models Must Be Taught to Know What They Don't Know

When using large language models (LLMs) in high-stakes applications, we need to know when we can trust their predictions. Some works argue that prompting high-performance LLMs is sufficient to produce calibrated uncertainties, while others introduce sampling methods that can be prohibitively expensive. In this work, we first argue that prompting on its own is insufficient to achieve good calibration and then show that fine-tuning on a small dataset of correct and incorrect answers can create an uncertainty estimate with good generalization and small computational overhead. We show that a thousand graded examples are sufficient to outperform baseline methods and that training through the features of a model is necessary for good performance and tractable for large open-source models when using LoRA. We also investigate the mechanisms that enable reliable LLM uncertainty estimation, finding that many models can be used as general-purpose uncertainty estimators, applicable not just to their own uncertainties but also the uncertainty of other models. Lastly, we show that uncertainty estimates inform human use of LLMs in human-AI collaborative settings through a user study.

cs.LG↗
arXiv2024

Multi-objective Differentiable Neural Architecture Search

Pareto front profiling in multi-objective optimization (MOO), i.e., finding a diverse set of Pareto optimal solutions, is challenging, especially with expensive objectives that require training a neural network. Typically, in MOO for neural architecture search (NAS), we aim to balance performance and hardware metrics across devices. Prior NAS approaches simplify this task by incorporating hardware constraints into the objective function, but profiling the Pareto front necessitates a computationally expensive search for each constraint. In this work, we propose a novel NAS algorithm that encodes user preferences to trade-off performance and hardware metrics, yielding representative and diverse architectures across multiple devices in just a single search run. To this end, we parameterize the joint architectural distribution across devices and multiple objectives via a hypernetwork that can be conditioned on hardware features and preference vectors, enabling zero-shot transferability to new devices. Extensive experiments involving up to 19 hardware devices and 3 different objectives demonstrate the effectiveness and scalability of our method. Finally, we show that, without any additional costs, our method outperforms existing MOO NAS methods across a broad range of qualitatively different search spaces and datasets, including MobileNetV3 on ImageNet-1k, an encoder-decoder transformer space for machine translation and a decoder-only space for language modelling.

cs.LG↗
arXiv2024

Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive

Direct Preference Optimisation (DPO) is effective at significantly improving the performance of large language models (LLMs) on downstream tasks such as reasoning, summarisation, and alignment. Using pairs of preferred and dispreferred data, DPO models the relative probability of picking one response over another. In this work, first we show theoretically that the standard DPO loss can lead to a reduction of the model's likelihood of the preferred examples, as long as the relative probability between the preferred and dispreferred classes increases. We then show empirically that this phenomenon occurs when fine-tuning LLMs on common datasets, especially datasets in which the edit distance between pairs of completions is low. Using these insights, we design DPO-Positive (DPOP), a new loss function and training procedure which avoids this failure mode. Surprisingly, we find that DPOP outperforms DPO and other fine-tuning procedures across a wide variety of datasets and downstream tasks, including datasets with high edit distances between completions. Furthermore, we find that the DPOP-tuned model outperforms the DPO-tuned model (all else equal) on benchmarks independent of the fine-tuning data, such as MT-Bench. Finally, using DPOP, we create and open-source Smaug-34B and Smaug-72B, with the latter becoming the first open-source LLM to surpass an average accuracy of 80% on the HuggingFace Open LLM Leaderboard.

cs.CL↗
arXiv2023

ForecastPFN: Synthetically-Trained Zero-Shot Forecasting

The vast majority of time-series forecasting approaches require a substantial training dataset. However, many real-life forecasting applications have very little initial observations, sometimes just 40 or fewer. Thus, the applicability of most forecasting methods is restricted in data-sparse commercial applications. While there is recent work in the setting of very limited initial data (so-called `zero-shot' forecasting), its performance is inconsistent depending on the data used for pretraining. In this work, we take a different approach and devise ForecastPFN, the first zero-shot forecasting model trained purely on a novel synthetic data distribution. ForecastPFN is a prior-data fitted network, trained to approximate Bayesian inference, which can make predictions on a new time series dataset in a single forward pass. Through extensive experiments, we show that zero-shot predictions made by ForecastPFN are more accurate and faster compared to state-of-the-art forecasting methods, even when the other methods are allowed to train on hundreds of additional in-distribution data points.

cs.LG↗
arXiv2023

Fairer and More Accurate Tabular Models Through NAS

Making models algorithmically fairer in tabular data has been long studied, with techniques typically oriented towards fixes which usually take a neural model with an undesirable outcome and make changes to how the data are ingested, what the model weights are, or how outputs are processed. We employ an emergent and different strategy where we consider updating the model's architecture and training hyperparameters to find an entirely new model with better outcomes from the beginning of the debiasing procedure. In this work, we propose using multi-objective Neural Architecture Search (NAS) and Hyperparameter Optimization (HPO) in the first application to the very challenging domain of tabular data. We conduct extensive exploration of architectural and hyperparameter spaces (MLP, ResNet, and FT-Transformer) across diverse datasets, demonstrating the dependence of accuracy and fairness metrics of model predictions on hyperparameter combinations. We show that models optimized solely for accuracy with NAS often fail to inherently address fairness concerns. We propose a novel approach that jointly optimizes architectural and training hyperparameters in a multi-objective constraint of both accuracy and fairness. We produce architectures that consistently Pareto dominate state-of-the-art bias mitigation methods either in fairness, accuracy or both, all of this while being Pareto-optimal over hyperparameters achieved through single-objective (accuracy) optimization runs. This research underscores the promise of automating fairness and accuracy optimization in deep learning models.

cs.LG↗
arXiv2023

Data Contamination Through the Lens of Time

Recent claims about the impressive abilities of large language models (LLMs) are often supported by evaluating publicly available benchmarks. Since LLMs train on wide swaths of the internet, this practice raises concerns of data contamination, i.e., evaluating on examples that are explicitly or implicitly included in the training data. Data contamination remains notoriously challenging to measure and mitigate, even with partial attempts like controlled experimentation of training data, canary strings, or embedding similarities. In this work, we conduct the first thorough longitudinal analysis of data contamination in LLMs by using the natural experiment of training cutoffs in GPT models to look at benchmarks released over time. Specifically, we consider two code/mathematical problem-solving datasets, Codeforces and Project Euler, and find statistically significant trends among LLM pass rate vs. GitHub popularity and release date that provide strong evidence of contamination. By open-sourcing our dataset, raw results, and evaluation framework, our work paves the way for rigorous analyses of data contamination in modern models. We conclude with a discussion of best practices and future steps for publicly releasing benchmarks in the age of LLMs that train on webscale data.

cs.CL↗
arXiv2023

Giraffe: Adventures in Expanding Context Lengths in LLMs

Modern large language models (LLMs) that rely on attention mechanisms are typically trained with fixed context lengths which enforce upper limits on the length of input sequences that they can handle at evaluation time. To use these models on sequences longer than the train-time context length, one might employ techniques from the growing family of context length extrapolation methods -- most of which focus on modifying the system of positional encodings used in the attention mechanism to indicate where tokens or activations are located in the input sequence. We conduct a wide survey of existing methods of context length extrapolation on a base LLaMA or LLaMA 2 model, and introduce some of our own design as well -- in particular, a new truncation strategy for modifying the basis for the position encoding. We test these methods using three new evaluation tasks (FreeFormQA, AlteredNumericQA, and LongChat-Lines) as well as perplexity, which we find to be less fine-grained as a measure of long context performance of LLMs. We release the three tasks publicly as datasets on HuggingFace. We discover that linear scaling is the best method for extending context length, and show that further gains can be achieved by using longer scales at evaluation time. We also discover promising extrapolation capabilities in the truncated basis. To support further research in this area, we release three new 13B parameter long-context models which we call Giraffe: 4k and 16k context models trained from base LLaMA-13B, and a 32k context model trained from base LLaMA2-13B. We also release the code to replicate our results.

cs.AI↗
arXiv2022

Robustness Disparities in Face Detection

Facial analysis systems have been deployed by large companies and critiqued by scholars and activists for the past decade. Many existing algorithmic audits examine the performance of these systems on later stage elements of facial analysis systems like facial recognition and age, emotion, or perceived gender prediction; however, a core component to these systems has been vastly understudied from a fairness perspective: face detection, sometimes called face localization. Since face detection is a pre-requisite step in facial analysis systems, the bias we observe in face detection will flow downstream to the other components like facial recognition and emotion prediction. Additionally, no prior work has focused on the robustness of these systems under various perturbations and corruptions, which leaves open the question of how various people are impacted by these phenomena. We present the first of its kind detailed benchmark of face detection systems, specifically examining the robustness to noise of commercial and academic models. We use both standard and recently released academic facial datasets to quantitatively analyze trends in face detection robustness. Across all the datasets and systems, we generally find that photos of individuals who are $\textit{masculine presenting}$, $\textit{older}$, of $\textit{darker skin type}$, or have $\textit{dim lighting}$ are more susceptible to errors than their counterparts in other identities.

cs.CY↗
arXiv2022

How Technology Impacts and Compares to Humans in Socially Consequential Arenas

One of the main promises of technology development is for it to be adopted by people, organizations, societies, and governments -- incorporated into their life, work stream, or processes. Often, this is socially beneficial as it automates mundane tasks, frees up more time for other more important things, or otherwise improves the lives of those who use the technology. However, these beneficial results do not apply in every scenario and may not impact everyone in a system the same way. Sometimes a technology is developed which produces both benefits and inflicts some harm. These harms may come at a higher cost to some people than others, raising the question: {\it how are benefits and harms weighed when deciding if and how a socially consequential technology gets developed?} The most natural way to answer this question, and in fact how people first approach it, is to compare the new technology to what used to exist. As such, in this work, I make comparative analyses between humans and machines in three scenarios and seek to understand how sentiment about a technology, performance of that technology, and the impacts of that technology combine to influence how one decides to answer my main research question.

cs.AI↗
arXiv2022

Rethinking Bias Mitigation: Fairer Architectures Make for Fairer Face Recognition

Face recognition systems are widely deployed in safety-critical applications, including law enforcement, yet they exhibit bias across a range of socio-demographic dimensions, such as gender and race. Conventional wisdom dictates that model biases arise from biased training data. As a consequence, previous works on bias mitigation largely focused on pre-processing the training data, adding penalties to prevent bias from effecting the model during training, or post-processing predictions to debias them, yet these approaches have shown limited success on hard problems such as face recognition. In our work, we discover that biases are actually inherent to neural network architectures themselves. Following this reframing, we conduct the first neural architecture search for fairness, jointly with a search for hyperparameters. Our search outputs a suite of models which Pareto-dominate all other high-performance architectures and existing bias mitigation methods in terms of accuracy and fairness, often by large margins, on the two most widely used datasets for face identification, CelebA and VGGFace2. Furthermore, these models generalize to other datasets and sensitive attributes. We release our code, models and raw data files at https://github.com/dooleys/FR-NAS.

cs.CV↗
arXiv2022

A Deep Dive into Dataset Imbalance and Bias in Face Identification

As the deployment of automated face recognition (FR) systems proliferates, bias in these systems is not just an academic question, but a matter of public concern. Media portrayals often center imbalance as the main source of bias, i.e., that FR models perform worse on images of non-white people or women because these demographic groups are underrepresented in training data. Recent academic research paints a more nuanced picture of this relationship. However, previous studies of data imbalance in FR have focused exclusively on the face verification setting, while the face identification setting has been largely ignored, despite being deployed in sensitive applications such as law enforcement. This is an unfortunate omission, as 'imbalance' is a more complex matter in identification; imbalance may arise in not only the training data, but also the testing data, and furthermore may affect the proportion of identities belonging to each demographic group or the number of images belonging to each identity. In this work, we address this gap in the research by thoroughly exploring the effects of each kind of imbalance possible in face identification, and discuss other factors which may impact bias in this setting.

cs.CV↗
arXiv2022

Topological Data Analysis for Word Sense Disambiguation

We develop and test a novel unsupervised algorithm for word sense induction and disambiguation which uses topological data analysis. Typical approaches to the problem involve clustering, based on simple low level features of distance in word embeddings. Our approach relies on advanced mathematical concepts in the field of topology which provides a richer conceptualization of clusters for the word sense induction tasks. We use a persistent homology barcode algorithm on the SemCor dataset and demonstrate that our approach gives low relative error on word sense induction. This shows the promise of topological algorithms for natural language processing and we advocate for future work in this promising area.

cs.CL↗
arXiv2022

The Dichotomous Affiliate Stable Matching Problem: Approval-Based Matching with Applicant-Employer Relations

While the stable marriage problem and its variants model a vast range of matching markets, they fail to capture complex agent relationships, such as the affiliation of applicants and employers in an interview marketplace. To model this problem, the existing literature on matching with externalities permits agents to provide complete and total rankings over matchings based off of both their own and their affiliates' matches. This complete ordering restriction is unrealistic, and further the model may have an empty core. To address this, we introduce the Dichotomous Affiliate Stable Matching (DASM) Problem, where agents' preferences indicate dichotomous acceptance or rejection of another agent in the marketplace, both for themselves and their affiliates. We also assume the agent's preferences over entire matchings are determined by a general weighted valuation function of their (and their affiliates') matches. Our results are threefold: (1) we use a human study to show that real-world matching rankings follow our assumed valuation function; (2) we prove that there always exists a stable solution by providing an efficient, easily-implementable algorithm that finds such a solution; and (3) we experimentally validate the efficiency of our algorithm versus a linear-programming-based approach.

cs.GT↗
arXiv2022

Are Commercial Face Detection Models as Biased as Academic Models?

As facial recognition systems are deployed more widely, scholars and activists have studied their biases and harms. Audits are commonly used to accomplish this and compare the algorithmic facial recognition systems' performance against datasets with various metadata labels about the subjects of the images. Seminal works have found discrepancies in performance by gender expression, age, perceived race, skin type, etc. These studies and audits often examine algorithms which fall into two categories: academic models or commercial models. We present a detailed comparison between academic and commercial face detection systems, specifically examining robustness to noise. We find that state-of-the-art academic face detection models exhibit demographic disparities in their noise robustness, specifically by having statistically significant decreased performance on older individuals and those who present their gender in a masculine manner. When we compare the size of these disparities to that of commercial models, we conclude that commercial models - in contrast to their relatively larger development budget and industry-level fairness commitments - are always as biased or more biased than an academic model.

cs.CV↗