Large language models (LLMs) can solve complex multi-step problems, but little is known about how these computations are implemented internally. Motivated by this, we study how LLMs answer multi-hop queries such as “The spouse of the performer of Imagine is”. These queries require two information extraction steps: a latent one for resolving the first hop (“the performer of Imagine”) into the bridge entity (John Lennon), and another for resolving the second hop (“the spouse of John Lennon”) into the target entity (Yoko Ono). Understanding how the latent step is computed internally is key to understanding the overall computation. By carefully analyzing the internal computations of transformer-based LLMs, we discover that the bridge entity is resolved in the early layers of the model. Then, only after this resolution, the two-hop query is solved in the later layers. Because the second hop commences in later layers, there could be cases where these layers no longer encode the necessary knowledge for correctly predicting the answer. Motivated by this, we propose a novel “back-patching” analysis method whereby a hidden representation from a later layer is patched back to an earlier layer. We find that in up to 66% of previously incorrect cases there exists a back-patch that results in the correct generation of the answer, showing that the later layers indeed sometimes lack the needed functionality. Overall our methods and findings open further opportunities for understanding and improving latent reasoning in transformer-based LLMs.

Modern language models capture a large body of factual knowledge. However, some facts can be incorrectly induced or become obsolete over time, resulting in factually incorrect generations. This has led to the development of various editing methods that allow updating facts encoded by the model. Evaluation of these methods has primarily focused on testing whether an individual fact has been successfully injected, and if similar predictions for other subjects have not changed. Here we argue that such evaluation is limited, since injecting one fact (e.g., “Jack Depp is the son of Johnny Depp”) introduces a “ripple effect” in the form of additional facts that the model needs to update (e.g., “Jack Depp is the sibling of Lily-Rose Depp”). To address this, we propose novel evaluation criteria that consider the implications of an edit on related facts. Using these criteria, we then construct RippleEdits, a diagnostic benchmark of 5K factual edits, capturing various types of ripple effects. We evaluate prominent editing methods on RippleEdits, showing that they fail to introduce consistent changes in the model’s knowledge. In addition, we find that a simple in-context editing baseline obtains the best scores on our benchmark, suggesting a promising research direction for model editing.1

Dual encoders are now the dominant architecture for dense retrieval. Yet, we have little understanding of how they represent text, and why this leads to good performance. In this work, we shed light on this question via distributions over the vocabulary. We propose to interpret the vector representations produced by dual encoders by projecting them into the model’s vocabulary space. We show that the resulting projections contain rich semantic information, and draw connection between them and sparse retrieval. We find that this view can offer an explanation for some of the failure cases of dense retrievers. For example, we observe that the inability of models to handle tail entities is correlated with a tendency of the token distributions to forget some of the tokens of those entities. We leverage this insight and propose a simple way to enrich query and passage representations with lexical information at inference time, and show that this significantly improves performance compared to the original model in zero-shot settings, and specifically on the BEIR benchmark.

Human communication often involves information gaps between the interlocutors. For example, in an educational dialogue a student often provides an answer that is incomplete, and there is a gap between this answer and the perfect one expected by the teacher. Successful dialogue then hinges on the teacher asking about this gap in an effective manner, thus creating a rich and interactive educational experience. We focus on the problem of generating such gap-focused questions (GFQs) automatically. We define the task, highlight key desired aspects of a good GFQ, and propose a model that satisfies these. Finally, we provide an evaluation by human annotators of our generated questions compared against human generated ones, demonstrating competitive performance.

Language models are trained on large volumes of text, and as a result their parameters might contain a significant body of factual knowledge. Any downstream task performed by these models implicitly builds on these facts, and thus it is highly desirable to have means for representing this body of knowledge in an interpretable way. However, there is currently no mechanism for such a representation. Here, we propose to address this goal by extracting a knowledge-graph of facts from a given language model. We describe a procedure for “crawling” the internal knowledge-base of a language model. Specifically, given a seed entity, we expand a knowledge-graph around it. The crawling procedure is decomposed into sub-tasks, realized through specially designed prompts that control for both precision (i.e., that no wrong facts are generated) and recall (i.e., the number of facts generated). We evaluate our approach on graphs crawled starting from dozens of seed entities, and show it yields high precision graphs (82-92%), while emitting a reasonable number of facts per entity.

In-context learning (ICL) in Large Language Models (LLMs) has emerged as a powerful new learning paradigm. However, its underlying mechanism is still not well understood. In particular, it is challenging to map it to the “standard’ machine learning framework, where one uses a training set S to find a best-fitting function f(x) in some hypothesis class. Here we make progress on this problem by showing that the functions learned by ICL often have a very simple structure: they correspond to the transformer LLM whose only inputs are the query x and a single “task vector’ calculated from the training set. Thus, ICL can be seen as compressing S into a single task vector 𝜃(S) and then using this task vector to modulate the transformer to produce the output. We support the above claim via comprehensive experiments across a range of models and tasks.

Transformer-based language models (LMs) are known to capture factual knowledge in their parameters. While previous work looked into where factual associations are stored, only little is known about how they are retrieved internally during inference. We investigate this question through the lens of information flow. Given a subject-relation query, we study how the model aggregates information about the subject and relation to predict the correct attribute. With interventions on attention edges, we first identify two critical points where information propagates to the prediction: one from the relation positions followed by another from the subject positions. Next, by analyzing the information at these points, we unveil a three-step internal mechanism for attribute extraction. First, the representation at the last-subject position goes through an enrichment process, driven by the early MLP sublayers, to encode many subject-related attributes. Second, information from the relation propagates to the prediction. Third, the prediction representation “queries” the enriched subject to extract the attribute. Perhaps surprisingly, this extraction is typically done via attention heads, which often encode subject-attribute mappings in their parameters. Overall, our findings introduce a comprehensive view of how factual associations are stored and extracted internally in LMs, facilitating future research on knowledge localization and editing.

A prominent weakness of modern language models (LMs) is their tendency to generate factually incorrect text, which hinders their usability. A natural question is whether such factual errors can be detected automatically. Inspired by truth-seeking mechanisms in law, we propose a factuality evaluation framework for LMs that is based on cross-examination. Our key idea is that an incorrect claim is likely to result in inconsistency with other claims that the model generates. To discover such inconsistencies, we facilitate a multi-turn interaction between the LM that generated the claim and another LM (acting as an examiner) which introduces questions to discover inconsistencies. We empirically evaluate our method on factual claims made by multiple recent LMs on four benchmarks, finding that it outperforms existing methods and baselines, often by a large gap. Our results demonstrate the potential of using interacting LMs for capturing factual errors.

Vision and language models (VLMs) have demonstrated remarkable zero-shot (ZS) performance in a variety of tasks. However, recent works have shown that even the best VLMs struggle to capture aspects of compositional scene understanding, such as object attributes, relations, and action states. In contrast, obtaining structured annotations, such as scene graphs (SGs), that could improve these models is time-consuming and costly, and thus cannot be used on a large scale. Here we ask whether small SG datasets can provide sufficient information for enhancing structured understanding of pretrained VLMs. We show that it is indeed possible to improve VLMs when learning from SGs by integrating components that incorporate structured information into both visual and textual representations. For the visual side, we incorporate a special “SG Component” in the image transformer trained to predict SG information, while for the textual side, we utilize SGs to generate fine-grained captions that highlight different compositional aspects of the scene. Our method improves the performance of several popular VLMs on multiple VL datasets with only a mild degradation in ZS capabilities.

Dense retrievers for open-domain question answering (ODQA) have been shown to achieve impressive performance by training on large datasets of question-passage pairs. In this work we ask whether this dependence on labeled data can be reduced via unsupervised pretraining that is geared towards ODQA. We show this is in fact possible, via a novel pretraining scheme designed for retrieval. Our “recurring span retrieval” approach uses recurring spans across passages in a document to create pseudo examples for contrastive learning. Our pretraining scheme directly controls for term overlap across pseudo queries and relevant passages, thus allowing to model both lexical and semantic relations between them. The resulting model, named Spider, performs surprisingly well without any labeled training examples on a wide range of ODQA datasets. Specifically, it significantly outperforms all other pretrained baselines in a zero-shot setting, and is competitive with BM25, a strong sparse baseline. Moreover, a hybrid retriever over Spider and BM25 improves over both, and is often competitive with DPR models, which are trained on tens of thousands of examples. Last, notable gains are observed when using Spider as an initialization for supervised training.

We consider the task of image-captioning using only the CLIP model and additional text data at training time and no additional captioned images. Our approach relies on the fact that CLIP is trained to make visual and textual embeddings similar. Therefore, we only need to learn how to translate CLIP textual embeddings back into text, and we can learn how to do this by learning a decoder for the frozen CLIP text encoder using only text. We argue that this intuition is “almost correct” because of a gap between the embedding spaces, and propose to rectify this via noise injection during training. We demonstrate the effectiveness of our approach by showing SOTA zero-shot image captioning across four benchmarks, including style transfer. Code, data, and models are available at https://github.com/DavidHuji/CapDec.

In several question answering benchmarks, pretrained models have reached human parity through fine-tuning on an order of 100,000 annotated questions and answers. We explore the more realistic few-shot setting, where only a few hundred training examples are available, and observe that standard models perform poorly, highlighting the discrepancy between current pretraining objectives and question answering. We propose a new pretraining scheme tailored for question answering: recurring span selection. Given a passage with multiple sets of recurring spans, we mask in each set all recurring spans but one, and ask the model to select the correct span in the passage for each masked span. Masked spans are replaced with a special token, viewed as a question representation, that is later used during fine-tuning to select the answer span. The resulting model obtains surprisingly good results on multiple benchmarks (e.g., 72.7 F1 on SQuAD with only 128 training examples), while maintaining competitive performance in the high-resource setting.

Large pre-trained language models have been shown to encode large amounts of world and commonsense knowledge in their parameters, leading to substantial interest in methods for extracting that knowledge. In past work, knowledge was extracted by taking manually-authored queries and gathering paraphrases for them using a separate pipeline. In this work, we propose a method for automatically rewriting queries into “BERTese”, a paraphrase query that is directly optimized towards better knowledge extraction. To encourage meaningful rewrites, we add auxiliary loss functions that encourage the query to correspond to actual language tokens. We empirically show our approach outperforms competing baselines, obviating the need for complex pipelines. Moreover, BERTese provides some insight into the type of language that helps language models perform knowledge extraction.

Models for reading comprehension (RC) commonly restrict their output space to the set of all single contiguous spans from the input, in order to alleviate the learning problem and avoid the need for a model that generates text explicitly. However, forcing an answer to be a single span can be restrictive, and some recent datasets also include multi-span questions, i.e., questions whose answer is a set of non-contiguous spans in the text. Naturally, models that return single spans cannot answer these questions. In this work, we propose a simple architecture for answering multi-span questions by casting the task as a sequence tagging problem, namely, predicting for each input token whether it should be part of the output or not. Our model substantially improves performance on span extraction questions from DROP and Quoref by 9.9 and 5.5 EM points respectively.

Collecting labeled data for coreference resolution is a challenging task, requiring skilled annotators. It is thus desirable to develop coreference resolution models that can make use of unlabeled data. Here we provide such an approach for the powerful class of neural coreference models. These models rely on representations of mentions, and we show these representations can be learned in a self-supervised manner towards improving resolution accuracy. We propose two self-supervised tasks that are closely related to coreference resolution and thus improve mention representation. Applying this approach to the GAP dataset results in new state of the arts results.

A key challenge in coreference resolution is to capture properties of entity clusters, and use those in the resolution process. Here we provide a simple and effective approach for achieving this, via an “Entity Equalization” mechanism. The Equalization approach represents each mention in a cluster via an approximation of the sum of all mentions in the cluster. We show how this can be done in a fully differentiable end-to-end manner, thus enabling high-order inferences in the resolution process. Our approach, which also employs BERT embeddings, results in new state-of-the-art results on the CoNLL-2012 coreference resolution task, improving average F1 by 3.6%.

We introduce a novel method for multilingual transfer that utilizes deep contextual embeddings, pretrained in an unsupervised fashion. While contextual embeddings have been shown to yield richer representations of meaning compared to their static counterparts, aligning them poses a challenge due to their dynamic nature. To this end, we construct context-independent variants of the original monolingual spaces and utilize their mapping to derive an alignment for the context-dependent spaces. This mapping readily supports processing of a target language, improving transfer by context-aware embeddings. Our experimental results demonstrate the effectiveness of this approach for zero-shot and few-shot learning of dependency parsing. Specifically, our method consistently outperforms the previous state-of-the-art on 6 tested languages, yielding an improvement of 6.8 LAS points on average.

Training semantic parsers from weak supervision (denotations) rather than strong supervision (programs) complicates training in two ways. First, a large search space of potential programs needs to be explored at training time to find a correct program. Second, spurious programs that accidentally lead to a correct denotation add noise to training. In this work we propose that in closed worlds with clear semantic types, one can substantially alleviate these problems by utilizing an abstract representation, where tokens in both the language utterance and program are lifted to an abstract form. We show that these abstractions can be defined with a handful of lexical rules and that they result in sharing between different examples that alleviates the difficulties in training. To test our approach, we develop the first semantic parser for CNLVR, a challenging visual reasoning dataset, where the search space is large and overcoming spuriousness is critical, because denotations are either TRUE or FALSE, and thus random programs are likely to lead to a correct denotation. Our method substantially improves performance, and reaches 82.5% accuracy, a 14.7% absolute accuracy improvement compared to the best reported accuracy so far.

Prepositional phrase (PP) attachment disambiguation is a known challenge in syntactic parsing. The lexical sparsity associated with PP attachments motivates research in word representations that can capture pertinent syntactic and semantic features of the word. One promising solution is to use word vectors induced from large amounts of raw text. However, state-of-the-art systems that employ such representations yield modest gains in PP attachment accuracy. In this paper, we show that word vector representations can yield significant PP attachment performance gains. This is achieved via a non-linear architecture that is discriminatively trained to maximize PP attachment accuracy. The architecture is initialized with word vectors trained from unlabeled data, and relearns those to maximize attachment accuracy. We obtain additional performance gains with alternative representations such as dependency-based word vectors. When tested on both English and Arabic datasets, our method outperforms both a strong SVM classifier and state-of-the-art parsers. For instance, we achieve 82.6% PP attachment accuracy on Arabic, while the Turbo and Charniak self-trained parsers obtain 76.7% and 80.8% respectively.