Yoav Goldberg

Mor Geva, Avi Caciularu, Guy Dar et al. · deepmind

The opaque nature and unexplained behavior of transformer-based language models (LMs) have spurred a wide interest in interpreting their predictions. However, current interpretation methods mostly focus on probing models from outside, executing behavioral tests, and analyzing salience input features, while the internal prediction construction process is largely not understood. In this work, we introduce LM-Debugger, an interactive debugger tool for transformer-based LMs, which provides a fine-grained interpretation of the model's internal prediction process, as well as a powerful framework for intervening in LM behavior. For its backbone, LM-Debugger relies on a recent method that interprets the inner token representations and their updates by the feed-forward layers in the vocabulary space. We demonstrate the utility of LM-Debugger for single-prediction debugging, by inspecting the internal disambiguation process done by GPT2. Moreover, we show how easily LM-Debugger allows to shift model behavior in a direction of the user's choice, by identifying a few vectors in the network and inducing effective interventions to the prediction process. We release LM-Debugger as an open-source tool and a demo over GPT2 models.

Shon Otmazgin, Arie Cattan, Yoav Goldberg · allen-ai

We introduce fastcoref, a python package for fast, accurate, and easy-to-use English coreference resolution. The package is pip-installable, and allows two modes: an accurate mode based on the LingMess architecture, providing state-of-the-art coreference accuracy, and a substantially faster model, F-coref, which is the focus of this work. F-coref allows to process 2.8K OntoNotes documents in 25 seconds on a V100 GPU (compared to 6 minutes for the LingMess model, and to 12 minutes of the popular AllenNLP coreference model) with only a modest drop in accuracy. The fast speed is achieved through a combination of distillation of a compact model from the LingMess model, and an efficient batching implementation using a technique we call leftover batching. Our code is available at https://github.com/shon-otmazgin/fastcoref

Shon Otmazgin, Arie Cattan, Yoav Goldberg · allen-ai

While coreference resolution typically involves various linguistic challenges, recent models are based on a single pairwise scorer for all types of pairs. We present LingMess, a new coreference model that defines different categories of coreference cases and optimize multiple pairwise scorers, where each scorer learns a specific set of linguistic challenges. Our model substantially improves pairwise scores for most categories and outperforms cluster-level performance on Ontonotes and 5 additional datasets. Our model is available in https://github.com/shon-otmazgin/lingmess-coref

Mor Geva, Avi Caciularu, Kevin Ro Wang et al. · deepmind

Transformer-based language models (LMs) are at the core of modern NLP, but their internal prediction construction process is opaque and largely not understood. In this work, we make a substantial step towards unveiling this underlying prediction process, by reverse-engineering the operation of the feed-forward network (FFN) layers, one of the building blocks of transformer models. We view the token representation as a changing distribution over the vocabulary, and the output from each FFN layer as an additive update to that distribution. Then, we analyze the FFN updates in the vocabulary space, showing that each update can be decomposed to sub-updates corresponding to single FFN parameter vectors, each promoting concepts that are often human-interpretable. We then leverage these findings for controlling LM predictions, where we reduce the toxicity of GPT2 by almost 50%, and for improving computation efficiency with a simple early exit rule, saving 20% of computation on average.

Yanai Elazar, Nora Kassner, Shauli Ravfogel et al. · cmu

Large amounts of training data are one of the major reasons for the high performance of state-of-the-art NLP models. But what exactly in the training data causes a model to make a certain prediction? We seek to answer this question by providing a language for describing how training data influences predictions, through a causal framework. Importantly, our framework bypasses the need to retrain expensive models and allows us to estimate causal effects based on observational data alone. Addressing the problem of extracting factual knowledge from pretrained language models (PLMs), we focus on simple data statistics such as co-occurrence counts and show that these statistics do influence the predictions of PLMs, suggesting that such models rely on shallow heuristics. Our causal framework and our results demonstrate the importance of studying datasets and the benefits of causality for understanding NLP models.

Adi Haviv, Ido Cohen, Jacob Gidron et al. · deepmind

To produce accurate predictions, language models (LMs) must balance between generalization and memorization. Yet, little is known about the mechanism by which transformer LMs employ their memorization capacity. When does a model decide to output a memorized phrase, and how is this phrase then retrieved from memory? In this work, we offer the first methodological framework for probing and characterizing recall of memorized sequences in transformer LMs. First, we lay out criteria for detecting model inputs that trigger memory recall, and propose idioms as inputs that typically fulfill these criteria. Next, we construct a dataset of English idioms and use it to compare model behavior on memorized vs. non-memorized inputs. Specifically, we analyze the internal prediction construction process by interpreting the model's hidden representations as a gradual refinement of the output probability distribution. We find that across different model sizes and architectures, memorized predictions are a two-step process: early layers promote the predicted token to the top of the output distribution, and upper layers increase model confidence. This suggests that memorized information is stored and retrieved in the early layers of the network. Last, we demonstrate the utility of our methodology beyond idioms in memorized factual statements. Overall, our work makes a first step towards understanding memory recall, and provides a methodological basis for future studies of transformer memorization.

Aryeh Tiktinsky, Vijay Viswanathan, Danna Niezni et al. · cmu

Combination therapies have become the standard of care for diseases such as cancer, tuberculosis, malaria and HIV. However, the combinatorial set of available multi-drug treatments creates a challenge in identifying effective combination therapies available in a situation. To assist medical professionals in identifying beneficial drug-combinations, we construct an expert-annotated dataset for extracting information about the efficacy of drug combinations from the scientific literature. Beyond its practical utility, the dataset also presents a unique NLP challenge, as the first relation extraction dataset consisting of variable-length relations. Furthermore, the relations in this dataset predominantly require language understanding beyond the sentence level, adding to the challenge of this task. We provide a promising baseline model and identify clear areas for further improvement. We release our dataset, code, and baseline models publicly to encourage the NLP community to participate in this task.

Oscar Sainz, Iker García-Ferrero, Alon Jacovi et al. · ibm-research

The 1st Workshop on Data Contamination (CONDA 2024) focuses on all relevant aspects of data contamination in natural language processing, where data contamination is understood as situations where evaluation data is included in pre-training corpora used to train large scale models, compromising evaluation results. The workshop fostered a shared task to collect evidence on data contamination in current available datasets and models. The goal of the shared task and associated database is to assist the community in understanding the extent of the problem and to assist researchers in avoiding reporting evaluation results on known contaminated resources. The shared task provides a structured, centralized public database for the collection of contamination evidence, open to contributions from the community via GitHub pool requests. This first compilation paper is based on 566 reported entries over 91 contaminated sources from a total of 23 contributors. The details of the individual contamination events are available in the platform. The platform continues to be online, open to contributions from the community.

Royi Rassin, Eran Hirsch, Daniel Glickman et al.

Text-conditioned image generation models often generate incorrect associations between entities and their visual attributes. This reflects an impaired mapping between linguistic binding of entities and modifiers in the prompt and visual binding of the corresponding elements in the generated image. As one notable example, a query like "a pink sunflower and a yellow flamingo" may incorrectly produce an image of a yellow sunflower and a pink flamingo. To remedy this issue, we propose SynGen, an approach which first syntactically analyses the prompt to identify entities and their modifiers, and then uses a novel loss function that encourages the cross-attention maps to agree with the linguistic binding reflected by the syntax. Specifically, we encourage large overlap between attention maps of entities and their modifiers, and small overlap with other entities and modifier words. The loss is optimized during inference, without retraining or fine-tuning the model. Human evaluation on three datasets, including one new and challenging set, demonstrate significant improvements of SynGen compared with current state of the art methods. This work highlights how making use of sentence structure during inference can efficiently and substantially improve the faithfulness of text-to-image generation.

Elron Bandel, Yoav Goldberg, Yanai Elazar · ibm-research

While fine-tuned language models perform well on many tasks, they were also shown to rely on superficial surface features such as lexical overlap. Excessive utilization of such heuristics can lead to failure on challenging inputs. We analyze the use of lexical overlap heuristics in natural language inference, paraphrase detection, and reading comprehension (using a novel contrastive dataset), and find that larger models are much less susceptible to adopting lexical overlap heuristics. We also find that longer training leads models to abandon lexical overlap heuristics. Finally, we provide evidence that the disparity between models size has its source in the pre-trained model

Royi Rassin, Shauli Ravfogel, Yoav Goldberg

We study the way DALLE-2 maps symbols (words) in the prompt to their references (entities or properties of entities in the generated image). We show that in stark contrast to the way human process language, DALLE-2 does not follow the constraint that each word has a single role in the interpretation, and sometimes re-use the same symbol for different purposes. We collect a set of stimuli that reflect the phenomenon: we show that DALLE-2 depicts both senses of nouns with multiple senses at once; and that a given word can modify the properties of two distinct entities in the image, or can be depicted as one object and also modify the properties of another object, creating a semantic leakage of properties between entities. Taken together, our study highlights the differences between DALLE-2 and human language processing and opens an avenue for future study on the inductive biases of text-to-image models.

Hongming Zhang, Yintong Huo, Yanai Elazar et al.

Recently, the community has achieved substantial progress on many commonsense reasoning benchmarks. However, it is still unclear what is learned from the training process: the knowledge, inference capability, or both? We argue that due to the large scale of commonsense knowledge, it is infeasible to annotate a large enough training set for each task to cover all commonsense for learning. Thus we should separate the commonsense knowledge acquisition and inference over commonsense knowledge as two separate tasks. In this work, we focus on investigating models' commonsense inference capabilities from two perspectives: (1) Whether models can know if the knowledge they have is enough to solve the task; (2) Whether models can develop commonsense inference capabilities that generalize across commonsense tasks. We first align commonsense tasks with relevant knowledge from commonsense knowledge bases and ask humans to annotate whether the knowledge is enough or not. Then, we convert different commonsense tasks into a unified question answering format to evaluate models' generalization capabilities. We name the benchmark as Commonsense Inference with Knowledge-in-the-loop Question Answering (CIKQA).

Itay Yair, Hillel Taub-Tabib, Yoav Goldberg · allen-ai

Information extraction systems often produce hundreds to thousands of strings on a specific topic. We present a method that facilitates better consumption of these strings, in an exploratory setting in which a user wants to both get a broad overview of what's available, and a chance to dive deeper on some aspects. The system works by grouping similar items together and arranging the remaining items into a hierarchical navigable DAG structure. We apply the method to medical information extraction.

Aviv Weinstein, Yoav Goldberg · allen-ai

Deverbal nouns are nominal forms of verbs commonly used in written English texts to describe events or actions, as well as their arguments. However, many NLP systems, and in particular pattern-based ones, neglect to handle such nominalized constructions. The solutions that do exist for handling arguments of nominalized constructions are based on semantic annotation and require semantic ontologies, making their applications restricted to a small set of nouns. We propose to adopt instead a more syntactic approach, which maps the arguments of deverbal nouns to the universal-dependency relations of the corresponding verbal construction. We present an unsupervised mechanism -- based on contextualized word representations -- which allows to enrich universal-dependency trees with dependency arcs denoting arguments of deverbal nouns, using the same labels as the corresponding verbal cases. By sharing the same label set as in the verbal case, patterns that were developed for verbs can be applied without modification but with high accuracy also to the nominal constructions.

Hila Gonen, Shauli Ravfogel, Yoav Goldberg

Multilingual language models were shown to allow for nontrivial transfer across scripts and languages. In this work, we study the structure of the internal representations that enable this transfer. We focus on the representation of gender distinctions as a practical case study, and examine the extent to which the gender concept is encoded in shared subspaces across different languages. Our analysis shows that gender representations consist of several prominent components that are shared across languages, alongside language-specific components. The existence of language-independent and language-specific components provides an explanation for an intriguing empirical observation we make: while gender classification transfers well across languages, interventions for gender removal, trained on a single language, do not transfer easily to others.

Mosh Levy, Shauli Ravfogel, Yoav Goldberg

Recent applications of LLMs in Machine Reading Comprehension (MRC) systems have shown impressive results, but the use of shortcuts, mechanisms triggered by features spuriously correlated to the true label, has emerged as a potential threat to their reliability. We analyze the problem from two angles: LLMs as editors, guided to edit text to mislead LLMs; and LLMs as readers, who answer questions based on the edited text. We introduce a framework that guides an editor to add potential shortcuts-triggers to samples. Using GPT4 as the editor, we find it can successfully edit trigger shortcut in samples that fool LLMs. Analysing LLMs as readers, we observe that even capable LLMs can be deceived using shortcut knowledge. Strikingly, we discover that GPT4 can be deceived by its own edits (15% drop in F1). Our findings highlight inherent vulnerabilities of LLMs to shortcut manipulations. We publish ShortcutQA, a curated dataset generated by our framework for future research.

Moran Baruch, Nir Drucker, Gilad Ezov et al.

Training large-scale CNNs that during inference can be run under Homomorphic Encryption (HE) is challenging due to the need to use only polynomial operations. This limits HE-based solutions adoption. We address this challenge and pioneer in providing a novel training method for large polynomial CNNs such as ResNet-152 and ConvNeXt models, and achieve promising accuracy on encrypted samples on large-scale dataset such as ImageNet. Additionally, we provide optimization insights regarding activation functions and skip-connection latency impacts, enhancing HE-based evaluation efficiency. Finally, to demonstrate the robustness of our method, we provide a polynomial adaptation of the CLIP model for secure zero-shot prediction, unlocking unprecedented capabilities at the intersection of HE and transfer learning.

Amit Moryossef, Yanai Elazar, Yoav Goldberg

Piano fingering -- knowing which finger to use to play each note in a musical piece, is a hard and important skill to master when learning to play the piano. While some sheet music is available with expert-annotated fingering information, most pieces lack this information, and people often resort to learning the fingering from demonstrations in online videos. We consider the AI task of automating the extraction of fingering information from videos. This is a non-trivial task as fingers are often occluded by other fingers, and it is often not clear from the video which of the keys were pressed, requiring the synchronization of hand position information and knowledge about the notes that were played. We show how to perform this task with high-accuracy using a combination of deep-learning modules, including a GAN-based approach for fine-tuning on out-of-domain data. We extract the fingering information with an f1 score of 97\%. We run the resulting system on 90 videos, resulting in high-quality piano fingering information of 150K notes, the largest available dataset of piano-fingering to date.

Shauli Ravfogel, Yoav Goldberg, Ryan Cotterell

Methods for erasing human-interpretable concepts from neural representations that assume linearity have been found to be tractable and useful. However, the impact of this removal on the behavior of downstream classifiers trained on the modified representations is not fully understood. In this work, we formally define the notion of log-linear guardedness as the inability of an adversary to predict the concept directly from the representation, and study its implications. We show that, in the binary case, under certain assumptions, a downstream log-linear model cannot recover the erased concept. However, we demonstrate that a multiclass log-linear model \emph{can} be constructed that indirectly recovers the concept in some cases, pointing to the inherent limitations of log-linear guardedness as a downstream bias mitigation technique. These findings shed light on the theoretical limitations of linear erasure methods and highlight the need for further research on the connections between intrinsic and extrinsic bias in neural models.

Asaf Achi Mordechai, Yoav Goldberg, Reut Tsarfaty

Current Text-to-Code models demonstrate impressive capabilities in generating executable code from natural language snippets. However, current studies focus on technical instructions and programmer-oriented language, and it is an open question whether these models can effectively translate natural language descriptions given by non-technical users and express complex goals, to an executable program that contains an intricate flow - composed of API access and control structures as loops, conditions, and sequences. To unlock the challenge of generating a complete program from a plain non-technical description we present NoviCode, a novel NL Programming task, which takes as input an API and a natural language description by a novice non-programmer and provides an executable program as output. To assess the efficacy of models on this task, we provide a novel benchmark accompanied by test suites wherein the generated program code is assessed not according to their form, but according to their functional execution. Our experiments show that, first, NoviCode is indeed a challenging task in the code synthesis domain, and that generating complex code from non-technical instructions goes beyond the current Text-to-Code paradigm. Second, we show that a novel approach wherein we align the NL utterances with the compositional hierarchical structure of the code, greatly enhances the performance of LLMs on this task, compared with the end-to-end Text-to-Code counterparts.

Yoav Goldberg

It is an established assumption that pattern-based models are good at precision, while learning based models are better at recall. But is that really the case? I argue that there are two kinds of recall: d-recall, reflecting diversity, and e-recall, reflecting exhaustiveness. I demonstrate through experiments that while neural methods are indeed significantly better at d-recall, it is sometimes the case that pattern-based methods are still substantially better at e-recall. Ideal methods should aim for both kinds, and this ideal should in turn be reflected in our evaluations.

Amit Moryossef, Zifan Jiang, Mathias Müller et al.

Sign language segmentation is a crucial task in sign language processing systems. It enables downstream tasks such as sign recognition, transcription, and machine translation. In this work, we consider two kinds of segmentation: segmentation into individual signs and segmentation into phrases, larger units comprising several signs. We propose a novel approach to jointly model these two tasks. Our method is motivated by linguistic cues observed in sign language corpora. We replace the predominant IO tagging scheme with BIO tagging to account for continuous signing. Given that prosody plays a significant role in phrase boundaries, we explore the use of optical flow features. We also provide an extensive analysis of hand shapes and 3D hand normalization. We find that introducing BIO tagging is necessary to model sign boundaries. Explicitly encoding prosody by optical flow improves segmentation in shallow models, but its contribution is negligible in deeper models. Careful tuning of the decoding algorithm atop the models further improves the segmentation quality. We demonstrate that our final models generalize to out-of-domain video content in a different signed language, even under a zero-shot setting. We observe that including optical flow and 3D hand normalization enhances the robustness of the model in this context.

Dany Haddad, Dan Bareket, Joseph Chee Chang et al.

AI-powered scientific research tools are rapidly being integrated into research workflows, yet the field lacks a clear lens into how researchers use these systems in real-world settings. We present and analyze the Asta Interaction Dataset, a large-scale resource comprising over 200,000 user queries and interaction logs from two deployed tools (a literature discovery interface and a scientific question-answering interface) within an LLM-powered retrieval-augmented generation platform. Using this dataset, we characterize query patterns, engagement behaviors, and how usage evolves with experience. We find that users submit longer and more complex queries than in traditional search, and treat the system as a collaborative research partner, delegating tasks such as drafting content and identifying research gaps. Users treat generated responses as persistent artifacts, revisiting and navigating among outputs and cited evidence in non-linear ways. With experience, users issue more targeted queries and engage more deeply with supporting citations, although keyword-style queries persist even among experienced users. We release the anonymized dataset and analysis with a new query intent taxonomy to inform future designs of real-world AI research assistants and to support realistic evaluation.

Tal Ishon, Yoav Goldberg, Uri Shaham

Topic modeling seeks to uncover latent semantic structure in text, with LDA providing a foundational probabilistic framework. While recent methods often incorporate external knowledge (e.g., pre-trained embeddings), such reliance limits applicability in emerging or underexplored domains. We introduce \textbf{PRISM}, a corpus-intrinsic method that derives a Dirichlet parameter from word co-occurrence statistics to initialize LDA without altering its generative process. Experiments on text and single cell RNA-seq data show that PRISM improves topic coherence and interpretability, rivaling models that rely on external knowledge. These results underscore the value of corpus-driven initialization for topic modeling in resource-constrained settings. Code is available at: https://github.com/shaham-lab/PRISM.

Uri Katz, Matan Vetzler, Amir DN Cohen et al.

Recognizing entities in texts is a central need in many information-seeking scenarios, and indeed, Named Entity Recognition (NER) is arguably one of the most successful examples of a widely adopted NLP task and corresponding NLP technology. Recent advances in large language models (LLMs) appear to provide effective solutions (also) for NER tasks that were traditionally handled with dedicated models, often matching or surpassing the abilities of the dedicated models. Should NER be considered a solved problem? We argue to the contrary: the capabilities provided by LLMs are not the end of NER research, but rather an exciting beginning. They allow taking NER to the next level, tackling increasingly more useful, and increasingly more challenging, variants. We present three variants of the NER task, together with a dataset to support them. The first is a move towards more fine-grained -- and intersectional -- entity types. The second is a move towards zero-shot recognition and extraction of these fine-grained types based on entity-type labels. The third, and most challenging, is the move from the recognition setup to a novel retrieval setup, where the query is a zero-shot entity type, and the expected result is all the sentences from a large, pre-indexed corpus that contain entities of these types, and their corresponding spans. We show that all of these are far from being solved. We provide a large, silver-annotated corpus of 4 million paragraphs covering 500 entity types, to facilitate research towards all of these three goals.

Zihao Yang, Mosh Levy, Yoav Goldberg et al.

Counterfactual prompting (i.e., perturbing a single factor and measuring output change) is widely used to evaluate things like LLM bias and CoT faithfulness. But in this work we argue that observed effects cannot be attributed to the targeted factor without accounting for baseline ``meaning-preserving'' modifications to text that establish general model sensitivity. This is because every counterfactual edit is a compound treatment that bundles the variable of interest with incidental surface-form variation; this violates treatment variation irrelevance. We observe prediction flip rates on MedQA of 14.9% when we surgically change patient gender. However, this is statistically indistinguishable from the flip rates induced by simply paraphrasing inputs (14.1%). In this case, it would therefore be unwarranted to conclude that the LLM is especially sensitive to patient gender. To account for this and robustly measure the effects of targeted interventions, we propose a framework in which we compare (via statistical testing) differences observed under target interventions to those induced by paraphrasing inputs. We then use this framework to revisit a analysis done on the MedPerturb dataset, which reported evidence of model sensitivity to patient demographics and stylistic cues. We find that these effects largely dissipate when we account for general model sensitivity, with only 5 of 120 tests reaching statistical significance. Applying the same framework to occupational biography classification, we detect clearly significant directional gender bias, showing that the framework identifies real directional effects even when they are small. We evaluate a range of metrics -- aggregate, per-sample distributional, and regression -- and find that per-sample metrics are dramatically more powerful than aggregate metrics and regression powerfully and uniquely characterizes effect direction and magnitude.

Uri Katz, Mosh Levy, Yoav Goldberg

The exponential growth of scientific literature necessitates advanced tools for effective knowledge exploration. We present Knowledge Navigator, a system designed to enhance exploratory search abilities by organizing and structuring the retrieved documents from broad topical queries into a navigable, two-level hierarchy of named and descriptive scientific topics and subtopics. This structured organization provides an overall view of the research themes in a domain, while also enabling iterative search and deeper knowledge discovery within specific subtopics by allowing users to refine their focus and retrieve additional relevant documents. Knowledge Navigator combines LLM capabilities with cluster-based methods to enable an effective browsing method. We demonstrate our approach's effectiveness through automatic and manual evaluations on two novel benchmarks, CLUSTREC-COVID and SCITOC. Our code, prompts, and benchmarks are made publicly available.

Or Shachar, Uri Katz, Yoav Goldberg et al.

We present NER Retriever, a zero-shot retrieval framework for ad-hoc Named Entity Retrieval, a variant of Named Entity Recognition (NER), where the types of interest are not provided in advance, and a user-defined type description is used to retrieve documents mentioning entities of that type. Instead of relying on fixed schemas or fine-tuned models, our method builds on internal representations of large language models (LLMs) to embed both entity mentions and user-provided open-ended type descriptions into a shared semantic space. We show that internal representations, specifically the value vectors from mid-layer transformer blocks, encode fine-grained type information more effectively than commonly used top-layer embeddings. To refine these representations, we train a lightweight contrastive projection network that aligns type-compatible entities while separating unrelated types. The resulting entity embeddings are compact, type-aware, and well-suited for nearest-neighbor search. Evaluated on three benchmarks, NER Retriever significantly outperforms both lexical and dense sentence-level retrieval baselines. Our findings provide empirical support for representation selection within LLMs and demonstrate a practical solution for scalable, schema-free entity retrieval. The NER Retriever Codebase is publicly available at https://github.com/ShacharOr100/ner_retriever

Mosh Levy, Alon Jacoby, Yoav Goldberg

This paper explores the impact of extending input lengths on the capabilities of Large Language Models (LLMs). Despite LLMs advancements in recent times, their performance consistency across different input lengths is not well understood. We investigate this aspect by introducing a novel QA reasoning framework, specifically designed to assess the impact of input length. We isolate the effect of input length using multiple versions of the same sample, each being extended with padding of different lengths, types and locations. Our findings show a notable degradation in LLMs' reasoning performance at much shorter input lengths than their technical maximum. We show that the degradation trend appears in every version of our dataset, although at different intensities. Additionally, our study reveals that the traditional metric of next word prediction correlates negatively with performance of LLMs' on our reasoning dataset. We analyse our results and identify failure modes that can serve as useful guides for future research, potentially informing strategies to address the limitations observed in LLMs.

Amit Moryossef, Mathias Müller, Anne Göhring et al.

Sign language translation systems are complex and require many components. As a result, it is very hard to compare methods across publications. We present an open-source implementation of a text-to-gloss-to-pose-to-video pipeline approach, demonstrating conversion from German to Swiss German Sign Language, French to French Sign Language of Switzerland, and Italian to Italian Sign Language of Switzerland. We propose three different components for the text-to-gloss translation: a lemmatizer, a rule-based word reordering and dropping component, and a neural machine translation system. Gloss-to-pose conversion occurs using data from a lexicon for three different signed languages, with skeletal poses extracted from videos. To generate a sentence, the text-to-gloss system is first run, and the pose representations of the resulting signs are stitched together.

Eric Lehman, Sarthak Jain, Karl Pichotta et al.

Large Transformers pretrained over clinical notes from Electronic Health Records (EHR) have afforded substantial gains in performance on predictive clinical tasks. The cost of training such models (and the necessity of data access to do so) coupled with their utility motivates parameter sharing, i.e., the release of pretrained models such as ClinicalBERT. While most efforts have used deidentified EHR, many researchers have access to large sets of sensitive, non-deidentified EHR with which they might train a BERT model (or similar). Would it be safe to release the weights of such a model if they did? In this work, we design a battery of approaches intended to recover Personal Health Information (PHI) from a trained BERT. Specifically, we attempt to recover patient names and conditions with which they are associated. We find that simple probing methods are not able to meaningfully extract sensitive information from BERT trained over the MIMIC-III corpus of EHR. However, more sophisticated "attacks" may succeed in doing so: To facilitate such research, we make our experimental setup and baseline probing models available at https://github.com/elehman16/exposing_patient_data_release

Aryeh Tiktinsky, Yoav Goldberg, Reut Tsarfaty

Syntactic dependencies can be predicted with high accuracy, and are useful for both machine-learned and pattern-based information extraction tasks. However, their utility can be improved. These syntactic dependencies are designed to accurately reflect syntactic relations, and they do not make semantic relations explicit. Therefore, these representations lack many explicit connections between content words, that would be useful for downstream applications. Proposals like English Enhanced UD improve the situation by extending universal dependency trees with additional explicit arcs. However, they are not available to Python users, and are also limited in coverage. We introduce a broad-coverage, data-driven and linguistically sound set of transformations, that makes event-structure and many lexical relations explicit. We present pyBART, an easy-to-use open-source Python library for converting English UD trees either to Enhanced UD graphs or to our representation. The library can work as a standalone package or be integrated within a spaCy NLP pipeline. When evaluated in a pattern-based relation extraction scenario, our representation results in higher extraction scores than Enhanced UD, while requiring fewer patterns.

Guy Kushilevitz, Shaul Markovitch, Yoav Goldberg

We tackle the task of Term Set Expansion (TSE): given a small seed set of example terms from a semantic class, finding more members of that class. The task is of great practical utility, and also of theoretical utility as it requires generalization from few examples. Previous approaches to the TSE task can be characterized as either distributional or pattern-based. We harness the power of neural masked language models (MLM) and propose a novel TSE algorithm, which combines the pattern-based and distributional approaches. Due to the small size of the seed set, fine-tuning methods are not effective, calling for more creative use of the MLM. The gist of the idea is to use the MLM to first mine for informative patterns with respect to the seed set, and then to obtain more members of the seed class by generalizing these patterns. Our method outperforms state-of-the-art TSE algorithms. Implementation is available at: https://github.com/ guykush/TermSetExpansion-MPB/

Asaf Amrami, Yoav Goldberg

Word sense induction (WSI) is the task of unsupervised clustering of word usages within a sentence to distinguish senses. Recent work obtain strong results by clustering lexical substitutes derived from pre-trained RNN language models (ELMo). Adapting the method to BERT improves the scores even further. We extend the previous method to support a dynamic rather than a fixed number of clusters as supported by other prominent methods, and propose a method for interpreting the resulting clusters by associating them with their most informative substitutes. We then perform extensive error analysis revealing the remaining sources of errors in the WSI task. Our code is available at https://github.com/asafamr/bertwsi.

Alon Jacovi, Oren Sar Shalom, Yoav Goldberg

We present an analysis into the inner workings of Convolutional Neural Networks (CNNs) for processing text. CNNs used for computer vision can be interpreted by projecting filters into image space, but for discrete sequence inputs CNNs remain a mystery. We aim to understand the method by which the networks process and classify text. We examine common hypotheses to this problem: that filters, accompanied by global max-pooling, serve as ngram detectors. We show that filters may capture several different semantic classes of ngrams by using different activation patterns, and that global max-pooling induces behavior which separates important ngrams from the rest. Finally, we show practical use cases derived from our findings in the form of model interpretability (explaining a trained model by deriving a concrete identity for each filter, bridging the gap between visualization tools in vision tasks and NLP) and prediction interpretability (explaining predictions). Code implementation is available online at github.com/sayaendo/interpreting-cnn-for-text.

Graham Neubig, Chris Dyer, Yoav Goldberg et al.

We describe DyNet, a toolkit for implementing neural network models based on dynamic declaration of network structure. In the static declaration strategy that is used in toolkits like Theano, CNTK, and TensorFlow, the user first defines a computation graph (a symbolic representation of the computation), and then examples are fed into an engine that executes this computation and computes its derivatives. In DyNet's dynamic declaration strategy, computation graph construction is mostly transparent, being implicitly constructed by executing procedural code that computes the network outputs, and the user is free to use different network structures for each input. Dynamic declaration thus facilitates the implementation of more complicated network architectures, and DyNet is specifically designed to allow users to implement their models in a way that is idiomatic in their preferred programming language (C++ or Python). One challenge with dynamic declaration is that because the symbolic computation graph is defined anew for every training example, its construction must have low overhead. To achieve this, DyNet has an optimized C++ backend and lightweight graph representation. Experiments show that DyNet's speeds are faster than or comparable with static declaration toolkits, and significantly faster than Chainer, another dynamic declaration toolkit. DyNet is released open-source under the Apache 2.0 license and available at http://github.com/clab/dynet.

Eliyahu Kiperwasser, Yoav Goldberg

We suggest a compositional vector representation of parse trees that relies on a recursive combination of recurrent-neural network encoders. To demonstrate its effectiveness, we use the representation as the backbone of a greedy, bottom-up dependency parser, achieving state-of-the-art accuracies for English and Chinese, without relying on external word embeddings. The parser's implementation is available for download at the first author's webpage.

Noga Peleg Pelc, Gal A. Kaminka, Yoav Goldberg

Large language models are increasingly used within larger systems ("LLM agents"). These make a sequence of LLM calls, each call providing the LLM with a combination of instructions, observations, and interaction history. The design of the encoded information and its structure play a central role in the quality of the resulting system, leading to efforts spent on context engineering. It is therefore critical to communicate the composition of the LLM context in a system, and how it evolves over time. Yet, no standard exists for doing so: context construction is typically conveyed through informal prose, ad hoc diagrams, or direct inspection of code, none of which precisely capture how a prompt evolves across interaction steps or how two context representation strategies differ. To remedy this, we introduce the Agentic Context Description Language (ACDL), a language for specifying the structure and dynamics of LLM input contexts in a precise, readable, and standard manner, along with visualizations. ACDL provides constructs for specifying context aspects such as role message sequences, dynamic content, time-indexed references, and conditional or iterative structure, capturing the full architecture of a prompt independently of any particular implementation. ACDL diagrams can be hand drawn on a whiteboard, or written in formal language which can then be rendered. We describe the language, demonstrate it by documenting several existing systems and their variants, and encourage the community to adopt it for describing LLM systems context, both in day-to-day communication and in papers. Tooling, examples and documentation are available at www.acdlang.org.

Victoria Basmov, Yoav Goldberg, Reut Tsarfaty

The task of reading comprehension (RC), often implemented as context-based question answering (QA), provides a primary means to assess language models' natural language understanding (NLU) capabilities. Yet, when applied to large language models (LLMs) with extensive built-in world knowledge, this method can be deceptive. If the context aligns with the LLMs' internal knowledge, it is hard to discern whether the models' answers stem from context comprehension or from LLMs' internal information. Conversely, using data that conflicts with the models' knowledge creates erroneous trends which distort the results. To address this issue, we suggest to use RC on imaginary data, based on fictitious facts and entities. This task is entirely independent of the models' world knowledge, enabling us to evaluate LLMs' linguistic abilities without the interference of parametric knowledge. Testing ChatGPT, GPT-4, LLaMA 2 and Mixtral on such imaginary data, we uncover a class of linguistic phenomena posing a challenge to current LLMs, involving thinking in terms of alternative, hypothetical scenarios. While all the models handle simple affirmative and negative contexts with high accuracy, they are much more prone to error when dealing with modal and conditional contexts. Crucially, these phenomena also trigger the LLMs' vulnerability to knowledge-conflicts again. In particular, while some models prove virtually unaffected by knowledge conflicts in affirmative and negative contexts, when faced with more semantically involved modal and conditional environments, they often fail to separate the text from their internal knowledge.

Amir DN Cohen, Hilla Merhav, Yoav Goldberg et al.

Current benchmarks for Hebrew Natural Language Processing (NLP) focus mainly on morpho-syntactic tasks, neglecting the semantic dimension of language understanding. To bridge this gap, we set out to deliver a Hebrew Machine Reading Comprehension (MRC) dataset, where MRC is to be realized as extractive Question Answering. The morphologically rich nature of Hebrew poses a challenge to this endeavor: the indeterminacy and non-transparency of span boundaries in morphologically complex forms lead to annotation inconsistencies, disagreements, and flaws in standard evaluation metrics. To remedy this, we devise a novel set of guidelines, a controlled crowdsourcing protocol, and revised evaluation metrics that are suitable for the morphologically rich nature of the language. Our resulting benchmark, HeQ (Hebrew QA), features 30,147 diverse question-answer pairs derived from both Hebrew Wikipedia articles and Israeli tech news. Our empirical investigation reveals that standard evaluation metrics such as F1 scores and Exact Match (EM) are not appropriate for Hebrew (and other MRLs), and we propose a relevant enhancement. In addition, our experiments show low correlation between models' performance on morpho-syntactic tasks and on MRC, which suggests that models designed for the former might underperform on semantics-heavy tasks. The development and exploration of HeQ illustrate some of the challenges MRLs pose in natural language understanding (NLU), fostering progression towards more and better NLU models for Hebrew and other MRLs.

Tomer Wolfson, Harsh Trivedi, Mor Geva et al. · deepmind

Automated agents, powered by Large language models (LLMs), are emerging as the go-to tool for querying information. However, evaluation benchmarks for LLM agents rarely feature natural questions that are both information-seeking and genuinely time-consuming for humans. To address this gap we introduce MoNaCo, a benchmark of 1,315 natural and time-consuming questions that require dozens, and at times hundreds, of intermediate steps to solve -- far more than any existing QA benchmark. To build MoNaCo, we developed a decomposed annotation pipeline to elicit and manually answer real-world time-consuming questions at scale. Frontier LLMs evaluated on MoNaCo achieve at most 61.2% F1, hampered by low recall and hallucinations. Our results underscore the limitations of LLM-powered agents in handling the complexity and sheer breadth of real-world information-seeking tasks -- with MoNaCo providing an effective resource for tracking such progress. The MoNaCo benchmark, codebase, prompts and models predictions are all publicly available at: https://tomerwolgithub.github.io/monaco

Royi Rassin, Aviv Slobodkin, Shauli Ravfogel et al.

We introduce GRADE, an automatic method for quantifying sample diversity in text-to-image models. Our method leverages the world knowledge embedded in large language models and visual question-answering systems to identify relevant concept-specific axes of diversity (e.g., ``shape'' for the concept ``cookie''). It then estimates frequency distributions of concepts and their attributes and quantifies diversity using entropy. We use GRADE to measure the diversity of 12 models over a total of 720K images, revealing that all models display limited variation, with clear deterioration in stronger models. Further, we find that models often exhibit default behaviors, a phenomenon where a model consistently generates concepts with the same attributes (e.g., 98% of the cookies are round). Lastly, we show that a key reason for low diversity is underspecified captions in training data. Our work proposes an automatic, semantically-driven approach to measure sample diversity and highlights the stunning homogeneity in text-to-image models.

Gili Lior, Yoav Goldberg, Gabriel Stanovsky

Document collections of various domains, e.g., legal, medical, or financial, often share some underlying collection-wide structure, which captures information that can aid both human users and structure-aware models. We propose to identify the typical structure of document within a collection, which requires to capture recurring topics across the collection, while abstracting over arbitrary header paraphrases, and ground each topic to respective document locations. These requirements pose several challenges: headers that mark recurring topics frequently differ in phrasing, certain section headers are unique to individual documents and do not reflect the typical structure, and the order of topics can vary between documents. Subsequently, we develop an unsupervised graph-based method which leverages both inter- and intra-document similarities, to extract the underlying collection-wide structure. Our evaluations on three diverse domains in both English and Hebrew indicate that our method extracts meaningful collection-wide structure, and we hope that future work will leverage our method for multi-document applications and structure-aware models.

Matan Avitan, Moran Baruch, Nir Drucker et al.

Large language models (LLMs) power modern AI applications, but processing sensitive data on untrusted servers raises privacy concerns. Homomorphic encryption (HE) enables computation on encrypted data for secure inference. However, neural text generation requires decoding methods like argmax and sampling, which are non-polynomial and thus computationally expensive under encryption, creating a significant performance bottleneck. We introduce cutmax, an HE-friendly argmax algorithm that reduces ciphertext operations compared to prior methods, enabling practical greedy decoding under encryption. We also propose the first HE-compatible nucleus (top-p) sampling method, leveraging cutmax for efficient stochastic decoding with provable privacy guarantees. Both techniques are polynomial, supporting efficient inference in privacy-preserving settings. Moreover, their differentiability facilitates gradient-based sequence-level optimization as a polynomial alternative to straight-through estimators. We further provide strong theoretical guarantees for cutmax, proving its convergence via exponential amplification of the gap ratio between the maximum and runner-up elements. Evaluations on realistic LLM outputs show latency reductions of 24x-35x over baselines, advancing secure text generation.

Mosh Levy, Zohar Elyoseph, Yoav Goldberg

A new generation of AI models generates step-by-step reasoning text before producing an answer. This text appears to offer a human-readable window into their computation process, and is increasingly relied upon for transparency and interpretability. However, it is unclear whether human understanding of this text matches the model's actual computational process. In this paper, we investigate a necessary condition for correspondence: the ability of humans to identify which steps in a reasoning text causally influence later steps. We evaluated humans on this ability by composing questions based on counterfactual measurements and found a significant discrepancy: participant accuracy was only 29%, barely above chance (25%), and remained low (42%) even when evaluating the majority vote on questions with high agreement. Our results reveal a fundamental gap between how humans interpret reasoning texts and how models use it, challenging its utility as a simple interpretability tool. We argue that reasoning texts should be treated as an artifact to be investigated, not taken at face value, and that understanding the non-human ways these models use language is a critical research direction.

Shir Ashury-Tahan, Amir David Nissan Cohen, Nadav Cohen et al.

While coreference resolution is traditionally used as a component in individual document understanding, in this work we take a more global view and explore what can we learn about a domain from the set of all document-level coreference relations that are present in a large corpus. We derive coreference chains from a corpus of 30 million biomedical abstracts and construct a graph based on the string phrases within these chains, establishing connections between phrases if they co-occur within the same coreference chain. We then use the graph structure and the betweeness centrality measure to distinguish between edges denoting hierarchy, identity and noise, assign directionality to edges denoting hierarchy, and split nodes (strings) that correspond to multiple distinct concepts. The result is a rich, data-driven ontology over concepts in the biomedical domain, parts of which overlaps significantly with human-authored ontologies. We release the coreference chains and resulting ontology under a creative-commons license, along with the code.

Matan Avitan, Ryan Cotterell, Yoav Goldberg et al.

Interventions targeting the representation space of language models (LMs) have emerged as an effective means to influence model behavior. Such methods are employed, for example, to eliminate or alter the encoding of demographic information such as gender within the model's representations and, in so doing, create a counterfactual representation. However, because the intervention operates within the representation space, understanding precisely what aspects of the text it modifies poses a challenge. In this paper, we give a method to convert representation counterfactuals into string counterfactuals. We demonstrate that this approach enables us to analyze the linguistic alterations corresponding to a given representation space intervention and to interpret the features utilized to encode a specific concept. Moreover, the resulting counterfactuals can be used to mitigate bias in classification through data augmentation.

Mosh Levy, Zohar Elyoseph, Shauli Ravfogel et al.

Large Language Models (LLMs) can generate reasoning tokens before their final answer to boost performance on complex tasks. While these sequences seem like human thought processes, empirical evidence reveals that they are not a faithful explanation of the model's actual reasoning process. To address this gap between appearance and function, we introduce the State over Tokens (SoT) conceptual framework. SoT reframes reasoning tokens not as a linguistic narrative, but as an externalized computational state -- the sole persistent information carrier across the model's stateless generation cycles. This explains how the tokens can drive correct reasoning without being a faithful explanation when read as text and surfaces previously overlooked research questions on these tokens. We argue that to truly understand the process that LLMs do, research must move beyond reading the reasoning tokens as text and focus on decoding them as state.

Jonathan Bragg, Mike D'Arcy, Nishant Balepur et al. · allen-ai

AI agents hold the potential to revolutionize scientific productivity by automating literature reviews, replicating experiments, analyzing data, and even proposing new directions of inquiry; indeed, there are now many such agents, ranging from general-purpose "deep research" systems to specialized science-specific agents, such as AI Scientist and AIGS. Rigorous evaluation of these agents is critical for progress. Yet existing benchmarks fall short on several fronts: they (1) fail to provide holistic, product-informed measures of real-world use cases such as science research; (2) lack reproducible agent tools necessary for a controlled comparison of core agentic capabilities; (3) do not account for confounding variables such as model cost and tool access; (4) do not provide standardized interfaces for quick agent prototyping and evaluation; and (5) lack comprehensive baseline agents necessary to identify true advances. In response, we define principles and tooling for more rigorously benchmarking agents. Using these, we present AstaBench, a suite that provides the first holistic measure of agentic ability to perform scientific research, comprising 2400+ problems spanning the entire scientific discovery process and multiple scientific domains, and including many problems inspired by actual user requests to deployed Asta agents. Our suite comes with the first scientific research environment with production-grade search tools that enable controlled, reproducible evaluation, better accounting for confounders. Alongside, we provide a comprehensive suite of nine science-optimized classes of Asta agents and numerous baselines. Our extensive evaluation of 57 agents across 22 agent classes reveals several interesting findings, most importantly that despite meaningful progress on certain individual aspects, AI remains far from solving the challenge of science research assistance.

Amir DN Cohen, Shauli Ravfogel, Shaltiel Shmidman et al.

In few-shot relation classification (FSRC), models must generalize to novel relations with only a few labeled examples. While much of the recent progress in NLP has focused on scaling data size, we argue that diversity in relation types is more crucial for FSRC performance. In this work, we demonstrate that training on a diverse set of relations significantly enhances a model's ability to generalize to unseen relations, even when the overall dataset size remains fixed. We introduce REBEL-FS, a new FSRC benchmark that incorporates an order of magnitude more relation types than existing datasets. Through systematic experiments, we show that increasing the diversity of relation types in the training data leads to consistent gains in performance across various few-shot learning scenarios, including high-negative settings. Our findings challenge the common assumption that more data alone leads to better performance and suggest that targeted data curation focused on diversity can substantially reduce the need for large-scale datasets in FSRC.