Anna Korhonen

Abdullatif Köksal, Timo Schick, Anna Korhonen et al. · meta-ai

Instruction tuning enables language models to more effectively generalize and better follow user intent. However, obtaining instruction data is costly and challenging. Prior work employs methods such as expensive human annotation, crowd-sourced datasets with alignment issues, and generating noisy examples via LLMs. We introduce the LongForm-C dataset, which is created by reverse instructions. We generate instructions via LLMs for human-written corpus examples using reverse instructions. First we select a diverse set of human-written documents from corpora such as C4 and Wikipedia; then we generate instructions for these documents via LLMs. This approach provides a cheaper and cleaner instruction-tuning dataset with natural output and one suitable for long text generation. Our models outperform 10x larger language models without instruction tuning on tasks such as story/recipe generation and long-form question answering. Moreover, LongForm models outperform prior instruction-tuned models such as FLAN-T5 and Alpaca by a large margin, and improve language understanding capabilities further. We publicly release our data and models: https://github.com/akoksal/LongForm.

Yifu Qiu, Zheng Zhao, Yftah Ziser et al. · cambridge

Are Large language models (LLMs) temporally grounded? Since LLMs cannot perceive and interact with the environment, it is impossible to answer this question directly. Instead, we provide LLMs with textual narratives and probe them with respect to their common-sense knowledge of the structure and duration of events, their ability to order events along a timeline, and self-consistency within their temporal model (e.g., temporal relations such as after and before are mutually exclusive for any pair of events). We evaluate state-of-the-art LLMs (such as LLaMA 2 and GPT-4) on three tasks reflecting these abilities. Generally, we find that LLMs lag significantly behind both human performance as well as small-scale, specialised LMs. In-context learning, instruction tuning, and chain-of-thought prompting reduce this gap only to a limited degree. Crucially, LLMs struggle the most with self-consistency, displaying incoherent behaviour in at least 27.23% of their predictions. Contrary to expectations, we also find that scaling the model size does not guarantee positive gains in performance. To explain these results, we study the sources from which LLMs may gather temporal information: we find that sentence ordering in unlabelled texts, available during pre-training, is only weakly correlated with event ordering. Moreover, public instruction tuning mixtures contain few temporal tasks. Hence, we conclude that current LLMs lack a consistent temporal model of textual narratives. Code, datasets, and LLM outputs are available at https://github.com/yfqiu-nlp/temporal-llms.

Yaoyiran Li, Fangyu Liu, Nigel Collier et al. · cambridge, deepmind

Word translation or bilingual lexicon induction (BLI) is a key cross-lingual task, aiming to bridge the lexical gap between different languages. In this work, we propose a robust and effective two-stage contrastive learning framework for the BLI task. At Stage C1, we propose to refine standard cross-lingual linear maps between static word embeddings (WEs) via a contrastive learning objective; we also show how to integrate it into the self-learning procedure for even more refined cross-lingual maps. In Stage C2, we conduct BLI-oriented contrastive fine-tuning of mBERT, unlocking its word translation capability. We also show that static WEs induced from the `C2-tuned' mBERT complement static WEs from Stage C1. Comprehensive experiments on standard BLI datasets for diverse languages and different experimental setups demonstrate substantial gains achieved by our framework. While the BLI method from Stage C1 already yields substantial gains over all state-of-the-art BLI methods in our comparison, even stronger improvements are met with the full two-stage framework: e.g., we report gains for 112/112 BLI setups, spanning 28 language pairs.

Yaoyiran Li, Fangyu Liu, Ivan Vulić et al. · cambridge, deepmind

Bilingual lexicon induction (BLI) with limited bilingual supervision is a crucial yet challenging task in multilingual NLP. Current state-of-the-art BLI methods rely on the induction of cross-lingual word embeddings (CLWEs) to capture cross-lingual word similarities; such CLWEs are obtained 1) via traditional static models (e.g., VecMap), or 2) by extracting type-level CLWEs from multilingual pretrained language models (mPLMs), or 3) through combining the former two options. In this work, we propose a novel semi-supervised post-hoc reranking method termed BLICEr (BLI with Cross-Encoder Reranking), applicable to any precalculated CLWE space, which improves their BLI capability. The key idea is to 'extract' cross-lingual lexical knowledge from mPLMs, and then combine it with the original CLWEs. This crucial step is done via 1) creating a word similarity dataset, comprising positive word pairs (i.e., true translations) and hard negative pairs induced from the original CLWE space, and then 2) fine-tuning an mPLM (e.g., mBERT or XLM-R) in a cross-encoder manner to predict the similarity scores. At inference, we 3) combine the similarity score from the original CLWE space with the score from the BLI-tuned cross-encoder. BLICEr establishes new state-of-the-art results on two standard BLI benchmarks spanning a wide spectrum of diverse languages: it substantially outperforms a series of strong baselines across the board. We also validate the robustness of BLICEr with different CLWEs.

Ivan Vulić, Goran Glavaš, Fangyu Liu et al. · deepmind

Pretrained multilingual language models (LMs) can be successfully transformed into multilingual sentence encoders (SEs; e.g., LaBSE, xMPNet) via additional fine-tuning or model distillation with parallel data. However, it remains unclear how to best leverage them to represent sub-sentence lexical items (i.e., words and phrases) in cross-lingual lexical tasks. In this work, we probe SEs for the amount of cross-lingual lexical knowledge stored in their parameters, and compare them against the original multilingual LMs. We also devise a simple yet efficient method for exposing the cross-lingual lexical knowledge by means of additional fine-tuning through inexpensive contrastive learning that requires only a small amount of word translation pairs. Using bilingual lexical induction (BLI), cross-lingual lexical semantic similarity, and cross-lingual entity linking as lexical probing tasks, we report substantial gains on standard benchmarks (e.g., +10 Precision@1 points in BLI). The results indicate that the SEs such as LaBSE can be 'rewired' into effective cross-lingual lexical encoders via the contrastive learning procedure, and that they contain more cross-lingual lexical knowledge than what 'meets the eye' when they are used as off-the-shelf SEs. This way, we also provide an effective tool for harnessing 'covert' multilingual lexical knowledge hidden in multilingual sentence encoders.

Han Zhou, Xingchen Wan, Ivan Vulić et al. · cambridge

Large pretrained language models are widely used in downstream NLP tasks via task-specific fine-tuning, but such procedures can be costly. Recently, Parameter-Efficient Fine-Tuning (PEFT) methods have achieved strong task performance while updating much fewer parameters than full model fine-tuning (FFT). However, it is non-trivial to make informed design choices on the PEFT configurations, such as their architecture, the number of tunable parameters, and even the layers in which the PEFT modules are inserted. Consequently, it is highly likely that the current, manually designed configurations are suboptimal in terms of their performance-efficiency trade-off. Inspired by advances in neural architecture search, we propose AutoPEFT for automatic PEFT configuration selection: we first design an expressive configuration search space with multiple representative PEFT modules as building blocks. Using multi-objective Bayesian optimisation in a low-cost setup, we then discover a Pareto-optimal set of configurations with strong performance-cost trade-offs across different numbers of parameters that are also highly transferable across different tasks. Empirically, on GLUE and SuperGLUE tasks, we show that AutoPEFT-discovered configurations significantly outperform existing PEFT methods and are on par or better than FFT without incurring substantial training efficiency costs.

Alan Ansell, Edoardo Maria Ponti, Anna Korhonen et al.

Massively multilingual Transformers (MMTs), such as mBERT and XLM-R, are widely used for cross-lingual transfer learning. While these are pretrained to represent hundreds of languages, end users of NLP systems are often interested only in individual languages. For such purposes, the MMTs' language coverage makes them unnecessarily expensive to deploy in terms of model size, inference time, energy, and hardware cost. We thus propose to extract compressed, language-specific models from MMTs which retain the capacity of the original MMTs for cross-lingual transfer. This is achieved by distilling the MMT bilingually, i.e., using data from only the source and target language of interest. Specifically, we use a two-phase distillation approach, termed BiStil: (i) the first phase distils a general bilingual model from the MMT, while (ii) the second, task-specific phase sparsely fine-tunes the bilingual "student" model using a task-tuned variant of the original MMT as its "teacher". We evaluate this distillation technique in zero-shot cross-lingual transfer across a number of standard cross-lingual benchmarks. The key results indicate that the distilled models exhibit minimal degradation in target language performance relative to the base MMT despite being significantly smaller and faster. Furthermore, we find that they outperform multilingually distilled models such as DistilmBERT and MiniLMv2 while having a very modest training budget in comparison, even on a per-language basis. We also show that bilingual models distilled from MMTs greatly outperform bilingual models trained from scratch. Our code and models are available at https://github.com/AlanAnsell/bistil.

Songbo Hu, Han Zhou, Mete Hergul et al. · cambridge

Creating high-quality annotated data for task-oriented dialog (ToD) is known to be notoriously difficult, and the challenges are amplified when the goal is to create equitable, culturally adapted, and large-scale ToD datasets for multiple languages. Therefore, the current datasets are still very scarce and suffer from limitations such as translation-based non-native dialogs with translation artefacts, small scale, or lack of cultural adaptation, among others. In this work, we first take stock of the current landscape of multilingual ToD datasets, offering a systematic overview of their properties and limitations. Aiming to reduce all the detected limitations, we then introduce Multi3WOZ, a novel multilingual, multi-domain, multi-parallel ToD dataset. It is large-scale and offers culturally adapted dialogs in 4 languages to enable training and evaluation of multilingual and cross-lingual ToD systems. We describe a complex bottom-up data collection process that yielded the final dataset, and offer the first sets of baseline scores across different ToD-related tasks for future reference, also highlighting its challenging nature.

Yifu Qiu, Yftah Ziser, Anna Korhonen et al.

Can unified vision-language models (VLMs) perform forward dynamics prediction (FDP), i.e., predicting the future state (in image form) given the previous observation and an action (in language form)? We find that VLMs struggle to generate physically plausible transitions between frames from instructions. Nevertheless, we identify a crucial asymmetry in multimodal grounding: fine-tuning a VLM to learn inverse dynamics prediction (IDP)-effectively captioning the action between frames-is significantly easier than learning FDP. In turn, IDP can be used to bootstrap FDP through two main strategies: 1) weakly supervised learning from synthetic data and 2) inference time verification. Firstly, IDP can annotate actions for unlabelled pairs of video frame observations to expand the training data scale for FDP. Secondly, IDP can assign rewards to multiple samples of FDP to score them, effectively guiding search at inference time. We evaluate the FDP resulting from both strategies through the task of action-centric image editing on Aurora-Bench with two families of VLMs. Despite remaining general-purpose, our best model achieves a performance competitive with state-of-the-art image editing models, improving on them by a margin between 7% and 13% according to GPT4o-as-judge, and achieving the best average human evaluation across all subsets of Aurora-Bench.

Zhangdie Yuan, Songbo Hu, Ivan Vulić et al. · cambridge

Acquiring factual knowledge with Pretrained Language Models (PLMs) has attracted increasing attention, showing promising performance in many knowledge-intensive tasks. Their good performance has led the community to believe that the models do possess a modicum of reasoning competence rather than merely memorising the knowledge. In this paper, we conduct a comprehensive evaluation of the learnable deductive (also known as explicit) reasoning capability of PLMs. Through a series of controlled experiments, we posit two main findings. (i) PLMs inadequately generalise learned logic rules and perform inconsistently against simple adversarial surface form edits. (ii) While the deductive reasoning fine-tuning of PLMs does improve their performance on reasoning over unseen knowledge facts, it results in catastrophically forgetting the previously learnt knowledge. Our main results suggest that PLMs cannot yet perform reliable deductive reasoning, demonstrating the importance of controlled examinations and probing of PLMs' reasoning abilities; we reach beyond (misleading) task performance, revealing that PLMs are still far from human-level reasoning capabilities, even for simple deductive tasks.

Yaoyiran Li, Anna Korhonen, Ivan Vulić · cambridge

Bilingual Lexicon Induction (BLI) is a core task in multilingual NLP that still, to a large extent, relies on calculating cross-lingual word representations. Inspired by the global paradigm shift in NLP towards Large Language Models (LLMs), we examine the potential of the latest generation of LLMs for the development of bilingual lexicons. We ask the following research question: Is it possible to prompt and fine-tune multilingual LLMs (mLLMs) for BLI, and how does this approach compare against and complement current BLI approaches? To this end, we systematically study 1) zero-shot prompting for unsupervised BLI and 2) few-shot in-context prompting with a set of seed translation pairs, both without any LLM fine-tuning, as well as 3) standard BLI-oriented fine-tuning of smaller LLMs. We experiment with 18 open-source text-to-text mLLMs of different sizes (from 0.3B to 13B parameters) on two standard BLI benchmarks covering a range of typologically diverse languages. Our work is the first to demonstrate strong BLI capabilities of text-to-text mLLMs. The results reveal that few-shot prompting with in-context examples from nearest neighbours achieves the best performance, establishing new state-of-the-art BLI scores for many language pairs. We also conduct a series of in-depth analyses and ablation studies, providing more insights on BLI with (m)LLMs, also along with their limitations.

Songbo Hu, Ivan Vulić, Fangyu Liu et al. · cambridge, deepmind

End-to-end (E2E) task-oriented dialogue (ToD) systems are prone to fall into the so-called "likelihood trap", resulting in generated responses which are dull, repetitive, and often inconsistent with dialogue history. Comparing ranked lists of multiple generated responses against the "gold response" (from evaluation data) reveals a wide diversity in response quality, with many good responses placed lower in the ranked list. The main challenge, addressed in this work, is then how to reach beyond greedily generated system responses, that is, how to obtain and select such high-quality responses from the list of overgenerated responses at inference without availability of the gold response. To this end, we propose a simple yet effective reranking method which aims to select high-quality items from the lists of responses initially overgenerated by the system. The idea is to use any sequence-level (similarity) scoring function to divide the semantic space of responses into high-scoring versus low-scoring partitions. At training, the high-scoring partition comprises all generated responses whose similarity to the gold response is higher than the similarity of the greedy response to the gold response. At inference, the aim is to estimate the probability that each overgenerated response belongs to the high-scoring partition, given only previous dialogue history. We validate the robustness and versatility of our proposed method on the standard MultiWOZ dataset: our methods improve a state-of-the-art E2E ToD system by 2.0 BLEU, 1.6 ROUGE, and 1.3 METEOR scores, achieving new peak results. Additional experiments on the BiTOD dataset and human evaluation further ascertain the generalisability and effectiveness of the proposed framework.

Arda Yüksel, Abdullatif Köksal, Lütfi Kerem Şenel et al.

Multiple choice question answering tasks evaluate the reasoning, comprehension, and mathematical abilities of Large Language Models (LLMs). While existing benchmarks employ automatic translation for multilingual evaluation, this approach is error-prone and potentially introduces culturally biased questions, especially in social sciences. We introduce the first multitask, multiple-choice Turkish QA benchmark, TurkishMMLU, to evaluate LLMs' understanding of the Turkish language. TurkishMMLU includes over 10,000 questions, covering 9 different subjects from Turkish high-school education curricula. These questions are written by curriculum experts, suitable for the high-school curricula in Turkey, covering subjects ranging from natural sciences and math questions to more culturally representative topics such as Turkish Literature and the history of the Turkish Republic. We evaluate over 20 LLMs, including multilingual open-source (e.g., Gemma, Llama, MT5), closed-source (GPT 4o, Claude, Gemini), and Turkish-adapted (e.g., Trendyol) models. We provide an extensive evaluation, including zero-shot and few-shot evaluation of LLMs, chain-of-thought reasoning, and question difficulty analysis along with model performance. We provide an in-depth analysis of the Turkish capabilities and limitations of current LLMs to provide insights for future LLMs for the Turkish language. We publicly release our code for the dataset and evaluation: https://github.com/ArdaYueksel/TurkishMMLU.

Abdullatif Köksal, Marion Thaler, Ayyoob Imani et al.

Instruction tuning enhances large language models (LLMs) by aligning them with human preferences across diverse tasks. Traditional approaches to create instruction tuning datasets face serious challenges for low-resource languages due to their dependence on data annotation. This work introduces a novel method, Multilingual Reverse Instructions (MURI), which generates high-quality instruction tuning datasets for low-resource languages without requiring human annotators or pre-existing multilingual models. Utilizing reverse instructions and a translation pipeline, MURI produces instruction-output pairs from existing human-written texts in low-resource languages. This method ensures cultural relevance and diversity by sourcing texts from different native domains and applying filters to eliminate inappropriate content. Our dataset, MURI-IT, includes more than 2 million instruction-output pairs across 200 languages. Evaluation by native speakers and fine-tuning experiments with mT5 models demonstrate the approach's effectiveness for both NLU and open-ended generation. We publicly release datasets and models at https://github.com/akoksal/muri.

Han Zhou, Xingchen Wan, Ivan Vulić et al. · cambridge

Prompt-based learning has been an effective paradigm for large pretrained language models (LLM), enabling few-shot or even zero-shot learning. Black-box prompt search has received growing interest recently for its distinctive properties of gradient-free optimization, proven particularly useful and powerful for model-as-a-service usage. However, the discrete nature and the complexity of combinatorial optimization hinder the efficiency of modern black-box approaches. Despite extensive research on search algorithms, the crucial aspect of search space design and optimization has been largely overlooked. In this paper, we first conduct a sensitivity analysis by prompting LLM, revealing that only a small number of tokens exert a disproportionate amount of influence on LLM predictions. Leveraging this insight, we propose the Clustering and Pruning for Efficient Black-box Prompt Search (ClaPS), a simple black-box search method that first clusters and prunes the search space to focus exclusively on influential prompt tokens. By employing even simple search methods within the pruned search space, ClaPS achieves state-of-the-art performance across various tasks and LLMs, surpassing the performance of complex approaches while significantly reducing search costs. Our findings underscore the critical role of search space design and optimization in enhancing both the usefulness and the efficiency of black-box prompt-based learning.

Nikita Moghe, Evgeniia Razumovskaia, Liane Guillou et al.

Task-oriented dialogue (TOD) systems have been widely deployed in many industries as they deliver more efficient customer support. These systems are typically constructed for a single domain or language and do not generalise well beyond this. To support work on Natural Language Understanding (NLU) in TOD across multiple languages and domains simultaneously, we constructed MULTI3NLU++, a multilingual, multi-intent, multi-domain dataset. MULTI3NLU++ extends the English only NLU++ dataset to include manual translations into a range of high, medium, and low resource languages (Spanish, Marathi, Turkish and Amharic), in two domains (BANKING and HOTELS). Because of its multi-intent property, MULTI3NLU++ represents complex and natural user goals, and therefore allows us to measure the realistic performance of TOD systems in a varied set of the world's languages. We use MULTI3NLU++ to benchmark state-of-the-art multilingual models for the NLU tasks of intent detection and slot labelling for TOD systems in the multilingual setting. The results demonstrate the challenging nature of the dataset, particularly in the low-resource language setting, offering ample room for future experimentation in multi-domain multilingual TOD setups.

Raoyuan Zhao, Abdullatif Köksal, Yihong Liu et al.

Traditional benchmarking in NLP typically involves using static held-out test sets. However, this approach often results in an overestimation of performance and lacks the ability to offer comprehensive, interpretable, and dynamic assessments of NLP models. Recently, works like DynaBench (Kiela et al., 2021) and CheckList (Ribeiro et al., 2020) have addressed these limitations through behavioral testing of NLP models with test types generated by a multistep human-annotated pipeline. Unfortunately, manually creating a variety of test types requires much human labor, often at prohibitive cost. In this work, we propose SYNTHEVAL, a hybrid behavioral testing framework that leverages large language models (LLMs) to generate a wide range of test types for a comprehensive evaluation of NLP models. SYNTHEVAL first generates sentences via LLMs using controlled generation, and then identifies challenging examples by comparing the predictions made by LLMs with task-specific NLP models. In the last stage, human experts investigate the challenging examples, manually design templates, and identify the types of failures the taskspecific models consistently exhibit. We apply SYNTHEVAL to two classification tasks, sentiment analysis and toxic language detection, and show that our framework is effective in identifying weaknesses of strong models on these tasks. We share our code in https://github.com/Loreley99/SynthEval_CheckList.

Marinela Parović, Alan Ansell, Ivan Vulić et al.

Adapters have emerged as a modular and parameter-efficient approach to (zero-shot) cross-lingual transfer. The established MAD-X framework employs separate language and task adapters which can be arbitrarily combined to perform the transfer of any task to any target language. Subsequently, BAD-X, an extension of the MAD-X framework, achieves improved transfer at the cost of MAD-X's modularity by creating "bilingual" adapters specific to the source-target language pair. In this work, we aim to take the best of both worlds by (i) fine-tuning task adapters adapted to the target language(s) (so-called "target language-ready" (TLR) adapters) to maintain high transfer performance, but (ii) without sacrificing the highly modular design of MAD-X. The main idea of "target language-ready" adapters is to resolve the training-vs-inference discrepancy of MAD-X: the task adapter "sees" the target language adapter for the very first time during inference, and thus might not be fully compatible with it. We address this mismatch by exposing the task adapter to the target language adapter during training, and empirically validate several variants of the idea: in the simplest form, we alternate between using the source and target language adapters during task adapter training, which can be generalized to cycling over any set of language adapters. We evaluate different TLR-based transfer configurations with varying degrees of generality across a suite of standard cross-lingual benchmarks, and find that the most general (and thus most modular) configuration consistently outperforms MAD-X and BAD-X on most tasks and languages.

Mor Ventura, Eyal Ben-David, Anna Korhonen et al.

Text-To-Image (TTI) models, such as DALL-E and StableDiffusion, have demonstrated remarkable prompt-based image generation capabilities. Multilingual encoders may have a substantial impact on the cultural agency of these models, as language is a conduit of culture. In this study, we explore the cultural perception embedded in TTI models by characterizing culture across three hierarchical tiers: cultural dimensions, cultural domains, and cultural concepts. Based on this ontology, we derive prompt templates to unlock the cultural knowledge in TTI models, and propose a comprehensive suite of evaluation techniques, including intrinsic evaluations using the CLIP space, extrinsic evaluations with a Visual-Question-Answer (VQA) model and human assessments, to evaluate the cultural content of TTI-generated images. To bolster our research, we introduce the CulText2I dataset, derived from six diverse TTI models and spanning ten languages. Our experiments provide insights regarding Do, What, Which and How research questions about the nature of cultural encoding in TTI models, paving the way for cross-cultural applications of these models.

Huanyu Zhang, Xuehai Bai, Chengzu Li et al.

Recent generative models have achieved remarkable progress in image editing. However, existing systems and benchmarks remain largely text-guided. In contrast, human communication is inherently multimodal, where visual instructions such as sketches efficiently convey spatial and structural intent. To address this gap, we introduce VIBE, the Visual Instruction Benchmark for Image Editing with a three-level interaction hierarchy that captures deictic grounding, morphological manipulation, and causal reasoning. Across these levels, we curate high-quality and diverse test cases that reflect progressively increasing complexity in visual instruction following. We further propose a robust LMM-as-a-judge evaluation framework with task-specific metrics to enable scalable and fine-grained assessment. Through a comprehensive evaluation of 17 representative open-source and proprietary image editing models, we find that proprietary models exhibit early-stage visual instruction-following capabilities and consistently outperform open-source models. However, performance degrades markedly with increasing task difficulty even for the strongest systems, highlighting promising directions for future research.

Beiduo Chen, Tiancheng Hu, Caiqi Zhang et al. · cambridge

Reasoning-tuned LLMs utilizing long Chain-of-Thought (CoT) excel at single-answer tasks, yet their ability to model Human Label Variation--which requires capturing probabilistic ambiguity rather than resolving it--remains underexplored. We investigate this through systematic disentanglement experiments on distribution-based tasks, employing Cross-CoT experiments to isolate the effect of reasoning text from intrinsic model priors. We observe a distinct "decoupled mechanism": while CoT improves distributional alignment, final accuracy is dictated by CoT content (99% variance contribution), whereas distributional ranking is governed by model priors (over 80%). Step-wise analysis further shows that while CoT's influence on accuracy grows monotonically during the reasoning process, distributional structure is largely determined by LLM's intrinsic priors. These findings suggest that long CoT serves as a decisive LLM decision-maker for the top option but fails to function as a granular distribution calibrator for ambiguous tasks.

Yifu Qiu, Zheng Zhao, Waylon Li et al. · cambridge

Internal modelling of the world -- predicting transitions between previous states $X$ and next states $Y$ under actions $Z$ -- is essential to reasoning and planning for LLMs and VLMs. Learning such models typically requires costly action-labelled trajectories. We propose SWIRL, a self-improvement framework that learns from state-only sequences by treating actions as a latent variable and alternating between Forward World Modelling (FWM) $P_θ(Y|X,Z)$ and an Inverse Dynamics Modelling (IDM) $Q_φ(Z|X,Y)$. SWIRL iterates two phases: (1) Variational Information Maximisation, which updates the FWM to generate next states that maximise conditional mutual information with latent actions given prior states, encouraging identifiable consistency; and (2) ELBO Maximisation, which updates the IDM to explain observed transitions, effectively performing coordinate ascent. Both models are trained with reinforcement learning (specifically, GRPO) with the opposite frozen model's log-probability as a reward signal. We provide theoretical learnability guarantees for both updates, and evaluate SWIRL on LLMs and VLMs across multiple environments: single-turn and multi-turn open-world visual dynamics and synthetic textual environments for physics, web, and tool calling. SWIRL achieves gains of 16% on AURORABench, 28% on ByteMorph, 16% on WorldPredictionBench, and 14% on StableToolBench.

Songbo Hu, Yinhong Liu, Ej Zhou et al.

Creating spoken dialogue datasets is methodologically challenging, and these challenges are amplified when the goal is to build multilingual, multi-parallel datasets at scale. This work introduces HEALTHDIAL, a large-scale, multilingual, and multi-parallel dataset for developing and evaluating retrieval-augmented generation (RAG)-based spoken dialogue systems. The dataset comprises 6,000 information-seeking dialogues (1,500 per language) grounded in trusted content from the World Health Organization (WHO) and 163 hours of user speech recorded from native speakers of diverse dialects across four official WHO languages: Arabic, Chinese, English, and Spanish. Each speaker is annotated with demographic (e.g., gender, age) and sociolinguistic (e.g., primary language, region of origin) variables. We report benchmark results across key dialogue tasks, which reveal consistent performance disparities across languages, even among high-resource ones. To support future research, we release the dataset, a prototype system, and a toolkit for data collection and system evaluation.

Songbo Hu, Han Zhou, Moy Yuan et al. · cambridge

Achieving robust language technologies that can perform well across the world's many languages is a central goal of multilingual NLP. In this work, we take stock of and empirically analyse task performance disparities that exist between multilingual task-oriented dialogue (ToD) systems. We first define new quantitative measures of absolute and relative equivalence in system performance, capturing disparities across languages and within individual languages. Through a series of controlled experiments, we demonstrate that performance disparities depend on a number of factors: the nature of the ToD task at hand, the underlying pretrained language model, the target language, and the amount of ToD annotated data. We empirically prove the existence of the adaptation and intrinsic biases in current ToD systems: e.g., ToD systems trained for Arabic or Turkish using annotated ToD data fully parallel to English ToD data still exhibit diminished ToD task performance. Beyond providing a series of insights into the performance disparities of ToD systems in different languages, our analyses offer practical tips on how to approach ToD data collection and system development for new languages.

Elwin Huaman, Adrian Gamarra Lafuente, Johanna Cordova et al.

The preservation of under-resourced languages requires digital tools and resources shaped by and for their speakers. We present the first dedicated ASR resources for Puno Quechua (ISO 639-3: qxp): (1) the largest speech corpus for any single Quechua variety, consisting in 66 hours of recordings for scripted and spontaneous speech (including 36 hours of manually transcribed and validated data), collected via a participatory design campaign; (2) the first systematic ASR benchmark for Puno Quechua, evaluating state-of-the-art models and fine-tuning Whisper-base, wav2vec2-base, and XLS-R-300M, with and without continued pre-training (CPT); (3) an open release of all datasets and fine-tuned models.

Beiduo Chen, Pingjun Hong, Ziyun Zhang et al.

Free-text explanations extend human label variation (HLV) beyond label disagreement by revealing the reasoning and preferences behind annotators' decisions. We study whether large language models (LLMs) can learn and reproduce such annotator-specific label-explanation behavior. Using two sentence-pair tasks with four annotators each -- natural language inference and paraphrase judgment -- we first analyze whether annotators exhibit stable individual patterns. We find that such patterns are weak at the single-annotation level due to strong input-content effects, but become detectable after input-content reduction and annotator-level aggregation. We then compare prompting and supervised fine-tuning (SFT) baselines and propose cross-annotator preference optimization (CAPO), which contrasts a target annotator's response with other valid but less target-specific annotations for the same input. Experiments show that prompting is limited and unstable, SFT better captures annotator-specific behavior, and CAPO further improves aggregation-aware imitation and judge-based attribution while preserving target-specific reasoning patterns under human validation. Overall, our results show that HLV can be learned as annotator-specific label-explanation behavior, suggesting a path toward scalable explanation-based annotation grounded in annotator histories rather than labels alone.

Lester James V. Miranda, Ivan Vulić, Anna Korhonen · cambridge

Synthesizing supervised finetuning (SFT) data from language models (LMs) to teach smaller models multilingual tasks has become increasingly common. However, teacher model selection is often ad hoc, typically defaulting to the largest available option, even though such models may have significant capability gaps in non-English languages. This practice can result in poor-quality synthetic data and suboptimal student downstream performance. In this work, we systematically characterize what makes an effective multilingual teacher. We measure intrinsic measures of data quality with extrinsic student model performance in a metric we call Polyglot Score; evaluating 10 LMs across 6 typologically diverse languages, generating over 1.4M SFT examples and training 240 student models. Among the models tested, Gemma 3 27B and Aya Expanse 32B emerge as consistently effective teachers across different student base model families. Further analyses reveal that model scale alone does not significantly predict teacher effectiveness; instead, data qualities such as prompt diversity, length, and response fluency capture over 93.3% of variance in intrinsic data quality and predict student performance. Finally, we provide practical recommendations, including matching the model families of teacher-student pairs and translating from or responding to existing prompts, which can yield improvements for less-resourced languages. We hope that our work advances data-centric research in multilingual synthetic data and LM development.

Anjali Kantharuban, Ivan Vulić, Anna Korhonen

Historically, researchers and consumers have noticed a decrease in quality when applying NLP tools to minority variants of languages (i.e. Puerto Rican Spanish or Swiss German), but studies exploring this have been limited to a select few languages. Additionally, past studies have mainly been conducted in a monolingual context, so cross-linguistic trends have not been identified and tied to external factors. In this work, we conduct a comprehensive evaluation of the most influential, state-of-the-art large language models (LLMs) across two high-use applications, machine translation and automatic speech recognition, to assess their functionality on the regional dialects of several high- and low-resource languages. Additionally, we analyze how the regional dialect gap is correlated with economic, social, and linguistic factors. The impact of training data, including related factors like dataset size and its construction procedure, is shown to be significant but not consistent across models or languages, meaning a one-size-fits-all approach cannot be taken in solving the dialect gap. This work will lay the foundation for furthering the field of dialectal NLP by laying out evident disparities and identifying possible pathways for addressing them through mindful data collection.

Yifu Qiu, Zheng Zhao, Yftah Ziser et al. · cambridge

Large language models (LLMs) often exhibit undesirable behaviours, such as generating untruthful or biased content. Editing their internal representations has been shown to be effective in mitigating such behaviours on top of the existing alignment methods. We propose a novel inference-time editing method, namely spectral editing of activations (SEA), to project the input representations into directions with maximal covariance with the positive demonstrations (e.g., truthful) while minimising covariance with the negative demonstrations (e.g., hallucinated). We also extend our method to non-linear editing using feature functions. We run extensive experiments on benchmarks concerning truthfulness and bias with six open-source LLMs of different sizes and model families. The results demonstrate the superiority of SEA in effectiveness, generalisation to similar tasks, as well as computation and data efficiency. We also show that SEA editing only has a limited negative impact on other model capabilities.

Alan Ansell, Ivan Vulić, Hannah Sterz et al.

Large Language Models (LLMs) are difficult to fully fine-tune (e.g., with instructions or human feedback) due to their sheer number of parameters. A family of parameter-efficient sparse fine-tuning methods have proven promising in terms of performance but their memory requirements increase proportionally to the size of the LLMs. In this work, we scale sparse fine-tuning to state-of-the-art LLMs like LLaMA 2 7B and 13B. We propose SpIEL, a novel sparse fine-tuning method which, for a desired density level, maintains an array of parameter indices and the deltas of these parameters relative to their pretrained values. It iterates over: (a) updating the active deltas, (b) pruning indices (based on the change of magnitude of their deltas) and (c) regrowth of indices. For regrowth, we explore two criteria based on either the accumulated gradients of a few candidate parameters or their approximate momenta estimated using the efficient SM3 optimizer. We experiment with instruction-tuning of LLMs on standard dataset mixtures, finding that SpIEL is often superior to popular parameter-efficient fine-tuning methods like LoRA (low-rank adaptation) in terms of performance and comparable in terms of run time. We additionally show that SpIEL is compatible with both quantization and efficient optimizers, to facilitate scaling to ever-larger model sizes. We release the code for SpIEL at https://github.com/AlanAnsell/peft and for the instruction-tuning experiments at https://github.com/ducdauge/sft-llm.

Evgeniia Razumovskaia, Goran Glavaš, Anna Korhonen et al.

Task-oriented dialogue (ToD) systems help users execute well-defined tasks across a variety of domains (e.g., $\textit{flight booking}$ or $\textit{food ordering}$), with their Natural Language Understanding (NLU) components being dedicated to the analysis of user utterances, predicting users' intents ($\textit{Intent Detection}$, ID) and extracting values for informational slots ($\textit{Value Extraction}$, VE). In most domains, labelled NLU data is scarce, making sample-efficient learning -- enabled with effective transfer paradigms -- paramount. In this work, we introduce SQATIN, a new framework for dialog NLU based on (i) instruction tuning and (ii) question-answering-based formulation of ID and VE tasks. According to the evaluation on established NLU benchmarks, SQATIN sets the new state of the art in dialogue NLU, substantially surpassing the performance of current models based on standard fine-tuning objectives in both in-domain training and cross-domain transfer. SQATIN yields particularly large performance gains in cross-domain transfer, owing to the fact that our QA-based instruction tuning leverages similarities between natural language descriptions of classes (i.e., slots and intents) across domains.

Shun Shao, Binxu Wang, Shay B. Cohen et al.

Mechanistic interpretability has made it possible to localize circuits underlying specific behaviors in language models, but existing methods are expensive, model-specific, and difficult to scale to larger architectures. We introduce \textbf{Differentiable Faithfulness Alignment (DFA)}, a framework that transfers circuit information from a smaller source model to a larger target model through a learned differentiable alignment. DFA projects source-model node importance scores into the target model and trains this mapping with a soft faithfulness objective, avoiding full circuit discovery on the target model. We evaluate DFA on Llama-3 and Qwen-2.5 across six tasks spanning factual retrieval, multiple-choice reasoning, and arithmetic. The strongest results occur on Llama-3 $1$B$\rightarrow3$B, where aligned circuits are often competitive with direct node attribution and zero-shot transfer remains effective. Recovery weakens for larger source--target gaps and is substantially lower on Qwen-2.5, suggesting that transfer becomes harder as architectural and scaling differences increase. Overall, DFA consistently outperforms simple baselines and, in some settings, recovers target-model circuits with faithfulness comparable to or stronger than direct attribution. These results suggest that smaller models can provide useful mechanistic priors for larger ones, while highlighting both the promise and the limits of node-level cross-model circuit alignment.\footnote{Code is available at https://github.com/jasonshaoshun/dfa-circuits.

Han Zhou, Xingchen Wan, Ivan Vulić et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has advanced the reasoning capability of large language models (LLMs), enabling autonomous agents that can conduct effective multi-turn and tool-integrated reasoning. While instructions serve as the primary protocol for defining agents, RLVR typically relies on static and manually designed instructions. However, those instructions may be suboptimal for the base model, and the optimal instruction may change as the agent's policy improves and explores the interaction with the environment. To bridge the gap, we introduce INSPO, a novel Instruction-Policy co-evolution framework that integrates instruction optimization as a dynamic component of the reinforcement learning (RL) loop. INSPO maintains a dynamic population of instruction candidates that are sampled with questions, where reward signals in RL loops are automatically attributed to each instruction, and low performers are periodically pruned. New instructions are generated and verified through an on-policy reflection mechanism, where an LLM-based optimizer analyzes past experience from a replay buffer and evolves more effective strategies given the current policy. We conduct extensive experiments on multi-turn retrieval and reasoning tasks, demonstrating that INSPO substantially outperforms strong baselines relying on static instructions. INSPO discovers innovative instructions that guide the agent toward more strategic reasoning paths, achieving substantial performance gains with only a marginal increase in computational overhead.

Giulia Occhini, Kumiko Tanaka-Ishii, Anna Barford et al.

Artificial intelligence (AI) has the potential to transform healthcare, education, governance and socioeconomic equity, but its benefits remain concentrated in a small number of languages (Bender, 2019; Blasi et al., 2022; Joshi et al., 2020; Ranathunga and de Silva, 2022; Young, 2015). Language AI - the technologies that underpin widely-used conversational systems such as ChatGPT - could provide major benefits if available in people's native languages, yet most of the world's 7,000+ linguistic communities currently lack access and face persistent digital marginalization. Here we present a global longitudinal analysis of social, economic and infrastructural conditions across languages to assess systemic inequalities in language AI. We first analyze the existence of AI resources for 6003 languages. We find that despite efforts of the community to broaden the reach of language technologies (Bapna et al., 2022; Costa-Jussà et al., 2022), the dominance of a handful of languages is exacerbating disparities on an unprecedented scale, with divides widening exponentially rather than narrowing. Further, we contrast the longitudinal diffusion of AI with that of earlier IT technologies, revealing a distinctive hype-driven pattern of spread. To translate our findings into practical insights and guide prioritization efforts, we introduce the Language AI Readiness Index (EQUATE), which maps the state of technological, socio-economic, and infrastructural prerequisites for AI deployment across languages. The index highlights communities where capacity exists but remains underutilized, and provides a framework for accelerating more equitable diffusion of language AI. Our work contributes to setting the baseline for a transition towards more sustainable and equitable language technologies.

Songbo Hu, Xiaobin Wang, Zhangdie Yuan et al. · cambridge

We present DIALIGHT, a toolkit for developing and evaluating multilingual Task-Oriented Dialogue (ToD) systems which facilitates systematic evaluations and comparisons between ToD systems using fine-tuning of Pretrained Language Models (PLMs) and those utilising the zero-shot and in-context learning capabilities of Large Language Models (LLMs). In addition to automatic evaluation, this toolkit features (i) a secure, user-friendly web interface for fine-grained human evaluation at both local utterance level and global dialogue level, and (ii) a microservice-based backend, improving efficiency and scalability. Our evaluations reveal that while PLM fine-tuning leads to higher accuracy and coherence, LLM-based systems excel in producing diverse and likeable responses. However, we also identify significant challenges of LLMs in adherence to task-specific instructions and generating outputs in multiple languages, highlighting areas for future research. We hope this open-sourced toolkit will serve as a valuable resource for researchers aiming to develop and properly evaluate multilingual ToD systems and will lower, currently still high, entry barriers in the field.

Yinhong Liu, Han Zhou, Zhijiang Guo et al. · cambridge

Large Language Models (LLMs) have demonstrated promising capabilities as automatic evaluators in assessing the quality of generated natural language. However, LLMs still exhibit biases in evaluation and often struggle to generate coherent evaluations that align with human assessments. In this work, we first conduct a systematic study of the misalignment between LLM evaluators and human evaluation, revealing that existing calibration methods aimed at mitigating biases of LLMs are insufficient for effectively aligning LLM evaluators. Inspired by the use of preference data in RLHF, we formulate the evaluation as a ranking problem and introduce Pairwise-preference Search (PAIRS), an uncertainty-guided search-based rank aggregation method that employs LLMs to conduct pairwise comparisons locally and efficiently ranks candidate texts globally. PAIRS achieves state-of-the-art performance on representative evaluation tasks in long-form generations and demonstrates significant improvements over direct scoring. Furthermore, we provide insights into the role of pairwise preference in quantifying the transitivity of LLMs and demonstrate how PAIRS benefits from calibration using debiased pairwise evaluations.

Felicia Körner, Max Müller-Eberstein, Anna Korhonen et al.

Training Large Language Models (LLMs) with high multilingual coverage is becoming increasingly important -- especially when monolingual resources are scarce. Recent studies have found that LLMs process multilingual inputs in shared concept spaces, thought to support generalization and cross-lingual transfer. However, these prior studies often do not use causal methods, lack deeper error analysis or focus on the final model only, leaving open how these spaces emerge during training. We investigate the development of language-agnostic concept spaces during pretraining of EuroLLM through the causal interpretability method of activation patching. We isolate cross-lingual concept representations, then inject them into a translation prompt to investigate how consistently translations can be altered, independently of the language. We find that shared concept spaces emerge early} and continue to refine, but that alignment with them is language-dependent}. Furthermore, in contrast to prior work, our fine-grained manual analysis reveals that some apparent gains in translation quality reflect shifts in behavior -- like selecting senses for polysemous words or translating instead of copying cross-lingual homographs -- rather than improved translation ability. Our findings offer new insight into the training dynamics of cross-lingual alignment and the conditions under which causal interpretability methods offer meaningful insights in multilingual contexts.

Yifu Qiu, Yftah Ziser, Anna Korhonen et al.

Hallucinations pose a significant challenge to the reliability of neural models for abstractive summarisation. While automatically generated summaries may be fluent, they often lack faithfulness to the original document. This issue becomes even more pronounced in low-resource settings, such as cross-lingual transfer. With the existing faithful metrics focusing on English, even measuring the extent of this phenomenon in cross-lingual settings is hard. To address this, we first develop a novel metric, mFACT, evaluating the faithfulness of non-English summaries, leveraging translation-based transfer from multiple English faithfulness metrics. We then propose a simple but effective method to reduce hallucinations with a cross-lingual transfer, which weighs the loss of each training example by its faithfulness score. Through extensive experiments in multiple languages, we demonstrate that mFACT is the metric that is most suited to detect hallucinations. Moreover, we find that our proposed loss weighting method drastically increases both performance and faithfulness according to both automatic and human evaluation when compared to strong baselines for cross-lingual transfer such as MAD-X. Our code and dataset are available at https://github.com/yfqiu-nlp/mfact-summ.

Abdullatif Köksal, Omer Faruk Yalcin, Ahmet Akbiyik et al.

Pretrained language models (PLMs) are key components in NLP, but they contain strong social biases. Quantifying these biases is challenging because current methods focusing on fill-the-mask objectives are sensitive to slight changes in input. To address this, we propose a bias probing technique called LABDet, for evaluating social bias in PLMs with a robust and language-agnostic method. For nationality as a case study, we show that LABDet `surfaces' nationality bias by training a classifier on top of a frozen PLM on non-nationality sentiment detection. We find consistent patterns of nationality bias across monolingual PLMs in six languages that align with historical and political context. We also show for English BERT that bias surfaced by LABDet correlates well with bias in the pretraining data; thus, our work is one of the few studies that directly links pretraining data to PLM behavior. Finally, we verify LABDet's reliability and applicability to different templates and languages through an extensive set of robustness checks. We publicly share our code and dataset in https://github.com/akoksal/LABDet.

Alan Ansell, Edoardo Maria Ponti, Anna Korhonen et al.

Fine-tuning the entire set of parameters of a large pretrained model has become the mainstream approach for transfer learning. To increase its efficiency and prevent catastrophic forgetting and interference, techniques like adapters and sparse fine-tuning have been developed. Adapters are modular, as they can be combined to adapt a model towards different facets of knowledge (e.g., dedicated language and/or task adapters). Sparse fine-tuning is expressive, as it controls the behavior of all model components. In this work, we introduce a new fine-tuning method with both these desirable properties. In particular, we learn sparse, real-valued masks based on a simple variant of the Lottery Ticket Hypothesis. Task-specific masks are obtained from annotated data in a source language, and language-specific masks from masked language modeling in a target language. Both these masks can then be composed with the pretrained model. Unlike adapter-based fine-tuning, this method neither increases the number of parameters at inference time nor alters the original model architecture. Most importantly, it outperforms adapters in zero-shot cross-lingual transfer by a large margin in a series of multilingual benchmarks, including Universal Dependencies, MasakhaNER, and AmericasNLI. Based on an in-depth analysis, we additionally find that sparsity is crucial to prevent both 1) interference between the fine-tunings to be composed and 2) overfitting. We release the code and models at https://github.com/cambridgeltl/composable-sft.

Edoardo Maria Ponti, Goran Glavaš, Olga Majewska et al.

In order to simulate human language capacity, natural language processing systems must be able to reason about the dynamics of everyday situations, including their possible causes and effects. Moreover, they should be able to generalise the acquired world knowledge to new languages, modulo cultural differences. Advances in machine reasoning and cross-lingual transfer depend on the availability of challenging evaluation benchmarks. Motivated by both demands, we introduce Cross-lingual Choice of Plausible Alternatives (XCOPA), a typologically diverse multilingual dataset for causal commonsense reasoning in 11 languages, which includes resource-poor languages like Eastern Apurímac Quechua and Haitian Creole. We evaluate a range of state-of-the-art models on this novel dataset, revealing that the performance of current methods based on multilingual pretraining and zero-shot fine-tuning falls short compared to translation-based transfer. Finally, we propose strategies to adapt multilingual models to out-of-sample resource-lean languages where only a small corpus or a bilingual dictionary is available, and report substantial improvements over the random baseline. The XCOPA dataset is freely available at github.com/cambridgeltl/xcopa.

Edoardo M. Ponti, Ivan Vulić, Ryan Cotterell et al.

Most combinations of NLP tasks and language varieties lack in-domain examples for supervised training because of the paucity of annotated data. How can neural models make sample-efficient generalizations from task-language combinations with available data to low-resource ones? In this work, we propose a Bayesian generative model for the space of neural parameters. We assume that this space can be factorized into latent variables for each language and each task. We infer the posteriors over such latent variables based on data from seen task-language combinations through variational inference. This enables zero-shot classification on unseen combinations at prediction time. For instance, given training data for named entity recognition (NER) in Vietnamese and for part-of-speech (POS) tagging in Wolof, our model can perform accurate predictions for NER in Wolof. In particular, we experiment with a typologically diverse sample of 33 languages from 4 continents and 11 families, and show that our model yields comparable or better results than state-of-the-art, zero-shot cross-lingual transfer methods. Moreover, we demonstrate that approximate Bayesian model averaging results in smoother predictive distributions, whose entropy inversely correlates with accuracy. Hence, the proposed framework also offers robust estimates of prediction uncertainty. Our code is located at github.com/cambridgeltl/parameter-factorization

Panagiotis Fytas, Anna Breger, Ian Selby et al.

Developing imaging models capable of detecting pathologies from chest X-rays can be cost and time-prohibitive for large datasets as it requires supervision to attain state-of-the-art performance. Instead, labels extracted from radiology reports may serve as distant supervision since these are routinely generated as part of clinical practice. Despite their widespread use, current rule-based methods for label extraction rely on extensive rule sets that are limited in their robustness to syntactic variability. To alleviate these limitations, we introduce RadPert, a rule-based system that integrates an uncertainty-aware information schema with a streamlined set of rules, enhancing performance. Additionally, we have developed RadPrompt, a multi-turn prompting strategy that leverages RadPert to bolster the zero-shot predictive capabilities of large language models, achieving a statistically significant improvement in weighted average F1 score over GPT-4 Turbo. Most notably, RadPrompt surpasses both its underlying models, showcasing the synergistic potential of LLMs with rule-based models. We have evaluated our methods on two English Corpora: the MIMIC-CXR gold-standard test set and a gold-standard dataset collected from the Cambridge University Hospitals.

Kim Zierahn, Cristina Cachero, Anna Korhonen et al.

A growing body of research examines personality traits in Large Language Models (LLMs), particularly in human-agent collaboration. Prior work has frequently applied the Big Five inventory to assess LLM behavior analogous to human personality, without questioning the underlying assumptions. This paper critically evaluates whether LLM responses to personality tests satisfy six defining characteristics of personality. We find that none are fully met, indicating that such assessments do not measure a construct equivalent to human personality. We propose a research agenda for shifting from anthropomorphic trait attribution toward functional evaluations, clarifying what personality tests actually capture in LLMs and developing LLM-specific frameworks for characterizing stable, intrinsic behavior.

Han Zhou, Xingchen Wan, Ruoxi Sun et al. · cambridge

Large language models, employed as multiple agents that interact and collaborate with each other, have excelled at solving complex tasks. The agents are programmed with prompts that declare their functionality, along with the topologies that orchestrate interactions across agents. Designing prompts and topologies for multi-agent systems (MAS) is inherently complex. To automate the entire design process, we first conduct an in-depth analysis of the design space aiming to understand the factors behind building effective MAS. We reveal that prompts together with topologies play critical roles in enabling more effective MAS design. Based on the insights, we propose Multi-Agent System Search (MASS), a MAS optimization framework that efficiently exploits the complex MAS design space by interleaving its optimization stages, from local to global, from prompts to topologies, over three stages: 1) block-level (local) prompt optimization; 2) workflow topology optimization; 3) workflow-level (global) prompt optimization, where each stage is conditioned on the iteratively optimized prompts/topologies from former stages. We show that MASS-optimized multi-agent systems outperform a spectrum of existing alternatives by a substantial margin. Based on the MASS-found systems, we finally propose design principles behind building effective multi-agent systems.

Yaoyiran Li, Xiang Zhai, Moustafa Alzantot et al. · cambridge

Traditional recommender systems such as matrix factorization methods have primarily focused on learning a shared dense embedding space to represent both items and user preferences. Subsequently, sequence models such as RNN, GRUs, and, recently, Transformers have emerged and excelled in the task of sequential recommendation. This task requires understanding the sequential structure present in users' historical interactions to predict the next item they may like. Building upon the success of Large Language Models (LLMs) in a variety of tasks, researchers have recently explored using LLMs that are pretrained on vast corpora of text for sequential recommendation. To use LLMs for sequential recommendation, both the history of user interactions and the model's prediction of the next item are expressed in text form. We propose CALRec, a two-stage LLM finetuning framework that finetunes a pretrained LLM in a two-tower fashion using a mixture of two contrastive losses and a language modeling loss: the LLM is first finetuned on a data mixture from multiple domains followed by another round of target domain finetuning. Our model significantly outperforms many state-of-the-art baselines (+37% in Recall@1 and +24% in NDCG@10) and our systematic ablation studies reveal that (i) both stages of finetuning are crucial, and, when combined, we achieve improved performance, and (ii) contrastive alignment is effective among the target domains explored in our experiments.

Chengzu Li, Han Zhou, Goran Glavaš et al. · cambridge

When adapting ICL with or without fine-tuning, we are curious about whether the instruction-tuned language model is able to achieve well-calibrated results without suffering from the problem of overconfidence (i.e., miscalibration) considering its strong instruction following ability, especially in such limited data setups. In this work, we deliver an in-depth analysis of the behavior across different choices of learning methods from the perspective of both performance and calibration. Through extensive controlled experiments, we observe that the miscalibration problem exists across all learning methods in low-resource setups. To achieve simultaneous gain for both in-task performance and calibration, we then study the potential of self-ensembling applied at different modeling stages (e.g., variations of in-context examples or variations in prompts or different ensembling strategies) to make the predictions more calibrated and have comparable or even better performance. We find that self-ensembling with max probability produces robust and calibrated predictions. Our work reveals the potential calibration problem of using ICL despite the improvements in task performance and sheds light on which learning paradigm to choose. We also provide practical guidelines for choosing learning paradigms depending on whether the data has been seen by the model before and a worthwhile solution via self-ensembling on how to enhance both task performance and calibration of LMs, which we hope could encourage further study.

Lester James V. Miranda, Songbo Hu, Roi Reichart et al.

Where and how language models (LMs) are deployed determines who can benefit from them. However, there are several challenges that prevent effective deployment of LMs in non-English-speaking and hardware constrained communities in the Global South. We call this challenge the last mile: the intersection of multilinguality and edge deployment, where the goals are aligned but the technical requirements often compete. Studying these two fields together is both a need, as linguistically diverse communities often face the most severe infrastructure constraints, and an opportunity, as edge and multilingual NLP research remain largely siloed. To understand the state of the art and the challenges of combining the two areas, we survey 232 papers that tackle this problem across the language modelling pipeline, from data collection to development and deployment. We also discuss open questions and provide actionable recommendations for different stakeholders in the NLP ecosystem. Finally, we hope that this work contributes to the development of inclusive and equitable language technologies.

Evgeniia Razumovskaia, Ivan Vulić, Anna Korhonen

Supervised fine-tuning (SFT), supervised instruction tuning (SIT) and in-context learning (ICL) are three alternative, de facto standard approaches to few-shot learning. ICL has gained popularity recently with the advent of LLMs due to its simplicity and sample efficiency. Prior research has conducted only limited investigation into how these approaches work for multilingual few-shot learning, and the focus so far has been mostly on their performance. In this work, we present an extensive and systematic comparison of the three approaches, testing them on 6 high- and low-resource languages, three different NLU tasks, and a myriad of language and domain setups. Importantly, performance is only one aspect of the comparison, where we also analyse the approaches through the optics of their computational, inference and financial costs. Our observations show that supervised instruction tuning has the best trade-off between performance and resource requirements. As another contribution, we analyse the impact of target language adaptation of pretrained LLMs and find that the standard adaptation approaches can (superficially) improve target language generation capabilities, but language understanding elicited through ICL does not improve and remains limited, with low scores especially for low-resource languages.

Yaoyiran Li, Anna Korhonen, Ivan Vulić · cambridge

Recent work has shown that, while large language models (LLMs) demonstrate strong word translation or bilingual lexicon induction (BLI) capabilities in few-shot setups, they still cannot match the performance of 'traditional' mapping-based approaches in the unsupervised scenario where no seed translation pairs are available, especially for lower-resource languages. To address this challenge with LLMs, we propose self-augmented in-context learning (SAIL) for unsupervised BLI: starting from a zero-shot prompt, SAIL iteratively induces a set of high-confidence word translation pairs for in-context learning (ICL) from an LLM, which it then reapplies to the same LLM in the ICL fashion. Our method shows substantial gains over zero-shot prompting of LLMs on two established BLI benchmarks spanning a wide range of language pairs, also outperforming mapping-based baselines across the board. In addition to achieving state-of-the-art unsupervised BLI performance, we also conduct comprehensive analyses on SAIL and discuss its limitations.