Seraphina Goldfarb-Tarrant

Hadas Orgad, Seraphina Goldfarb-Tarrant, Yonatan Belinkov

Common studies of gender bias in NLP focus either on extrinsic bias measured by model performance on a downstream task or on intrinsic bias found in models' internal representations. However, the relationship between extrinsic and intrinsic bias is relatively unknown. In this work, we illuminate this relationship by measuring both quantities together: we debias a model during downstream fine-tuning, which reduces extrinsic bias, and measure the effect on intrinsic bias, which is operationalized as bias extractability with information-theoretic probing. Through experiments on two tasks and multiple bias metrics, we show that our intrinsic bias metric is a better indicator of debiasing than (a contextual adaptation of) the standard WEAT metric, and can also expose cases of superficial debiasing. Our framework provides a comprehensive perspective on bias in NLP models, which can be applied to deploy NLP systems in a more informed manner. Our code and model checkpoints are publicly available.

Preethi Seshadri, Samuel Cahyawijaya, Ayomide Odumakinde et al.

Agentic benchmarks increasingly rely on LLM-simulated users to scalably evaluate agent performance, yet the robustness, validity, and fairness of this approach remain unexamined. Through a user study with participants across the United States, India, Kenya, and Nigeria, we investigate whether LLM-simulated users serve as reliable proxies for real human users in evaluating agents on τ-Bench retail tasks. We find that user simulation lacks robustness, with agent success rates varying up to 9 percentage points across different user LLMs. Furthermore, evaluations using simulated users exhibit systematic miscalibration, underestimating agent performance on challenging tasks and overestimating it on moderately difficult ones. African American Vernacular English (AAVE) speakers experience consistently worse success rates and calibration errors than Standard American English (SAE) speakers, with disparities compounding significantly with age. We also find simulated users to be a differentially effective proxy for different populations, performing worst for AAVE and Indian English speakers. Additionally, simulated users introduce conversational artifacts and surface different failure patterns than human users. These findings demonstrate that current evaluation practices risk misrepresenting agent capabilities across diverse user populations and may obscure real-world deployment challenges.

Henry C. Conklin, Tom Hosking, Tan Yi-Chern et al.

Despite the increasing prevalence of large language models (LLMs), we still have a limited understanding of how their representational spaces are structured. This limits our ability to interpret how and what they learn or relate them to learning in humans. We argue LLMs are best seen as an instance of lossy compression, where over training they learn by retaining only information in their training data relevant to their objective(s). We show pre-training results in models that are optimally compressed for next-sequence prediction, approaching the Information Bottleneck bound on compression. Across an array of open weights models, each compresses differently, likely due to differences in the data and training recipes used. However even across different families of LLMs the optimality of a model's compression, and the information present in it, can predict downstream performance on across a wide array of benchmarks, letting us directly link representational structure to actionable insights about model performance. In the general case the work presented here offers a unified Information-Theoretic framing for how these models learn that is deployable at scale.

Omer Hofman, Jonathan Brokman, Oren Rachmil et al.

Agentic AI systems, which build on Large Language Models (LLMs) and interact with tools and memory, have rapidly advanced in capability and scope. Yet, since LLMs have been shown to struggle in multilingual settings, typically resulting in lower performance and reduced safety, agentic systems risk inheriting these limitations. This raises concerns about the accessibility of such systems, as users interacting in languages other than English may encounter unreliable or security-critical agent behavior. Despite growing interest in evaluating agentic AI, existing benchmarks focus exclusively on English, leaving multilingual settings unexplored. To address this gap, we propose MAPS, a multilingual benchmark suite designed to evaluate agentic AI systems across diverse languages and tasks. MAPS builds on four widely used agentic benchmarks - GAIA (real-world tasks), SWE-bench (code generation), MATH (mathematical reasoning), and the Agent Security Benchmark (security). We translate each dataset into eleven diverse languages, resulting in 805 unique tasks and 9,660 total language-specific instances - enabling a systematic analysis of the multilingual effect on AI agents' performance and robustness. Empirically, we observe degradation in both performance and security when transitioning from English to other languages, with severity varying by task and correlating with the amount of translated input. Building on these findings, we provide actionable recommendations to guide agentic AI systems development and assessment under multilingual settings. This work establishes the first standardized evaluation framework for multilingual agentic AI, encouraging future research towards equitable, reliable, and accessible agentic AI. MAPS benchmark suite is publicly available at https://huggingface.co/datasets/Fujitsu-FRE/MAPS

Team Cohere, Aakanksha, Arash Ahmadian et al. · mila

In this report we describe the development of Command A, a powerful large language model purpose-built to excel at real-world enterprise use cases. Command A is an agent-optimised and multilingual-capable model, with support for 23 languages of global business, and a novel hybrid architecture balancing efficiency with top of the range performance. It offers best-in-class Retrieval Augmented Generation (RAG) capabilities with grounding and tool use to automate sophisticated business processes. These abilities are achieved through a decentralised training approach, including self-refinement algorithms and model merging techniques. We also include results for Command R7B which shares capability and architectural similarities to Command A. Weights for both models have been released for research purposes. This technical report details our original training pipeline and presents an extensive evaluation of our models across a suite of enterprise-relevant tasks and public benchmarks, demonstrating excellent performance and efficiency.

Aakanksha, Arash Ahmadian, Seraphina Goldfarb-Tarrant et al.

Large Language Models (LLMs) have been adopted and deployed worldwide for a broad variety of applications. However, ensuring their safe use remains a significant challenge. Preference training and safety measures often overfit to harms prevalent in Western-centric datasets, and safety protocols frequently fail to extend to multilingual settings. In this work, we explore model merging in a diverse multi-task setting, combining safety and general-purpose tasks within a multilingual context. Each language introduces unique and varied learning challenges across tasks. We find that objective-based merging is more effective than mixing data, with improvements of up to 8% and 10% in general performance and safety respectively. We also find that language-based merging is highly effective -- by merging monolingually fine-tuned models, we achieve a 4% increase in general performance and 7% reduction in harm across all languages on top of the data mixtures method using the same available data. Overall, our comprehensive study of merging approaches provides a useful framework for building strong and safe multilingual models.

Neel Rajani, Aryo Pradipta Gema, Seraphina Goldfarb-Tarrant et al.

Training large language models (LLMs) for reasoning via maths and code datasets has become a major new focus in LLM post-training. Two particularly popular approaches are reinforcement learning (RL) and supervised fine-tuning (SFT), but their training dynamics are poorly understood. We present a comparative analysis of RL and SFT on the same maths problems with the same model and similar hyperparameters. We find that RL yields minor in-domain gains on maths and slight degradation on knowledge-intensive benchmarks like MMLU, while both trends are more pronounced in SFT. We also analyse model parameters across checkpoints, observing that both algorithms modify query and key weights the most. Meanwhile, SFT exhibits greater updates and also affects mid-layer MLPs more, leading us to hypothesise that this may have caused the out-of-domain degradation. We therefore investigate whether freezing parts of the model during training can mitigate the reduced performance on knowledge-intensive benchmarks. However, our results are inconclusive, with benefits on GPQA:Diamond and degradation on other benchmarks. Taken together, our observations provide a preliminary indication for why RL amplifies existing capabilities, while SFT replaces old skills with new ones.

Hongyu Chen, Seraphina Goldfarb-Tarrant

Large Language Models (LLMs) are increasingly employed as automated evaluators to assess the safety of generated content, yet their reliability in this role remains uncertain. This study evaluates a diverse set of 11 LLM judge models across critical safety domains, examining three key aspects: self-consistency in repeated judging tasks, alignment with human judgments, and susceptibility to input artifacts such as apologetic or verbose phrasing. Our findings reveal that biases in LLM judges can significantly distort the final verdict on which content source is safer, undermining the validity of comparative evaluations. Notably, apologetic language artifacts alone can skew evaluator preferences by up to 98\%. Contrary to expectations, larger models do not consistently exhibit greater robustness, while smaller models sometimes show higher resistance to specific artifacts. To mitigate LLM evaluator robustness issues, we investigate jury-based evaluations aggregating decisions from multiple models. Although this approach both improves robustness and enhances alignment to human judgements, artifact sensitivity persists even with the best jury configurations. These results highlight the urgent need for diversified, artifact-resistant methodologies to ensure reliable safety assessments.

Preethi Seshadri, Hongyu Chen, Sameer Singh et al.

Large language models (LLMs) are increasingly being deployed in high-stakes applications like hiring, yet their potential for unfair decision-making remains understudied in generative and retrieval settings. In this work, we examine the allocational fairness of LLM-based hiring systems through two tasks that reflect actual HR usage: resume summarization and applicant ranking. By constructing a synthetic resume dataset with controlled perturbations and curating job postings, we investigate whether model behavior differs across demographic groups. Our findings reveal that generated summaries exhibit meaningful differences more frequently for race than for gender perturbations. Models also display non-uniform retrieval selection patterns across demographic groups and exhibit high ranking sensitivity to both gender and race perturbations. Surprisingly, retrieval models can show comparable sensitivity to both demographic and non-demographic changes, suggesting that fairness issues may stem from broader model brittleness. Overall, our results indicate that LLM-based hiring systems, especially in the retrieval stage, can exhibit notable biases that lead to discriminatory outcomes in real-world contexts.

Seraphina Goldfarb-Tarrant, Pedro Rodriguez, Jane Dwivedi-Yu et al.

Dense retrievers compress source documents into (possibly lossy) vector representations, yet there is little analysis of what information is lost versus preserved, and how it affects downstream tasks. We conduct the first analysis of the information captured by dense retrievers compared to the language models they are based on (e.g., BERT versus Contriever). We use 25 MultiBert checkpoints as randomized initialisations to train MultiContrievers, a set of 25 contriever models. We test whether specific pieces of information -- such as gender and occupation -- can be extracted from contriever vectors of wikipedia-like documents. We measure this extractability via information theoretic probing. We then examine the relationship of extractability to performance and gender bias, as well as the sensitivity of these results to many random initialisations and data shuffles. We find that (1) contriever models have significantly increased extractability, but extractability usually correlates poorly with benchmark performance 2) gender bias is present, but is not caused by the contriever representations 3) there is high sensitivity to both random initialisation and to data shuffle, suggesting that future retrieval research should test across a wider spread of both.

Hadas Orgad, Boyi Wei, Kaden Zheng et al.

Large language models (LLMs) undergo alignment training to avoid harmful behaviors, yet the resulting safeguards remain brittle: jailbreaks routinely bypass them, and fine-tuning on narrow domains can induce ``emergent misalignment'' that generalizes broadly. Whether this brittleness reflects a fundamental lack of coherent internal organization for harmfulness remains unclear. Here we use targeted weight pruning as a causal intervention to probe the internal organization of harmfulness in LLMs. We find that harmful content generation depends on a compact set of weights that are general across harm types and distinct from benign capabilities. Aligned models exhibit a greater compression of harm generation weights than unaligned counterparts, indicating that alignment reshapes harmful representations internally--despite the brittleness of safety guardrails at the surface level. This compression explains emergent misalignment: if weights of harmful capabilities are compressed, fine-tuning that engages these weights in one domain can trigger broad misalignment. Consistent with this, pruning harm generation weights in a narrow domain substantially reduces emergent misalignment. Notably, LLMs harmful generation capability is dissociated from how they recognize and explain such content. Together, these results reveal a coherent internal structure for harmfulness in LLMs that may serve as a foundation for more principled approaches to safety.

Aidan Peppin, Julia Kreutzer, Alice Schoenauer Sebag et al.

Despite advances in large language model capabilities in recent years, a large gap remains in their capabilities and safety performance for many languages beyond a relatively small handful of globally dominant languages. This paper provides researchers, policymakers and governance experts with an overview of key challenges to bridging the "language gap" in AI and minimizing safety risks across languages. We provide an analysis of why the language gap in AI exists and grows, and how it creates disparities in global AI safety. We identify barriers to address these challenges, and recommend how those working in policy and governance can help address safety concerns associated with the language gap by supporting multilingual dataset creation, transparency, and research.

Nishant Balepur, Matthew Shu, Yoo Yeon Sung et al. · allen-ai, oxford

To assist users in complex tasks, LLMs generate plans: step-by-step instructions towards a goal. While alignment methods aim to ensure LLM plans are helpful, they train (RLHF) or evaluate (ChatbotArena) on what users prefer, assuming this reflects what helps them. We test this with Planorama: an interface where 126 users answer 300 multi-step questions with LLM plans. We get 4388 plan executions and 5584 comparisons to measure plan helpfulness (QA success) and user preferences on plans, and recreate the setup in agents and reward models to see if they simulate or prefer what helps users. We expose: 1) user/model preferences and agent success do not accurately predict which plans help users, so common alignment feedback can misalign with helpfulness; 2) this gap is not due to user-specific preferences, as users are similarly successful when using plans they prefer/disprefer; 3) surface-level cues like brevity and question similarity strongly link to preferences, but such biases fail to predict helpfulness. In all, we argue aligning helpful LLMs needs feedback from real user interactions, not just preferences of what looks helpful, so we discuss the plan NLP researchers can execute to solve this problem.

Aakanksha, Arash Ahmadian, Beyza Ermis et al.

A key concern with the concept of "alignment" is the implicit question of "alignment to what?". AI systems are increasingly used across the world, yet safety alignment is often focused on homogeneous monolingual settings. Additionally, preference training and safety measures often overfit to harms common in Western-centric datasets. Here, we explore the viability of different alignment approaches when balancing dual objectives: addressing and optimizing for a non-homogeneous set of languages and cultural preferences while minimizing both global and local harms. We collect the first set of human annotated red-teaming prompts in different languages distinguishing between global and local harm, which serve as a laboratory for understanding the reliability of alignment techniques when faced with preference distributions that are non-stationary across geographies and languages. While this setting is seldom covered by the literature to date, which primarily centers on English harm mitigation, it captures real-world interactions with AI systems around the world. We establish a new precedent for state-of-the-art alignment techniques across 6 languages with minimal degradation in general performance. Our work provides important insights into cross-lingual transfer and novel optimization approaches to safeguard AI systems designed to serve global populations.

Nishant Balepur, Matthew Shu, Alexander Hoyle et al.

Keyword mnemonics are memorable explanations that link new terms to simpler keywords. Prior work generates mnemonics for students, but they do not train models using mnemonics students prefer and aid learning. We build SMART, a mnemonic generator trained on feedback from real students learning new terms. To train SMART, we first fine-tune LLaMA-2 on a curated set of user-written mnemonics. We then use LLM alignment to enhance SMART: we deploy mnemonics generated by SMART in a flashcard app to find preferences on mnemonics students favor. We gather 2684 preferences from 45 students across two types: expressed (inferred from ratings) and observed (inferred from student learning), yielding three key findings. First, expressed and observed preferences disagree; what students think is helpful does not always capture what is truly helpful. Second, Bayesian models can synthesize complementary data from multiple preference types into a single effectiveness signal. SMART is tuned via Direct Preference Optimization on this signal, which resolves ties and missing labels in the typical method of pairwise comparisons, augmenting data for LLM output quality gains. Third, mnemonic experts assess SMART as matching GPT-4 at much lower deployment costs, showing the utility of capturing diverse student feedback to align LLMs in education.

Seraphina Goldfarb-Tarrant, Eddie Ungless, Esma Balkir et al.

Bias research in NLP seeks to analyse models for social biases, thus helping NLP practitioners uncover, measure, and mitigate social harms. We analyse the body of work that uses prompts and templates to assess bias in language models. We draw on a measurement modelling framework to create a taxonomy of attributes that capture what a bias test aims to measure and how that measurement is carried out. By applying this taxonomy to 90 bias tests, we illustrate qualitatively and quantitatively that core aspects of bias test conceptualisations and operationalisations are frequently unstated or ambiguous, carry implicit assumptions, or be mismatched. Our analysis illuminates the scope of possible bias types the field is able to measure, and reveals types that are as yet under-researched. We offer guidance to enable the community to explore a wider section of the possible bias space, and to better close the gap between desired outcomes and experimental design, both for bias and for evaluating language models more broadly.

Seraphina Goldfarb-Tarrant, Björn Ross, Adam Lopez

Sentiment analysis (SA) systems are widely deployed in many of the world's languages, and there is well-documented evidence of demographic bias in these systems. In languages beyond English, scarcer training data is often supplemented with transfer learning using pre-trained models, including multilingual models trained on other languages. In some cases, even supervision data comes from other languages. Does cross-lingual transfer also import new biases? To answer this question, we use counterfactual evaluation to test whether gender or racial biases are imported when using cross-lingual transfer, compared to a monolingual transfer setting. Across five languages, we find that systems using cross-lingual transfer usually become more biased than their monolingual counterparts. We also find racial biases to be much more prevalent than gender biases. To spur further research on this topic, we release the sentiment models we used for this study, and the intermediate checkpoints throughout training, yielding 1,525 distinct models; we also release our evaluation code.

Seraphina Goldfarb-Tarrant, Adam Lopez, Roi Blanco et al.

Sentiment analysis (SA) systems are used in many products and hundreds of languages. Gender and racial biases are well-studied in English SA systems, but understudied in other languages, with few resources for such studies. To remedy this, we build a counterfactual evaluation corpus for gender and racial/migrant bias in four languages. We demonstrate its usefulness by answering a simple but important question that an engineer might need to answer when deploying a system: What biases do systems import from pre-trained models when compared to a baseline with no pre-training? Our evaluation corpus, by virtue of being counterfactual, not only reveals which models have less bias, but also pinpoints changes in model bias behaviour, which enables more targeted mitigation strategies. We release our code and evaluation corpora to facilitate future research.

Seraphina Goldfarb-Tarrant, Rebecca Marchant, Ricardo Muñoz Sanchez et al.

Natural Language Processing (NLP) systems learn harmful societal biases that cause them to amplify inequality as they are deployed in more and more situations. To guide efforts at debiasing these systems, the NLP community relies on a variety of metrics that quantify bias in models. Some of these metrics are intrinsic, measuring bias in word embedding spaces, and some are extrinsic, measuring bias in downstream tasks that the word embeddings enable. Do these intrinsic and extrinsic metrics correlate with each other? We compare intrinsic and extrinsic metrics across hundreds of trained models covering different tasks and experimental conditions. Our results show no reliable correlation between these metrics that holds in all scenarios across tasks and languages. We urge researchers working on debiasing to focus on extrinsic measures of bias, and to make using these measures more feasible via creation of new challenge sets and annotated test data. To aid this effort, we release code, a new intrinsic metric, and an annotated test set focused on gender bias in hate speech.

Seraphina Goldfarb-Tarrant, Alexander Robertson, Jasmina Lazic et al.

Systematic reviews, which entail the extraction of data from large numbers of scientific documents, are an ideal avenue for the application of machine learning. They are vital to many fields of science and philanthropy, but are very time-consuming and require experts. Yet the three main stages of a systematic review are easily done automatically: searching for documents can be done via APIs and scrapers, selection of relevant documents can be done via binary classification, and extraction of data can be done via sequence-labelling classification. Despite the promise of automation for this field, little research exists that examines the various ways to automate each of these tasks. We construct a pipeline that automates each of these aspects, and experiment with many human-time vs. system quality trade-offs. We test the ability of classifiers to work well on small amounts of data and to generalise to data from countries not represented in the training data. We test different types of data extraction with varying difficulty in annotation, and five different neural architectures to do the extraction. We find that we can get surprising accuracy and generalisability of the whole pipeline system with only 2 weeks of human-expert annotation, which is only 15% of the time it takes to do the whole review manually and can be repeated and extended to new data with no additional effort.

Seraphina Goldfarb-Tarrant, Tuhin Chakrabarty, Ralph Weischedel et al.

Long-form narrative text generated from large language models manages a fluent impersonation of human writing, but only at the local sentence level, and lacks structure or global cohesion. We posit that many of the problems of story generation can be addressed via high-quality content planning, and present a system that focuses on how to learn good plot structures to guide story generation. We utilize a plot-generation language model along with an ensemble of rescoring models that each implement an aspect of good story-writing as detailed in Aristotle's Poetics. We find that stories written with our more principled plot-structure are both more relevant to a given prompt and higher quality than baselines that do not content plan, or that plan in an unprincipled way.

Seraphina Goldfarb-Tarrant, Haining Feng, Nanyun Peng

Story composition is a challenging problem for machines and even for humans. We present a neural narrative generation system that interacts with humans to generate stories. Our system has different levels of human interaction, which enables us to understand at what stage of story-writing human collaboration is most productive, both to improving story quality and human engagement in the writing process. We compare different varieties of interaction in story-writing, story-planning, and diversity controls under time constraints, and show that increased types of human collaboration at both planning and writing stages results in a 10-50% improvement in story quality as compared to less interactive baselines. We also show an accompanying increase in user engagement and satisfaction with stories as compared to our own less interactive systems and to previous turn-taking approaches to interaction. Finally, we find that humans tasked with collaboratively improving a particular characteristic of a story are in fact able to do so, which has implications for future uses of human-in-the-loop systems.