Javier Ferrando

Marta R. Costa-jussà, Eric Smith, Christophe Ropers et al. · meta-ai

Machine Translation systems can produce different types of errors, some of which are characterized as critical or catastrophic due to the specific negative impact that they can have on users. In this paper we focus on one type of critical error: added toxicity. We evaluate and analyze added toxicity when translating a large evaluation dataset (HOLISTICBIAS, over 472k sentences, covering 13 demographic axes) from English into 164 languages. An automatic toxicity evaluation shows that added toxicity across languages varies from 0% to 5%. The output languages with the most added toxicity tend to be low-resource ones, and the demographic axes with the most added toxicity include sexual orientation, gender and sex, and ability. We also perform human evaluation on a subset of 8 translation directions, confirming the prevalence of true added toxicity. We use a measurement of the amount of source contribution to the translation, where a low source contribution implies hallucination, to interpret what causes toxicity. Making use of the input attributions allows us to explain toxicity, because the source contributions significantly correlate with toxicity for 84% of languages studied. Given our findings, our recommendations to reduce added toxicity are to curate training data to avoid mistranslations, mitigate hallucination and check unstable translations.

Javier Ferrando, Gerard I. Gállego, Marta R. Costa-jussà

The Transformer architecture aggregates input information through the self-attention mechanism, but there is no clear understanding of how this information is mixed across the entire model. Additionally, recent works have demonstrated that attention weights alone are not enough to describe the flow of information. In this paper, we consider the whole attention block -- multi-head attention, residual connection, and layer normalization -- and define a metric to measure token-to-token interactions within each layer. Then, we aggregate layer-wise interpretations to provide input attribution scores for model predictions. Experimentally, we show that our method, ALTI (Aggregation of Layer-wise Token-to-token Interactions), provides more faithful explanations and increased robustness than gradient-based methods.

Javier Ferrando, Gerard I. Gállego, Belen Alastruey et al.

In Neural Machine Translation (NMT), each token prediction is conditioned on the source sentence and the target prefix (what has been previously translated at a decoding step). However, previous work on interpretability in NMT has mainly focused solely on source sentence tokens' attributions. Therefore, we lack a full understanding of the influences of every input token (source sentence and target prefix) in the model predictions. In this work, we propose an interpretability method that tracks input tokens' attributions for both contexts. Our method, which can be extended to any encoder-decoder Transformer-based model, allows us to better comprehend the inner workings of current NMT models. We apply the proposed method to both bilingual and multilingual Transformers and present insights into their behaviour.

Elena Voita, Javier Ferrando, Christoforos Nalmpantis

We analyze a family of large language models in such a lightweight manner that can be done on a single GPU. Specifically, we focus on the OPT family of models ranging from 125m to 66b parameters and rely only on whether an FFN neuron is activated or not. First, we find that the early part of the network is sparse and represents many discrete features. Here, many neurons (more than 70% in some layers of the 66b model) are "dead", i.e. they never activate on a large collection of diverse data. At the same time, many of the alive neurons are reserved for discrete features and act as token and n-gram detectors. Interestingly, their corresponding FFN updates not only promote next token candidates as could be expected, but also explicitly focus on removing the information about triggering them tokens, i.e., current input. To the best of our knowledge, this is the first example of mechanisms specialized at removing (rather than adding) information from the residual stream. With scale, models become more sparse in a sense that they have more dead neurons and token detectors. Finally, some neurons are positional: them being activated or not depends largely (or solely) on position and less so (or not at all) on textual data. We find that smaller models have sets of neurons acting as position range indicators while larger models operate in a less explicit manner.

Belen Alastruey, Javier Ferrando, Gerard I. Gállego et al.

Transformers have achieved state-of-the-art results across multiple NLP tasks. However, the self-attention mechanism complexity scales quadratically with the sequence length, creating an obstacle for tasks involving long sequences, like in the speech domain. In this paper, we discuss the usefulness of self-attention for Direct Speech Translation. First, we analyze the layer-wise token contributions in the self-attention of the encoder, unveiling local diagonal patterns. To prove that some attention weights are avoidable, we propose to substitute the standard self-attention with a local efficient one, setting the amount of context used based on the results of the analysis. With this approach, our model matches the baseline performance, and improves the efficiency by skipping the computation of those weights that standard attention discards.

Sergi Masip, Gido M. van de Ven, Javier Ferrando et al.

Catastrophic forgetting in continual learning is often measured at the performance or last-layer representation level, overlooking the underlying mechanisms. We introduce a mechanistic framework that offers a geometric interpretation of catastrophic forgetting as the result of transformations to the encoding of individual features. These transformations can lead to forgetting by reducing the allocated capacity of features (worse representation) and disrupting their readout by downstream computations. Analysis of a tractable model formalizes this view, allowing us to identify best- and worst-case scenarios. Through experiments on this model, we empirically test our formal analysis and highlight the detrimental effect of depth. Finally, we demonstrate how our framework can be used in the analysis of practical models through the use of Crosscoders. We present a case study of a Vision Transformer trained on sequential CIFAR-10. Our work provides a new, feature-centric vocabulary for continual learning.

Arnau Marin-Llobet, Javier Ferrando

We introduce an autonomous multiagent framework for mechanistic interpretability that automates both explaining and finding internal features in large language models. The system runs two coupled loops: (1) explanation refinement, where an agent proposes competing hypotheses and iteratively tests them with targeted prompt controls and a multi-metric evaluation; and (2) feature discovery, where an agent generates prompt sets, constructs a k-nearest-neighbor graph in activation space, and retrieves candidate features using statistical separability and semantic coherence criteria. On Gemma-2 family models and MLP neurons in weight-sparse transformers, our agent improves over one-shot auto-interpretations, discovers language-specific and safety-relevant features, and produces auditable explanation traces, showing that agent-driven empirical loops yield sharper and more falsifiable explanations than one-shot labels.

Javier Ferrando, Matthias Sperber, Hendra Setiawan et al.

Behavioral testing in NLP allows fine-grained evaluation of systems by examining their linguistic capabilities through the analysis of input-output behavior. Unfortunately, existing work on behavioral testing in Machine Translation (MT) is currently restricted to largely handcrafted tests covering a limited range of capabilities and languages. To address this limitation, we propose to use Large Language Models (LLMs) to generate a diverse set of source sentences tailored to test the behavior of MT models in a range of situations. We can then verify whether the MT model exhibits the expected behavior through matching candidate sets that are also generated using LLMs. Our approach aims to make behavioral testing of MT systems practical while requiring only minimal human effort. In our experiments, we apply our proposed evaluation framework to assess multiple available MT systems, revealing that while in general pass-rates follow the trends observable from traditional accuracy-based metrics, our method was able to uncover several important differences and potential bugs that go unnoticed when relying only on accuracy.

Igor Tufanov, Karen Hambardzumyan, Javier Ferrando et al.

We present the LM Transparency Tool (LM-TT), an open-source interactive toolkit for analyzing the internal workings of Transformer-based language models. Differently from previously existing tools that focus on isolated parts of the decision-making process, our framework is designed to make the entire prediction process transparent, and allows tracing back model behavior from the top-layer representation to very fine-grained parts of the model. Specifically, it (1) shows the important part of the whole input-to-output information flow, (2) allows attributing any changes done by a model block to individual attention heads and feed-forward neurons, (3) allows interpreting the functions of those heads or neurons. A crucial part of this pipeline is showing the importance of specific model components at each step. As a result, we are able to look at the roles of model components only in cases where they are important for a prediction. Since knowing which components should be inspected is key for analyzing large models where the number of these components is extremely high, we believe our tool will greatly support the interpretability community both in research settings and in practical applications.

Javier Ferrando, Enrique Lopez-Cuena, Pablo Agustin Martin-Torres et al.

Sparse Autoencoders uncover thousands of features in vision models, yet explaining these features without requiring human intervention remains an open challenge. While previous work has proposed generating correlation-based explanations based on top activating input examples, we present a fundamentally different alternative based on causal interventions. We leverage the structure of Vision-Language Models and steer individual SAE features in the vision encoder after providing an empty image. Then, we prompt the language model to explain what it ``sees'', effectively eliciting the visual concept represented by each feature. Results show that Steering offers an scalable alternative that complements traditional approaches based on input examples, serving as a new axis for automated interpretability in vision models. Moreover, the quality of explanations improves consistently with the scale of the language model, highlighting our method as a promising direction for future research. Finally, we propose Steering-informed Top-k, a hybrid approach that combines the strengths of causal interventions and input-based approaches to achieve state-of-the-art explanation quality without additional computational cost.

Javier Ferrando, Gabriele Sarti, Arianna Bisazza et al.

The rapid progress of research aimed at interpreting the inner workings of advanced language models has highlighted a need for contextualizing the insights gained from years of work in this area. This primer provides a concise technical introduction to the current techniques used to interpret the inner workings of Transformer-based language models, focusing on the generative decoder-only architecture. We conclude by presenting a comprehensive overview of the known internal mechanisms implemented by these models, uncovering connections across popular approaches and active research directions in this area.

Javier Ferrando, Oscar Obeso, Senthooran Rajamanoharan et al.

Hallucinations in large language models are a widespread problem, yet the mechanisms behind whether models will hallucinate are poorly understood, limiting our ability to solve this problem. Using sparse autoencoders as an interpretability tool, we discover that a key part of these mechanisms is entity recognition, where the model detects if an entity is one it can recall facts about. Sparse autoencoders uncover meaningful directions in the representation space, these detect whether the model recognizes an entity, e.g. detecting it doesn't know about an athlete or a movie. This suggests that models can have self-knowledge: internal representations about their own capabilities. These directions are causally relevant: capable of steering the model to refuse to answer questions about known entities, or to hallucinate attributes of unknown entities when it would otherwise refuse. We demonstrate that despite the sparse autoencoders being trained on the base model, these directions have a causal effect on the chat model's refusal behavior, suggesting that chat finetuning has repurposed this existing mechanism. Furthermore, we provide an initial exploration into the mechanistic role of these directions in the model, finding that they disrupt the attention of downstream heads that typically move entity attributes to the final token.

Javier Ferrando, Elena Voita

Information flows by routes inside the network via mechanisms implemented in the model. These routes can be represented as graphs where nodes correspond to token representations and edges to operations inside the network. We automatically build these graphs in a top-down manner, for each prediction leaving only the most important nodes and edges. In contrast to the existing workflows relying on activation patching, we do this through attribution: this allows us to efficiently uncover existing circuits with just a single forward pass. Additionally, the applicability of our method is far beyond patching: we do not need a human to carefully design prediction templates, and we can extract information flow routes for any prediction (not just the ones among the allowed templates). As a result, we can talk about model behavior in general, for specific types of predictions, or different domains. We experiment with Llama 2 and show that the role of some attention heads is overall important, e.g. previous token heads and subword merging heads. Next, we find similarities in Llama 2 behavior when handling tokens of the same part of speech. Finally, we show that some model components can be specialized on domains such as coding or multilingual texts.

Nurkhan Laiyk, Gerard I. Gállego, Javier Ferrando et al.

Function vectors (FVs) are vector representations of tasks extracted from model activations during in-context learning. While prior work has shown that multilingual model representations can be language-agnostic, it remains unclear whether the same holds for function vectors. We study whether FVs exhibit language-agnosticity, using machine translation as a case study. Across three decoder-only multilingual LLMs, we find that translation FVs extracted from a single English$\rightarrow$Target direction transfer to other target languages, consistently improving the rank of correct translation tokens across multiple unseen languages. Ablation results show that removing the FV degrades translation across languages with limited impact on unrelated tasks. We further show that base-model FVs transfer to instruction-tuned variants and partially generalize from word-level to sentence-level translation.

David Puertolas Merenciano, Ekaterina Vasyagina, Raghav Dixit et al.

LoRA adapters let users fine-tune large language models (LLMs) efficiently. However, LoRA adapters are shared through open repositories like Hugging Face Hub \citep{huggingface_hub_docs}, making them vulnerable to backdoor attacks. Current detection methods require running the model with test input data -- making them impractical for screening thousands of adapters where the trigger for backdoor behavior is unknown. We detect poisoned adapters by analyzing their weight matrices directly, without running the model -- making our method data-agnostic. Our method extracts simple statistics -- how concentrated the singular values are, their entropy, and the distribution shape -- and flags adapters that deviate from normal patterns. We evaluate the method on 500 LoRA adapters -- 400 clean, and 100 poisoned for Llama-3.2-3B on instruction and reasoning datasets: Alpaca, Dolly, GSM8K, ARC-Challenge, SQuADv2, NaturalQuestions, HumanEval, and GLUE dataset. We achieve 97\% detection accuracy with less than 2\% false positives.

Oscar Obeso, Andy Arditi, Javier Ferrando et al.

Large language models are now routinely used in high-stakes applications where hallucinations can cause serious harm, such as medical consultations or legal advice. Existing hallucination detection methods, however, are impractical for real-world use, as they are either limited to short factual queries or require costly external verification. We present a cheap, scalable method for real-time identification of hallucinated tokens in long-form generations, and scale it effectively to 70B parameter models. Our approach targets \emph{entity-level hallucinations} -- e.g., fabricated names, dates, citations -- rather than claim-level, thereby naturally mapping to token-level labels and enabling streaming detection. We develop an annotation methodology that leverages web search to annotate model responses with grounded labels indicating which tokens correspond to fabricated entities. This dataset enables us to train effective hallucination classifiers with simple and efficient methods such as linear probes. Evaluating across four model families, our classifiers consistently outperform baselines on long-form responses, including more expensive methods such as semantic entropy (e.g., AUC 0.90 vs 0.71 for Llama-3.3-70B), and are also an improvement in short-form question-answering settings. Moreover, despite being trained only with entity-level labels, our probes effectively detect incorrect answers in mathematical reasoning tasks, indicating generalization beyond entities. While our annotation methodology is expensive, we find that annotated responses from one model can be used to train effective classifiers on other models; accordingly, we publicly release our datasets to facilitate reuse. Overall, our work suggests a promising new approach for scalable, real-world hallucination detection.

Javier Ferrando, Gerard I. Gállego, Ioannis Tsiamas et al.

Language Generation Models produce words based on the previous context. Although existing methods offer input attributions as explanations for a model's prediction, it is still unclear how prior words affect the model's decision throughout the layers. In this work, we leverage recent advances in explainability of the Transformer and present a procedure to analyze models for language generation. Using contrastive examples, we compare the alignment of our explanations with evidence of the linguistic phenomena, and show that our method consistently aligns better than gradient-based and perturbation-based baselines. Then, we investigate the role of MLPs inside the Transformer and show that they learn features that help the model predict words that are grammatically acceptable. Lastly, we apply our method to Neural Machine Translation models, and demonstrate that they generate human-like source-target alignments for building predictions.

Javier Ferrando, Marta R. Costa-jussà

This work proposes an extensive analysis of the Transformer architecture in the Neural Machine Translation (NMT) setting. Focusing on the encoder-decoder attention mechanism, we prove that attention weights systematically make alignment errors by relying mainly on uninformative tokens from the source sequence. However, we observe that NMT models assign attention to these tokens to regulate the contribution in the prediction of the two contexts, the source and the prefix of the target sequence. We provide evidence about the influence of wrong alignments on the model behavior, demonstrating that the encoder-decoder attention mechanism is well suited as an interpretability method for NMT. Finally, based on our analysis, we propose methods that largely reduce the word alignment error rate compared to standard induced alignments from attention weights.

Marta R. Costa-jussà, Carlos Escolano, Christine Basta et al.

Multilingual Neural Machine Translation architectures mainly differ in the amount of sharing modules and parameters among languages. In this paper, and from an algorithmic perspective, we explore if the chosen architecture, when trained with the same data, influences the gender bias accuracy. Experiments in four language pairs show that Language-Specific encoders-decoders exhibit less bias than the Shared encoder-decoder architecture. Further interpretability analysis of source embeddings and the attention shows that, in the Language-Specific case, the embeddings encode more gender information, and its attention is more diverted. Both behaviors help in mitigating gender bias.

Javier Ferrando, Juan Luis Dominguez, Jordi Torres et al.

This paper presents a study showing the benefits of the EfficientNet models compared with heavier Convolutional Neural Networks (CNNs) in the Document Classification task, essential problem in the digitalization process of institutions. We show in the RVL-CDIP dataset that we can improve previous results with a much lighter model and present its transfer learning capabilities on a smaller in-domain dataset such as Tobacco3482. Moreover, we present an ensemble pipeline which is able to boost solely image input by combining image model predictions with the ones generated by BERT model on extracted text by OCR. We also show that the batch size can be effectively increased without hindering its accuracy so that the training process can be sped up by parallelizing throughout multiple GPUs, decreasing the computational time needed. Lastly, we expose the training performance differences between PyTorch and Tensorflow Deep Learning frameworks.