Constanza Fierro

Daniel Hershcovich, Stella Frank, Heather Lent et al.

Various efforts in the Natural Language Processing (NLP) community have been made to accommodate linguistic diversity and serve speakers of many different languages. However, it is important to acknowledge that speakers and the content they produce and require, vary not just by language, but also by culture. Although language and culture are tightly linked, there are important differences. Analogous to cross-lingual and multilingual NLP, cross-cultural and multicultural NLP considers these differences in order to better serve users of NLP systems. We propose a principled framework to frame these efforts, and survey existing and potential strategies.

Jiaang Li, Yova Kementchedjhieva, Constanza Fierro et al.

Large-scale pretrained language models (LMs) are said to ``lack the ability to connect utterances to the world'' (Bender and Koller, 2020), because they do not have ``mental models of the world' '(Mitchell and Krakauer, 2023). If so, one would expect LM representations to be unrelated to representations induced by vision models. We present an empirical evaluation across four families of LMs (BERT, GPT-2, OPT and LLaMA-2) and three vision model architectures (ResNet, SegFormer, and MAE). Our experiments show that LMs partially converge towards representations isomorphic to those of vision models, subject to dispersion, polysemy and frequency. This has important implications for both multi-modal processing and the LM understanding debate (Mitchell and Krakauer, 2023).

Constanza Fierro, Anders Søgaard

Pretrained language models can be queried for factual knowledge, with potential applications in knowledge base acquisition and tasks that require inference. However, for that, we need to know how reliable this knowledge is, and recent work has shown that monolingual English language models lack consistency when predicting factual knowledge, that is, they fill-in-the-blank differently for paraphrases describing the same fact. In this paper, we extend the analysis of consistency to a multilingual setting. We introduce a resource, mParaRel, and investigate (i) whether multilingual language models such as mBERT and XLM-R are more consistent than their monolingual counterparts; and (ii) if such models are equally consistent across languages. We find that mBERT is as inconsistent as English BERT in English paraphrases, but that both mBERT and XLM-R exhibit a high degree of inconsistency in English and even more so for all the other 45 languages.

Constanza Fierro, Reinald Kim Amplayo, Fantine Huot et al.

The increasing demand for the deployment of LLMs in information-seeking scenarios has spurred efforts in creating verifiable systems, which generate responses to queries along with supporting evidence. In this paper, we explore the attribution capabilities of plan-based models which have been recently shown to improve the faithfulness, grounding, and controllability of generated text. We conceptualize plans as a sequence of questions which serve as blueprints of the generated content and its organization. We propose two attribution models that utilize different variants of blueprints, an abstractive model where questions are generated from scratch, and an extractive model where questions are copied from the input. Experiments on long-form question-answering show that planning consistently improves attribution quality. Moreover, the citations generated by blueprint models are more accurate compared to those obtained from LLM-based pipelines lacking a planning component.

Constanza Fierro, Fabien Roger

Providing high-quality feedback to Large Language Models (LLMs) on a diverse training distribution can be difficult and expensive, and providing feedback only on a narrow distribution can result in unintended generalizations. To better leverage narrow training data, we propose contrastive weight steering, a simple post-training method that edits the model parameters using weight arithmetic. We isolate a behavior direction in weight-space by subtracting the weight deltas from two small fine-tunes -- one that induces the desired behavior and another that induces its opposite -- and then add or remove this direction to modify the model's weights. We apply this technique to mitigate sycophancy and induce misalignment, and find that weight steering often generalizes further than activation steering, achieving stronger out-of-distribution behavioral control before degrading general capabilities. We also show that, in the context of task-specific fine-tuning, weight steering can partially mitigate undesired behavioral drift: it can reduce sycophancy and under-refusals introduced during fine-tuning while preserving task performance gains. Finally, we provide preliminary evidence that emergent misalignment can be detected by measuring the similarity between fine-tuning updates and an "evil" weight direction, suggesting that it may be possible to monitor the evolution of weights during training and detect rare misaligned behaviors that never manifest during training or evaluations.

Constanza Fierro, Negar Foroutan, Desmond Elliott et al.

Large Language Models (LLMs) store and retrieve vast amounts of factual knowledge acquired during pre-training. Prior research has localized and identified mechanisms behind knowledge recall; however, it has only focused on English monolingual models. The question of how these mechanisms generalize to non-English languages and multilingual LLMs remains unexplored. In this paper, we address this gap by conducting a comprehensive analysis of three multilingual LLMs. First, we show that previously identified recall mechanisms in English largely apply to multilingual contexts, with nuances based on language and architecture. Next, through patching intermediate representations, we localize the role of language during recall, finding that subject enrichment is language-independent, while object extraction is language-dependent. Additionally, we discover that the last token representation acts as a Function Vector (FV), encoding both the language of the query and the content to be extracted from the subject. Furthermore, in decoder-only LLMs, FVs compose these two pieces of information in two separate stages. These insights reveal unique mechanisms in multilingual LLMs for recalling information, highlighting the need for new methodologies -- such as knowledge evaluation, fact editing, and knowledge acquisition -- that are specifically tailored for multilingual LLMs.

Constanza Fierro, Nicolas Garneau, Emanuele Bugliarello et al.

Facts are subject to contingencies and can be true or false in different circumstances. One such contingency is time, wherein some facts mutate over a given period, e.g., the president of a country or the winner of a championship. Trustworthy language models ideally identify mutable facts as such and process them accordingly. We create MuLan, a benchmark for evaluating the ability of English language models to anticipate time-contingency, covering both 1:1 and 1:N relations. We hypothesize that mutable facts are encoded differently than immutable ones, hence being easier to update. In a detailed evaluation of six popular large language models, we consistently find differences in the LLMs' confidence, representations, and update behavior, depending on the mutability of a fact. Our findings should inform future work on the injection of and induction of time-contingent knowledge to/from LLMs.

Constanza Fierro, Jiaang Li, Anders Søgaard

The purpose of instruction tuning is enabling zero-shot performance, but instruction tuning has also been shown to improve chain-of-thought reasoning and value alignment (Si et al., 2023). Here we consider the impact on $\textit{consistency}$, i.e., the sensitivity of language models to small perturbations in the input. We compare 10 instruction-tuned LLaMA models to the original LLaMA-7b model and show that almost across-the-board they become more consistent, both in terms of their representations and their predictions in zero-shot and downstream tasks. We explain these improvements through mechanistic analyses of factual recall.

Iwan Williams, Ninell Oldenburg, Ruchira Dhar et al.

Mechanistic interpretability (MI) aims to explain how neural networks work by uncovering their underlying causal mechanisms. As the field grows in influence, it is increasingly important to examine not just models themselves, but the assumptions, concepts and explanatory strategies implicit in MI research. We argue that mechanistic interpretability needs philosophy: not as an afterthought, but as an ongoing partner in clarifying its concepts, refining its methods, and assessing the epistemic and ethical stakes of interpreting AI systems. Taking three open problems from the MI literature as examples, this position paper illustrates the value philosophy can add to MI research, and outlines a path toward deeper interdisciplinary dialogue.

Hans Peter Lynsgøe Raaschou-jensen, Constanza Fierro, Anders Søgaard

Recent advances in reasoning language models -- particularly those that use long, latent chains of thought -- have demonstrated remarkable capabilities in complex, agentic tasks. However, as these models operate over increasingly extended time horizons, their internal progress becomes opaque to users, complicating expectation management and real-time oversight. In this work, we investigate whether real-time progress prediction is feasible. We discretize progress and train a linear probe to classify reasoning states. We then introduce a two-stage fine-tuning approach that enables reasoning models to generate progress estimates (0$\rightarrow$100\%) during inference. Our best fine-tuned model achieves an average error of 10\% for sequences less than 16,000 tokens, offering a practical mechanism for monitoring and interpreting model reasoning in real time.

Fantine Huot, Joshua Maynez, Chris Alberti et al.

Cross-lingual summarization consists of generating a summary in one language given an input document in a different language, allowing for the dissemination of relevant content across speakers of other languages. The task is challenging mainly due to the paucity of cross-lingual datasets and the compounded difficulty of summarizing and translating. This work presents $μ$PLAN, an approach to cross-lingual summarization that uses an intermediate planning step as a cross-lingual bridge. We formulate the plan as a sequence of entities capturing the summary's content and the order in which it should be communicated. Importantly, our plans abstract from surface form: using a multilingual knowledge base, we align entities to their canonical designation across languages and generate the summary conditioned on this cross-lingual bridge and the input. Automatic and human evaluation on the XWikis dataset (across four language pairs) demonstrates that our planning objective achieves state-of-the-art performance in terms of informativeness and faithfulness. Moreover, $μ$PLAN models improve the zero-shot transfer to new cross-lingual language pairs compared to baselines without a planning component.

Constanza Fierro, Jorge Pérez, Javier Mora

Deep learning techniques have been successfully applied to predict unplanned readmissions of patients in medical centers. The training data for these models is usually based on historical medical records that contain a significant amount of free-text from admission reports, referrals, exam notes, etc. Most of the models proposed so far are tailored to English text data and assume that electronic medical records follow standards common in developed countries. These two characteristics make them difficult to apply in developing countries that do not necessarily follow international standards for registering patient information, or that store text information in languages other than English. In this paper we propose a deep learning architecture for predicting unplanned readmissions that consumes data that is significantly less structured compared with previous models in the literature. We use it to present the first results for this task in a large clinical dataset that mainly contains Spanish text data. The dataset is composed of almost 10 years of records in a Chilean medical center. On this dataset, our model achieves results that are comparable to some of the most recent results obtained in US medical centers for the same task (0.76 AUROC).