Adrian de Wynter

Rishav Hada, Varun Gumma, Adrian de Wynter et al. · microsoft-research

Large Language Models (LLMs) excel in various Natural Language Processing (NLP) tasks, yet their evaluation, particularly in languages beyond the top $20$, remains inadequate due to existing benchmarks and metrics limitations. Employing LLMs as evaluators to rank or score other models' outputs emerges as a viable solution, addressing the constraints tied to human annotators and established benchmarks. In this study, we explore the potential of LLM-based evaluators, specifically GPT-4 in enhancing multilingual evaluation by calibrating them against $20$K human judgments across three text-generation tasks, five metrics, and eight languages. Our analysis reveals a bias in GPT4-based evaluators towards higher scores, underscoring the necessity of calibration with native speaker judgments, especially in low-resource and non-Latin script languages, to ensure accurate evaluation of LLM performance across diverse languages.

Dongjun Kim, Adrian de Wynter, Huancheng Chen et al.

While finetuning effectively adapts foundation models to specialized downstream tasks, it can degrade nontarget capabilities acquired during pretraining. Existing forgetting aware methods typically seek safer updates through specialized initialization or fixed constraints, but do not regulate the adaptation preservation trade-off during training. We propose Foundation Preserving LoRA (FoLoRA), a forgetting aware optimization framework. Guided by a first order preservation condition, FoLoRA defines a forgetting penalty over pretraining-proxy activations and a task utility over downstream task activations. It then scores update directions by task utility per unit forgetting penalty via a generalized Rayleigh quotient. The resulting spectral coordinate system enables direction wise gated Adam updates, attenuating low utility to penalty directions during training. To estimate the forgetting penalty, FoLoRA constructs pretraining proxy calibration data by sampling from the pretrained model rather than relying on a single proxy dataset. Experiments on math, code, and instruction following adaptation show that FoLoRA achieves the strongest preservation adaptation balance over baselines, improving target task performance with best aggregate preservation of non target capabilities.

Adrian de Wynter, Xun Wang, Alex Sokolov et al. · microsoft-research

We present an empirical evaluation of various outputs generated by nine of the most widely-available large language models (LLMs). Our analysis is done with off-the-shelf, readily-available tools. We find a correlation between percentage of memorized text, percentage of unique text, and overall output quality, when measured with respect to output pathologies such as counterfactual and logically-flawed statements, and general failures like not staying on topic. Overall, 80.0% of the outputs evaluated contained memorized data, but outputs containing the most memorized content were also more likely to be considered of high quality. We discuss and evaluate mitigation strategies, showing that, in the models evaluated, the rate of memorized text being output is reduced. We conclude with a discussion on potential implications around what it means to learn, to memorize, and to evaluate quality text.

Adrian de Wynter

Much research has been carried out on large language models (LLMs) and LLM-powered agentic workflows. However, many works within the field state emergence of, ascribe to, or assume, generalised anthropomorphic attributes to them (e.g., morality or understanding of natural language). Our goal is not to argue in favour or against the existence of these attributes, but to point out that these conclusions could be incorrect. For this we build and train a simple neural network on the videogame Age of Empires II, and note that any entity in a sufficiently-powerful substrate, such as LEGO or the Greater Boston Area, could also present such attributes. Hence, the purported anthropomorphic attributes of LLMs are empirically non-unique: although some properties (e.g., responses to prompts) could remain constant, others, such as the interpretation of their perceived behaviour, might change with the substrate. Thus, any empirically-grounded discussion requires explicit measurement criteria; otherwise the interpretation is left to the representation. We then show that assuming that these attributes exist or not in a system, independent of the substrate and in a generalised way, leads to either circular or uninformative conclusions, regardless of the experimenter's viewpoint on the subject. Finally we propose a 'null' assumption, where one assumes LLM non-uniqueness instead of assuming anthropomorphic attributes to set up an experiment, along with examples of it. We also discuss potential objections to our work, briefly survey the field, and prove that \textit{Age of Empires II} is functionally- and Turing-complete.

Amartya Roy, Elamparithy M, Kripabandhu Ghosh et al.

In context learning (ICL) underpins recent advances in large language models (LLMs), although its role and performance in causal reasoning remains unclear. Causal reasoning demands multihop composition and strict conjunctive control, and reliance on spurious lexical relations of the input could provide misleading results. We hypothesize that, due to their ability to project the input into a latent space, encoder and encoder decoder architectures are better suited for said multihop conjunctive reasoning versus decoder only models. To do this, we compare fine-tuned versions of all the aforementioned architectures with zero and few shot ICL in both natural language and non natural language scenarios. We find that ICL alone is insufficient for reliable causal reasoning, often overfocusing on irrelevant input features. In particular, decoder only models are noticeably brittle to distributional shifts, while finetuned encoder and encoder decoder models can generalize more robustly across our tests, including the non natural language split. Both architectures are only matched or surpassed by decoder only architectures at large scales. We conclude by noting that for cost effective, short horizon robust causal reasoning, encoder or encoder decoder architectures with targeted finetuning are preferable.

Adrian de Wynter, Tangming Yuan

We evaluate two large language models (LLMs) ability to perform argumentative reasoning. We experiment with argument mining (AM) and argument pair extraction (APE), and evaluate the LLMs' ability to recognize arguments under progressively more abstract input and output (I/O) representations (e.g., arbitrary label sets, graphs, etc.). Unlike the well-known evaluation of prompt phrasings, abstraction evaluation retains the prompt's phrasing but tests reasoning capabilities. We find that scoring-wise the LLMs match or surpass the SOTA in AM and APE, and under certain I/O abstractions LLMs perform well, even beating chain-of-thought--we call this symbolic prompting. However, statistical analysis on the LLMs outputs when subject to small, yet still human-readable, alterations in the I/O representations (e.g., asking for BIO tags as opposed to line numbers) showed that the models are not performing reasoning. This suggests that LLM applications to some tasks, such as data labelling and paper reviewing, must be done with care.

Adrian de Wynter

We perform a critical examination of the scientific methodology behind contemporary large language model (LLM) research. For this we assess over 2,000 research works released between 2020 and 2024 based on criteria typical of what is considered good research (e.g. presence of statistical tests and reproducibility), and cross-validate it with arguments that are at the centre of controversy (e.g., claims of emergent behaviour). We find multiple trends, such as declines in ethics disclaimers, a rise of LLMs as evaluators, and an increase on claims of LLM reasoning abilities without leveraging human evaluation. We note that conference checklists are effective at curtailing some of these issues, but balancing velocity and rigour in research cannot solely rely on these. We tie all these findings to findings from recent meta-reviews and extend recommendations on how to address what does, does not, and should work in LLM research.

Wei Li, Adrian de Wynter

It is known that large language models (LLMs) underperform in English dialects, and that improving them is difficult due to data scarcity. In this work we investigate how quality and availability impact the feasibility of improving LLMs in this context. For this, we evaluate three rarely-studied English dialects (Yorkshire, Geordie, and Cornish), plus African-American Vernacular English, and West Frisian as control. We find that human-human agreement when determining LLM generation quality directly impacts LLM-as-a-judge performance. That is, LLM-human agreement mimics the human-human agreement pattern, and so do metrics such as accuracy. It is an issue because LLM-human agreement measures an LLM's alignment with the human consensus; and hence raises questions about the feasibility of improving LLM performance in locales where low populations induce low agreement. We also note that fine-tuning does not eradicate, and might amplify, this pattern in English dialects. But also find encouraging signals, such as some LLMs' ability to generate high-quality data, thus enabling scalability. We argue that data must be carefully evaluated to ensure fair and inclusive LLM improvement; and, in the presence of scarcity, new tools are needed to handle the pattern found.

Adrian de Wynter, Anthony Hevia, Si-Qing Chen

We present an evaluation of text simplification (TS) in Spanish for a production system, by means of two corpora focused in both complex-sentence and complex-word identification. We compare the most prevalent Spanish-specific readability scores with neural networks, and show that the latter are consistently better at predicting user preferences regarding TS. As part of our analysis, we find that multilingual models underperform against equivalent Spanish-only models on the same task, yet all models focus too often on spurious statistical features, such as sentence length. We release the corpora in our evaluation to the broader community with the hopes of pushing forward the state-of-the-art in Spanish natural language processing.

Yadong Zhang, Shaoguang Mao, Tao Ge et al.

This paper presents a comprehensive survey of the current status and opportunities for Large Language Models (LLMs) in strategic reasoning, a sophisticated form of reasoning that necessitates understanding and predicting adversary actions in multi-agent settings while adjusting strategies accordingly. Strategic reasoning is distinguished by its focus on the dynamic and uncertain nature of interactions among multi-agents, where comprehending the environment and anticipating the behavior of others is crucial. We explore the scopes, applications, methodologies, and evaluation metrics related to strategic reasoning with LLMs, highlighting the burgeoning development in this area and the interdisciplinary approaches enhancing their decision-making performance. It aims to systematize and clarify the scattered literature on this subject, providing a systematic review that underscores the importance of strategic reasoning as a critical cognitive capability and offers insights into future research directions and potential improvements.

Adrian de Wynter, Ishaan Watts, Tua Wongsangaroonsri et al. · cmu, deepmind

Large language models (LLMs) and small language models (SLMs) are being adopted at remarkable speed, although their safety still remains a serious concern. With the advent of multilingual S/LLMs, the question now becomes a matter of scale: can we expand multilingual safety evaluations of these models with the same velocity at which they are deployed? To this end, we introduce RTP-LX, a human-transcreated and human-annotated corpus of toxic prompts and outputs in 28 languages. RTP-LX follows participatory design practices, and a portion of the corpus is especially designed to detect culturally-specific toxic language. We evaluate 10 S/LLMs on their ability to detect toxic content in a culturally-sensitive, multilingual scenario. We find that, although they typically score acceptably in terms of accuracy, they have low agreement with human judges when scoring holistically the toxicity of a prompt; and have difficulty discerning harm in context-dependent scenarios, particularly with subtle-yet-harmful content (e.g. microaggressions, bias). We release this dataset to contribute to further reduce harmful uses of these models and improve their safe deployment.

Adrian de Wynter

We show that GPT-4's reasoning and planning capabilities extend to the 1993 first-person shooter Doom. This large language model (LLM) is able to run and play the game with only a few instructions, plus a textual description--generated by the model itself from screenshots--about the state of the game being observed. We find that GPT-4 can play the game to a passable degree: it is able to manipulate doors, combat enemies, and perform pathing. More complex prompting strategies involving multiple model calls provide better results. While further work is required to enable the LLM to play the game as well as its classical, reinforcement learning-based counterparts, we note that GPT-4 required no training, leaning instead on its own reasoning and observational capabilities. We hope our work pushes the boundaries on intelligent, LLM-based agents in video games. We conclude by discussing the ethical implications of our work.

Adrian de Wynter, Xun Wang, Qilong Gu et al.

Modern large language models (LLMs) are capable of interpreting input strings as instructions, or prompts, and carry out tasks based on them. Unlike traditional learners, LLMs cannot use back-propagation to obtain feedback, and condition their output in situ in a phenomenon known as in-context learning (ICL). Many approaches to prompting and pre-training these models involve the automated generation of these prompts, also known as meta-prompting, or prompting to obtain prompts. However, they do not formally describe the properties and behavior of the LLMs themselves. We propose a theoretical framework based on category theory to generalize and describe ICL and LLM behavior when interacting with users. Our framework allows us to obtain formal results around task agnosticity and equivalence of various meta-prompting approaches. Using our framework and experimental results we argue that meta-prompting is more effective than basic prompting at generating desirable outputs.

Fangru Lin, Shaoguang Mao, Emanuele La Malfa et al. · oxford

Language is not monolithic. While benchmarks, including those designed for multiple languages, are often used as proxies to evaluate the performance of Large Language Models (LLMs), they tend to overlook the nuances of within-language variation and thus fail to model the experience of speakers of non-standard dialects. Focusing on African American Vernacular English (AAVE), we present the first study aimed at objectively assessing the fairness and robustness of LLMs in handling dialects across canonical reasoning tasks, including algorithm, math, logic, and integrated reasoning. We introduce ReDial (Reasoning with Dialect Queries), a benchmark containing 1.2K+ parallel query pairs in Standardized English and AAVE. We hire AAVE speakers, including experts with computer science backgrounds, to rewrite seven popular benchmarks, such as HumanEval and GSM8K. With ReDial, we evaluate widely used LLMs, including GPT, Claude, Llama, Mistral, and the Phi model families. Our findings reveal that almost all of these widely used models show significant brittleness and unfairness to queries in AAVE. Our work establishes a systematic and objective framework for analyzing LLM bias in dialectal queries. Moreover, it highlights how mainstream LLMs provide unfair service to dialect speakers in reasoning tasks, laying a critical foundation for future research.

Adrian de Wynter, Tangming Yuan

Large language models (LLMs) are excellent at maintaining high-level, convincing dialogues. They are being fast deployed as chatbots and evaluators in sensitive areas, such as peer review and mental health applications. This, along with the disparate accounts on their reasoning capabilities, calls for a closer examination of LLMs and their comprehension of dialogue. In this work we begin by evaluating LLMs' ability to maintain a debate--one of the purest yet most complex forms of human communication. Then we measure how this capability relates to their understanding of what is being talked about, namely, their comprehension of dialogical structures and the pragmatic context. We find that LLMs are capable of maintaining coherent, persuasive debates, often swaying the beliefs of participants and audiences alike. We also note that awareness or suspicion of AI involvement encourage people to be more critical of the arguments made. When polling LLMs on their comprehension of deeper structures of dialogue, however, they cannot demonstrate said understanding. Our findings tie the shortcomings of LLMs-as-evaluators to their (in)ability to understand the context. More broadly, for the field of argumentation theory we posit that, if an agent can convincingly maintain a dialogue, it is not necessary for it to know what it is talking about. Hence, the modelling of pragmatic context and coherence are secondary to effectiveness.

Adrian de Wynter

In-context learning (ICL) allows some autoregressive models to solve tasks via next-token prediction and without needing further training. This has led to claims about these model's ability to solve (learn) unseen tasks with only a few shots (exemplars) in the prompt. However, deduction does not always imply learning, as ICL does not explicitly encode a given observation. Instead, the models rely on their prior knowledge and the exemplars given, if any. We argue that, mathematically, ICL does constitute learning, but its full characterisation requires empirical work. We then carry out a large-scale analysis of ICL ablating out or accounting for memorisation, pretraining, distributional shifts, and prompting style and phrasing. We find that ICL is an effective learning paradigm, but limited in its ability to learn and generalise to unseen tasks. We note that, in the limit where exemplars become more numerous, accuracy is insensitive to exemplar distribution, model, prompt style, and the input's linguistic features. Instead, it deduces patterns from regularities in the prompt, which leads to distributional sensitivity, especially in prompting styles such as chain-of-thought. Given the varied accuracies on formally similar tasks, we conclude that autoregression's ad-hoc encoding is not a robust mechanism, and suggests limited all-purpose generalisability.

Adrian de Wynter

Warning: this paper discusses content related, but not limited to, violence, sex, and suicide. Loneliness, or the lack of fulfilling relationships, significantly impacts a person's mental and physical well-being and is prevalent worldwide. Previous research suggests that large language models (LLMs) may help mitigate loneliness. However, we argue that the use of widespread LLMs in services like ChatGPT is more prevalent--and riskier, as they are not designed for this purpose. To explore this, we analysed user interactions with ChatGPT outside of its marketed use as a task-oriented assistant. In dialogues classified as lonely, users frequently (37%) sought advice or validation, and received good engagement. However, ChatGPT failed in sensitive scenarios, like responding appropriately to suicidal ideation or trauma. We also observed a 35% higher incidence of toxic content, with women being 22x more likely to be targeted than men. Our findings underscore ethical and legal questions about this technology, and note risks like radicalisation or further isolation. We conclude with recommendations to research and industry to address loneliness.

Anubhav Jangra, Bahareh Sarrafzadeh, Silviu Cucerzan et al. · microsoft-research

With the surge of large language models (LLMs) and their ability to produce customized output, style-personalized text generation--"write like me"--has become a rapidly growing area of interest. However, style personalization is highly specific, relative to every user, and depends strongly on the pragmatic context, which makes it uniquely challenging. Although prior research has introduced benchmarks and metrics for this area, they tend to be non-standardized and have known limitations (e.g., poor correlation with human subjects). LLMs have been found to not capture author-specific style well, it follows that the metrics themselves must be scrutinized carefully. In this work we critically examine the effectiveness of the most common metrics used in the field, such as BLEU, embeddings, and LLMs-as-judges. We evaluate these metrics using our proposed style discrimination benchmark, which spans eight diverse writing tasks across three evaluation settings: domain discrimination, authorship attribution, and LLM-generated personalized vs non-personalized discrimination. We find strong evidence that employing ensembles of diverse evaluation metrics consistently outperforms single-evaluator methods, and conclude by providing guidance on how to reliably assess style-personalized text generation.

Adrian de Wynter

An evaluator is trustworthy when there exists some agreed-upon way to measure its performance as a labeller. The two ways to establish trustworthiness are either by testing it, or by assuming the evaluator `knows' somehow the way to label the corpus. However, if labelled references (e.g., a development set) are unavailable, neither of these approaches work: the former requires the data, and the latter is an assumption, not evidence. To address this, we introduce an algorithm (the `No-Data Algorithm') by which to establish trust in an evaluator without any existing references. Our algorithm works by successively posing challenges to said evaluator. We show that this is sufficient to establish trustworthiness w.h.p., in such a way that when the evaluator actually knows the way to label the corpus, the No-Data Algorithm accepts its output; and, conversely, flags untrustworthy evaluators when these are unable to prove it. We present formal proofs of correctness, empirical tests, and applications to LLMs-as-judges on low-resource languages.

Sunayana Sitaram, Adrian de Wynter, Isobel McCrum et al.

Misgendering is the act of referring to someone by a gender that does not match their chosen identity. It marginalizes and undermines a person's sense of self, causing significant harm. English-based approaches have clear-cut approaches to avoiding misgendering, such as the use of the pronoun ``they''. However, other languages pose unique challenges due to both grammatical and cultural constructs. In this work we develop methodologies to assess and mitigate misgendering across 42 languages and dialects using a participatory-design approach to design effective and appropriate guardrails across all languages. We test these guardrails in a standard LLM-based application (meeting transcript summarization), where both the data generation and the annotation steps followed a human-in-the-loop approach. We find that the proposed guardrails are very effective in reducing misgendering rates across all languages in the summaries generated, and without incurring loss of quality. Our human-in-the-loop approach demonstrates a method to feasibly scale inclusive and responsible AI-based solutions across multiple languages and cultures. We release the guardrails and synthetic dataset encompassing 42 languages, along with human and LLM-judge evaluations, to encourage further research on this subject.

Adrian de Wynter

We prove that three strategy video games from the Sid Meier's Civilization series: Sid Meier's Civilization: Beyond Earth, Sid Meier's Civilization V, and Sid Meier's Civilization VI, are Turing complete. We achieve this by building three universal Turing machines-one for each game-using only the elements present in the games, and using their internal rules and mechanics as the transition function. The existence of such machines imply that under the assumptions made, the games are undecidable. We show constructions of these machines within a running game session, and we provide a sample execution of an algorithm-the three-state Busy Beaver-with one of our machines.

Adrian de Wynter, Daniel J. Perry

We extract an optimal subset of architectural parameters for the BERT architecture from Devlin et al. (2018) by applying recent breakthroughs in algorithms for neural architecture search. This optimal subset, which we refer to as "Bort", is demonstrably smaller, having an effective (that is, not counting the embedding layer) size of $5.5\%$ the original BERT-large architecture, and $16\%$ of the net size. Bort is also able to be pretrained in $288$ GPU hours, which is $1.2\%$ of the time required to pretrain the highest-performing BERT parametric architectural variant, RoBERTa-large (Liu et al., 2019), and about $33\%$ of that of the world-record, in GPU hours, required to train BERT-large on the same hardware. It is also $7.9$x faster on a CPU, as well as being better performing than other compressed variants of the architecture, and some of the non-compressed variants: it obtains performance improvements of between $0.3\%$ and $31\%$, absolute, with respect to BERT-large, on multiple public natural language understanding (NLU) benchmarks.

Adrian de Wynter

We introduce Mischief, a simple and lightweight method to produce a class of human-readable, realistic adversarial examples for language models. We perform exhaustive experimentations of our algorithm on four transformer-based architectures, across a variety of downstream tasks, as well as under varying concentrations of said examples. Our findings show that the presence of Mischief-generated adversarial samples in the test set significantly degrades (by up to $20\%$) the performance of these models with respect to their reported baselines. Nonetheless, we also demonstrate that, by including similar examples in the training set, it is possible to restore the baseline scores on the adversarial test set. Moreover, for certain tasks, the models trained with Mischief set show a modest increase on performance with respect to their original, non-adversarial baseline.

Adrian de Wynter

We consider the problem of finding the set of architectural parameters for a chosen deep neural network which is optimal under three metrics: parameter size, inference speed, and error rate. In this paper we state the problem formally, and present an approximation algorithm that, for a large subset of instances behaves like an FPTAS with an approximation error of $ρ\leq |{1- ε}|$, and that runs in $O(|Ξ| + |{W^*_T}|(1 + |Θ||{B}||Ξ|/({ε\, s^{3/2})}))$ steps, where $ε$ and $s$ are input parameters; $|{B}|$ is the batch size; $|{W^*_T}|$ denotes the cardinality of the largest weight set assignment; and $|Ξ|$ and $|Θ|$ are the cardinalities of the candidate architecture and hyperparameter spaces, respectively.

Adrian de Wynter

We present an algorithm for solving binary classification problems when the dataset is not fully representative of the problem being solved, and obtaining more data is not possible. It relies on a trained model with loose accuracy constraints, an iterative hyperparameter searching-and-pruning procedure over a search space $Θ$, and a data-generating function. Our algorithm works by reconstructing up to homology the manifold on which lies the support of the underlying distribution. We provide an analysis on correctness and runtime complexity under ideal conditions and an extension to deep neural networks. In the former case, if $\sizeΘ$ is the number of hyperparameter sets in the search space, this algorithm returns a solution that is up to $2(1 - {2^{-\sizeΘ}})$ times better than simply training with an enumeration of $Θ$ and picking the best model. As part of our analysis we also prove that an open cover of a dataset has the same homology as the manifold on which lies the support of the underlying probability distribution, if and only said dataset is learnable. This latter result acts as a formal argument to explain the effectiveness of data expansion techniques.

Adrian de Wynter

Within machine learning, the subfield of Neural Architecture Search (NAS) has recently garnered research attention due to its ability to improve upon human-designed models. However, the computational requirements for finding an exact solution to this problem are often intractable, and the design of the search space still requires manual intervention. In this paper we attempt to establish a formalized framework from which we can better understand the computational bounds of NAS in relation to its search space. For this, we first reformulate the function approximation problem in terms of sequences of functions, and we call it the Function Approximation (FA) problem; then we show that it is computationally infeasible to devise a procedure that solves FA for all functions to zero error, regardless of the search space. We show also that such error will be minimal if a specific class of functions is present in the search space. Subsequently, we show that machine learning as a mathematical problem is a solution strategy for FA, albeit not an effective one, and further describe a stronger version of this approach: the Approximate Architectural Search Problem (a-ASP), which is the mathematical equivalent of NAS. We leverage the framework from this paper and results from the literature to describe the conditions under which a-ASP can potentially solve FA as well as an exhaustive search, but in polynomial time.

Adrian de Wynter, Lambert Mathias

Dealing with previously unseen slots is a challenging problem in a real-world multi-domain dialogue state tracking task. Other approaches rely on predefined mappings to generate candidate slot keys, as well as their associated values. This, however, may fail when the key, the value, or both, are not seen during training. To address this problem we introduce a neural network that leverages external knowledge bases (KBs) to better classify out-of-vocabulary slot keys and values. This network projects the slot into an attribute space derived from the KB, and, by leveraging similarities in this space, we propose candidate slot keys and values to the dialogue state tracker. We provide extensive experiments that demonstrate that our stratagem can improve upon a previous approach, which relies on predefined candidate mappings. In particular, we evaluate this approach by training a state-of-the-art model with candidates generated from our network, and obtained relative increases of 57.7% and 82.7% in F1 score and accuracy, respectively, for the aforementioned model, when compared to the current candidate generation strategy.