publications

2024

1. GoodIT

   Indian-BhED: A Dataset for Measuring India-Centric Biases in Large Language Models

   Khyati Khandelwal, Manuel Tonneau, Andrew M. Bean, Hannah Rose Kirk, and Scott A. Hale

   In Proceedings of the 2024 International Conference on Information Technology for Social Good, Sep 2024

   Abs arXiv Link

   Large Language Models (LLMs), now used daily by millions, can encode societal biases, exposing their users to representational harms. A large body of scholarship on LLM bias exists but it predominantly adopts a Western-centric frame and attends comparatively less to bias levels and potential harms in the Global South. In this paper, we quantify stereotypical bias in popular LLMs according to an Indian-centric frame through Indian-BhED, a first of its kind dataset, containing stereotypical and anti-stereotypical examples in the context of caste and religious stereotypes in India. We find that the majority of LLMs tested have a strong propensity to output stereotypes in the Indian context, especially when compared to axes of bias traditionally studied in the Western context, such as gender and race. Notably, we find that GPT-2, GPT-2 Large, and GPT 3.5 have a particularly high propensity for preferring stereotypical outputs as a percent of all sentences for the axes of caste (63–79%) and religion (69–72%). We finally investigate potential causes for such harmful behaviour in LLMs, and posit intervention techniques to reduce both stereotypical and anti-stereotypical biases. The findings of this work highlight the need for including more diverse voices when researching fairness in AI and evaluating LLMs.

2. NeurIPSWorkshop

   Do Large Language Models have Shared Weaknesses in Medical Question Answering?

   Andrew M. Bean, Karolina Korgul, Felix Krones, Robert McCraith, and Adam Mahdi

   Mar 2024

   Abs arXiv

   With the rapid development of new large language models (LLMs), each claiming to surpass previous models, an overall picture of medical LLM research can be elusive. To address this challenge, we benchmark a range of top LLMs and identify consistent patterns which appear across models. We test 8 well-known LLMs on 874 newly collected questions from Polish medical licensing exams. For each question, we score each model on the top-1 accuracy and the distribution of probabilities assigned. We then compare with factors including question difficulty for humans, question length, and the scores of the other models. LLM accuracies were positively correlated pairwise (0.29 to 0.62). Model performance was also correlated with human performance (0.07 to 0.16), but negatively correlated to the difference between the question-level accuracy of top-scoring and bottom-scoring humans (−0.16 to −0.23). The top output probability and question length were positive and negative predictors of accuracy respectively (p <0.05). The top scoring LLM, GPT-4 Turbo, scored 82%, followed by Med42, PaLM 2, Mixtral and GPT-3.5 around 63%. We found evidence of similarities between models in which questions they answer correctly, as well as similarities with human test takers. Larger models typically performed better, but differences in training methods were also highly impactful. Model accuracy was positively correlated with confidence, but negatively correlated with question length. We expect that similar training methods will lead these patterns to persist across future models. These patterns can therefore aid medical experts in forming expectations about LLMs as a category to support application research.

3. NeurIPSOral

   The PRISM Alignment Project: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models

   Hannah Rose Kirk, Alexander Whitefield, Paul Röttger, Andrew M. Bean, Katerina Margatina, Juan Ciro, Rafael Mosquera, Max Bartolo, Adina Williams, He He, Bertie Vidgen, and Scott A. Hale

   Apr 2024

   Abs arXiv

   Human feedback plays a central role in the alignment of Large Language Models (LLMs). However, open questions remain about the methods (how), domains (where), people (who) and objectives (to what end) of human feedback collection. To navigate these questions, we introduce PRISM, a new dataset which maps the sociodemographics and stated preferences of 1,500 diverse participants from 75 countries, to their contextual preferences and fine-grained feedback in 8,011 live conversations with 21 LLMs. PRISM contributes (i) wide geographic and demographic participation in human feedback data; (ii) two census-representative samples for understanding collective welfare (UK and US); and (iii) individualised feedback where every rating is linked to a detailed participant profile, thus permitting exploration of personalisation and attribution of sample artefacts. We focus on collecting conversations that centre subjective and multicultural perspectives on value-laden and controversial topics, where we expect the most interpersonal and cross-cultural disagreement. We demonstrate the usefulness of PRISM via three case studies of dialogue diversity, preference diversity, and welfare outcomes, showing that it matters which humans set alignment norms. As well as offering a rich community resource, we advocate for broader participation in AI development and a more inclusive approach to technology design.

4. NeurIPSOral

   LINGOLY: A Benchmark of Olympiad-Level Linguistic Reasoning Puzzles in Low-Resource and Extinct Languages

   Andrew M. Bean, Simi Hellsten, Harry Mayne, Jabez Magomere, Ethan A. Chi, Ryan Chi, Scott A. Hale, and Hannah Rose Kirk

   Jun 2024

   Abs arXiv Code

   In this paper, we present the LingOly benchmark, a novel benchmark for advanced reasoning abilities in large language models. Using challenging Linguistic Olympiad puzzles, we evaluate (i) capabilities for in-context identification and generalisation of linguistic patterns in very low-resource or extinct languages, and (ii) abilities to follow complex task instructions. The LingOly benchmark covers more than 90 mostly low-resource languages, minimising issues of data contamination, and contains 1,133 problems across 6 formats and 5 levels of human difficulty. We assess performance with both direct accuracy and comparison to a no-context baseline to penalise memorisation. Scores from 11 state-of-the-art LLMs demonstrate the benchmark to be challenging, and models perform poorly on the higher difficulty problems. On harder problems, even the top model only achieved 38.7% accuracy, 24.7% improvement over the no-context baseline. Large closed models typically outperform open models, and in general, the higher resource the language, the better the scores. These results indicate, in absence of memorisation, true multi-step out-of-domain reasoning remains a challenge for current language models.

5. NeurIPSWorkshop

   Fine-Tuning Large Language Models with Human-inspired Learning Strategies in Medical Question Answering

   Yushi Yang, Andrew M. Bean, Robert McCraith, and Adam Mahdi

   Aug 2024

   Abs arXiv

   Training Large Language Models (LLMs) incurs substantial data-related costs, motivating the development of data-efficient training methods through optimised data ordering and selection. Human-inspired learning strategies, such as curriculum learning, offer possibilities for efficient training by organising data according to common human learning practices. Despite evidence that fine-tuning with curriculum learning improves the performance of LLMs for natural language understanding tasks, its effectiveness is typically assessed using a single model. In this work, we extend previous research by evaluating both curriculum-based and non-curriculum-based learning strategies across multiple LLMs, using human-defined and automated data labels for medical question answering. Our results indicate a moderate impact of using human-inspired learning strategies for fine-tuning LLMs, with maximum accuracy gains of 1.77% per model and 1.81% per dataset. Crucially, we demonstrate that the effectiveness of these strategies varies significantly across different model-dataset combinations, emphasising that the benefits of a specific human-inspired strategy for fine-tuning LLMs do not generalise. Additionally, we find evidence that curriculum learning using LLM-defined question difficulty outperforms human-defined difficulty, highlighting the potential of using model-generated measures for optimal curriculum design.

2023

1. House of Lords

   Large Language Models - Written Evidence

   Andrew M. Bean, Hannah Rose Kirk, Jakob Mökander, Cailean Osborne, Huw Roberts, and Marta Ziosi

   Oct 2023

   Link
2. EMNLP

   The Past, Present and Better Future of Feedback Learning in Large Language Models for Subjective Human Preferences and Values

   Hannah Kirk, Andrew Bean, Bertie Vidgen, Paul Rottger, and Scott Hale

   In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Dec 2023

   Abs arXiv Link

   Human feedback is increasingly used to steer the behaviours of Large Language Models (LLMs). However, it is unclear how to collect and incorporate feedback in a way that is efficient, effective and unbiased, especially for highly subjective human preferences and values. In this paper, we survey existing approaches for learning from human feedback, drawing on 95 papers primarily from the ACL and arXiv repositories. First, we summarise the past, pre-LLM trends for integrating human feedback into language models. Second, we give an overview of present techniques and practices, as well as the motivations for using feedback; conceptual frameworks for defining values and preferences; and how feedback is collected and from whom. Finally, we encourage a better future of feedback learning in LLMs by raising five unresolved conceptual and practical challenges.