Mandar Joshi

Kenton Lee, Mandar Joshi, Iulia Turc et al. · deepmind

Visually-situated language is ubiquitous -- sources range from textbooks with diagrams to web pages with images and tables, to mobile apps with buttons and forms. Perhaps due to this diversity, previous work has typically relied on domain-specific recipes with limited sharing of the underlying data, model architectures, and objectives. We present Pix2Struct, a pretrained image-to-text model for purely visual language understanding, which can be finetuned on tasks containing visually-situated language. Pix2Struct is pretrained by learning to parse masked screenshots of web pages into simplified HTML. The web, with its richness of visual elements cleanly reflected in the HTML structure, provides a large source of pretraining data well suited to the diversity of downstream tasks. Intuitively, this objective subsumes common pretraining signals such as OCR, language modeling, image captioning. In addition to the novel pretraining strategy, we introduce a variable-resolution input representation and a more flexible integration of language and vision inputs, where language prompts such as questions are rendered directly on top of the input image. For the first time, we show that a single pretrained model can achieve state-of-the-art results in six out of nine tasks across four domains: documents, illustrations, user interfaces, and natural images.

Devendra Singh Sachan, Mike Lewis, Mandar Joshi et al. · mila, uw

We propose a simple and effective re-ranking method for improving passage retrieval in open question answering. The re-ranker re-scores retrieved passages with a zero-shot question generation model, which uses a pre-trained language model to compute the probability of the input question conditioned on a retrieved passage. This approach can be applied on top of any retrieval method (e.g. neural or keyword-based), does not require any domain- or task-specific training (and therefore is expected to generalize better to data distribution shifts), and provides rich cross-attention between query and passage (i.e. it must explain every token in the question). When evaluated on a number of open-domain retrieval datasets, our re-ranker improves strong unsupervised retrieval models by 6%-18% absolute and strong supervised models by up to 12% in terms of top-20 passage retrieval accuracy. We also obtain new state-of-the-art results on full open-domain question answering by simply adding the new re-ranker to existing models with no further changes.

Hexiang Hu, Yi Luan, Yang Chen et al. · deepmind, gatech

Large-scale multi-modal pre-training models such as CLIP and PaLI exhibit strong generalization on various visual domains and tasks. However, existing image classification benchmarks often evaluate recognition on a specific domain (e.g., outdoor images) or a specific task (e.g., classifying plant species), which falls short of evaluating whether pre-trained foundational models are universal visual recognizers. To address this, we formally present the task of Open-domain Visual Entity recognitioN (OVEN), where a model need to link an image onto a Wikipedia entity with respect to a text query. We construct OVEN-Wiki by re-purposing 14 existing datasets with all labels grounded onto one single label space: Wikipedia entities. OVEN challenges models to select among six million possible Wikipedia entities, making it a general visual recognition benchmark with the largest number of labels. Our study on state-of-the-art pre-trained models reveals large headroom in generalizing to the massive-scale label space. We show that a PaLI-based auto-regressive visual recognition model performs surprisingly well, even on Wikipedia entities that have never been seen during fine-tuning. We also find existing pretrained models yield different strengths: while PaLI-based models obtain higher overall performance, CLIP-based models are better at recognizing tail entities.

Fangyu Liu, Julian Martin Eisenschlos, Francesco Piccinno et al. · deepmind

Visual language such as charts and plots is ubiquitous in the human world. Comprehending plots and charts requires strong reasoning skills. Prior state-of-the-art (SOTA) models require at least tens of thousands of training examples and their reasoning capabilities are still much limited, especially on complex human-written queries. This paper presents the first one-shot solution to visual language reasoning. We decompose the challenge of visual language reasoning into two steps: (1) plot-to-text translation, and (2) reasoning over the translated text. The key in this method is a modality conversion module, named as DePlot, which translates the image of a plot or chart to a linearized table. The output of DePlot can then be directly used to prompt a pretrained large language model (LLM), exploiting the few-shot reasoning capabilities of LLMs. To obtain DePlot, we standardize the plot-to-table task by establishing unified task formats and metrics, and train DePlot end-to-end on this task. DePlot can then be used off-the-shelf together with LLMs in a plug-and-play fashion. Compared with a SOTA model finetuned on more than >28k data points, DePlot+LLM with just one-shot prompting achieves a 24.0% improvement over finetuned SOTA on human-written queries from the task of chart QA.

Fangyu Liu, Francesco Piccinno, Syrine Krichene et al. · deepmind

Visual language data such as plots, charts, and infographics are ubiquitous in the human world. However, state-of-the-art vision-language models do not perform well on these data. We propose MatCha (Math reasoning and Chart derendering pretraining) to enhance visual language models' capabilities in jointly modeling charts/plots and language data. Specifically, we propose several pretraining tasks that cover plot deconstruction and numerical reasoning which are the key capabilities in visual language modeling. We perform the MatCha pretraining starting from Pix2Struct, a recently proposed image-to-text visual language model. On standard benchmarks such as PlotQA and ChartQA, the MatCha model outperforms state-of-the-art methods by as much as nearly 20%. We also examine how well MatCha pretraining transfers to domains such as screenshots, textbook diagrams, and document figures and observe overall improvement, verifying the usefulness of MatCha pretraining on broader visual language tasks.

Mandar Joshi, Terra Blevins, Mike Lewis et al. · uw

Creating labeled natural language training data is expensive and requires significant human effort. We mine input output examples from large corpora using a supervised mining function trained using a small seed set of only 100 examples. The mining consists of two stages -- (1) a biencoder-based recall-oriented dense search which pairs inputs with potential outputs, and (2) a crossencoder-based filter which re-ranks the output of the biencoder stage for better precision. Unlike model-generated data augmentation, our method mines naturally occurring high-quality input output pairs to mimic the style of the seed set for multiple tasks. On SQuAD-style reading comprehension, augmenting the seed set with the mined data results in an improvement of 13 F1 over a BART-large baseline fine-tuned only on the seed set. Likewise, we see improvements of 1.46 ROUGE-L on Xsum abstractive summarization.

Wang Zhu, Alekh Agarwal, Mandar Joshi et al. · uw

Understanding visually situated language requires interpreting complex layouts of textual and visual elements. Pre-processing tools, such as optical character recognition (OCR), can map document image inputs to textual tokens, then large language models (LLMs) can reason over text. However, such methods have high computational and engineering complexity. Can small pretrained image-to-text models accurately understand visual documents through similar recognition and reasoning steps instead? We propose Rationale Distillation (RD), which incorporates the outputs of OCR tools, LLMs, and larger multimodal models as intermediate "rationales", and trains a small student model to predict both rationales and answers. On three visual document understanding benchmarks representing infographics, scanned documents, and figures, our Pix2Struct (282M parameters) student model finetuned with RD outperforms the base model by 4-5% absolute accuracy with only 1% higher computational cost.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Mandar Joshi, Farzana Zahid, Judy Bowen et al.

Industrial Water Treatment Systems (IWTS) are safety critical cyber-physical infrastructures and due to increased connectivity, these systems are exposed to cyber threats that can manipulate process behaviour without creating obvious devices outliers. In particular, logic-layer deception anomalies can preserve numerically plausible measurements while breaking expected cause-and-effect relationships in the control process. These attacks are difficult to detect using threshold-based monitoring or require heavy server-oriented anomaly detection models. This paper explores the potential of Tiny Deep Learning (TinyDL) to provide lightweight on-device logic-level anomaly detection for resource constrained Programmable Logic Controllers (PLCs). We propose a novel framework, TinyDL-based incremental LSTM (Ti-iLSTM) which optimises the memory and space foot print of Long Short-Term Memory (LSTM), to detect logic-layer inconsistencies in Programmable Logic Controller (PLC) based Industrial Water Treatment Systems (IWTS). Experiments on the publicly available SWaT dataset show that the optimised model achieves high detection performance (F1-score=0.983 and ROC-AUC=0.998). A deployment-style validation on the WADI dataset confirms that the proposed light-weight framework remains applicable beyond a single dataset. The research demonstrates that combining logic-aware supervision with Tiny Deep Learning (TinyDL) sequence learning creates an efficient and accurate anomaly detection suitable for resource constrained Programmable Logic Controllers (PLCs) in industrial environments.

Mandar Joshi, Eunsol Choi, Daniel S. Weld et al.

We present TriviaQA, a challenging reading comprehension dataset containing over 650K question-answer-evidence triples. TriviaQA includes 95K question-answer pairs authored by trivia enthusiasts and independently gathered evidence documents, six per question on average, that provide high quality distant supervision for answering the questions. We show that, in comparison to other recently introduced large-scale datasets, TriviaQA (1) has relatively complex, compositional questions, (2) has considerable syntactic and lexical variability between questions and corresponding answer-evidence sentences, and (3) requires more cross sentence reasoning to find answers. We also present two baseline algorithms: a feature-based classifier and a state-of-the-art neural network, that performs well on SQuAD reading comprehension. Neither approach comes close to human performance (23% and 40% vs. 80%), suggesting that TriviaQA is a challenging testbed that is worth significant future study. Data and code available at -- http://nlp.cs.washington.edu/triviaqa/

Archiki Prasad, Mandar Joshi, Kenton Lee et al.

Chain-of-thought (CoT) reasoning and its variants have substantially improved the performance of language models on complex reasoning tasks, yet the precise mechanisms by which different strategies facilitate generalization remain poorly understood. While current explanations often point to increased test-time computation or structural guidance, establishing a consistent, quantifiable link between these factors and generalization remains challenging. In this work, we identify intrinsic dimensionality as a quantitative measure for characterizing the effectiveness of reasoning chains. Intrinsic dimensionality quantifies the minimum number of model dimensions needed to reach a given accuracy threshold on a given task. By keeping the model architecture fixed and varying the task formulation through different reasoning strategies, we demonstrate that effective reasoning strategies consistently reduce the intrinsic dimensionality of the task. Validating this on GSM8K with Gemma-3 1B and 4B, we observe a strong inverse correlation between the intrinsic dimensionality of a reasoning strategy and its generalization performance on both in-distribution and out-of-distribution data. Our findings suggest that effective reasoning chains facilitate learning by better compressing the task using fewer parameters, offering a new quantitative metric for analyzing reasoning processes.

Peter Shaw, Mandar Joshi, James Cohan et al.

Much of the previous work towards digital agents for graphical user interfaces (GUIs) has relied on text-based representations (derived from HTML or other structured data sources), which are not always readily available. These input representations have been often coupled with custom, task-specific action spaces. This paper focuses on creating agents that interact with the digital world using the same conceptual interface that humans commonly use -- via pixel-based screenshots and a generic action space corresponding to keyboard and mouse actions. Building upon recent progress in pixel-based pretraining, we show, for the first time, that it is possible for such agents to outperform human crowdworkers on the MiniWob++ benchmark of GUI-based instruction following tasks.

Xi Chen, Josip Djolonga, Piotr Padlewski et al.

We present the training recipe and results of scaling up PaLI-X, a multilingual vision and language model, both in terms of size of the components and the breadth of its training task mixture. Our model achieves new levels of performance on a wide-range of varied and complex tasks, including multiple image-based captioning and question-answering tasks, image-based document understanding and few-shot (in-context) learning, as well as object detection, video question answering, and video captioning. PaLI-X advances the state-of-the-art on most vision-and-language benchmarks considered (25+ of them). Finally, we observe emerging capabilities, such as complex counting and multilingual object detection, tasks that are not explicitly in the training mix.

Armen Aghajanyan, Bernie Huang, Candace Ross et al.

We introduce CM3, a family of causally masked generative models trained over a large corpus of structured multi-modal documents that can contain both text and image tokens. Our new causally masked approach generates tokens left to right while also masking out a small number of long token spans that are generated at the end of the string, instead of their original positions. The casual masking object provides a type of hybrid of the more common causal and masked language models, by enabling full generative modeling while also providing bidirectional context when generating the masked spans. We train causally masked language-image models on large-scale web and Wikipedia articles, where each document contains all of the text, hypertext markup, hyperlinks, and image tokens (from a VQVAE-GAN), provided in the order they appear in the original HTML source (before masking). The resulting CM3 models can generate rich structured, multi-modal outputs while conditioning on arbitrary masked document contexts, and thereby implicitly learn a wide range of text, image, and cross modal tasks. They can be prompted to recover, in a zero-shot fashion, the functionality of models such as DALL-E, GENRE, and HTLM. We set the new state-of-the-art in zero-shot summarization, entity linking, and entity disambiguation while maintaining competitive performance in the fine-tuning setting. We can generate images unconditionally, conditioned on text (like DALL-E) and do captioning all in a zero-shot setting with a single model.

Armen Aghajanyan, Dmytro Okhonko, Mike Lewis et al.

We introduce HTLM, a hyper-text language model trained on a large-scale web crawl. Modeling hyper-text has a number of advantages: (1) it is easily gathered at scale, (2) it provides rich document-level and end-task-adjacent supervision (e.g. class and id attributes often encode document category information), and (3) it allows for new structured prompting that follows the established semantics of HTML (e.g. to do zero-shot summarization by infilling title tags for a webpage that contains the input text). We show that pretraining with a BART-style denoising loss directly on simplified HTML provides highly effective transfer for a wide range of end tasks and supervision levels. HTLM matches or exceeds the performance of comparably sized text-only LMs for zero-shot prompting and fine-tuning for classification benchmarks, while also setting new state-of-the-art performance levels for zero-shot summarization. We also find that hyper-text prompts provide more value to HTLM, in terms of data efficiency, than plain text prompts do for existing LMs, and that HTLM is highly effective at auto-prompting itself, by simply generating the most likely hyper-text formatting for any available training data. We will release all code and models to support future HTLM research.

Weijia Shi, Mandar Joshi, Luke Zettlemoyer

Short textual descriptions of entities provide summaries of their key attributes and have been shown to be useful sources of background knowledge for tasks such as entity linking and question answering. However, generating entity descriptions, especially for new and long-tail entities, can be challenging since relevant information is often scattered across multiple sources with varied content and style. We introduce DESCGEN: given mentions spread over multiple documents, the goal is to generate an entity summary description. DESCGEN consists of 37K entity descriptions from Wikipedia and Fandom, each paired with nine evidence documents on average. The documents were collected using a combination of entity linking and hyperlinks to the Wikipedia and Fandom entity pages, which together provide high-quality distant supervision. The resulting summaries are more abstractive than those found in existing datasets and provide a better proxy for the challenge of describing new and emerging entities. We also propose a two-stage extract-then-generate baseline and show that there exists a large gap (19.9% in ROUGE-L) between state-of-the-art models and human performance, suggesting that the data will support significant future work.

Arie Cattan, Alon Eirew, Gabriel Stanovsky et al.

We point out that common evaluation practices for cross-document coreference resolution have been unrealistically permissive in their assumed settings, yielding inflated results. We propose addressing this issue via two evaluation methodology principles. First, as in other tasks, models should be evaluated on predicted mentions rather than on gold mentions. Doing this raises a subtle issue regarding singleton coreference clusters, which we address by decoupling the evaluation of mention detection from that of coreference linking. Second, we argue that models should not exploit the synthetic topic structure of the standard ECB+ dataset, forcing models to confront the lexical ambiguity challenge, as intended by the dataset creators. We demonstrate empirically the drastic impact of our more realistic evaluation principles on a competitive model, yielding a score which is 33 F1 lower compared to evaluating by prior lenient practices.

Arie Cattan, Alon Eirew, Gabriel Stanovsky et al.

Coreference resolution has been mostly investigated within a single document scope, showing impressive progress in recent years based on end-to-end models. However, the more challenging task of cross-document (CD) coreference resolution remained relatively under-explored, with the few recent models applied only to gold mentions. Here, we introduce the first end-to-end model for CD coreference resolution from raw text, which extends the prominent model for within-document coreference to the CD setting. Our model achieves competitive results for event and entity coreference resolution on gold mentions. More importantly, we set first baseline results, on the standard ECB+ dataset, for CD coreference resolution over predicted mentions. Further, our model is simpler and more efficient than recent CD coreference resolution systems, while not using any external resources.

Terra Blevins, Mandar Joshi, Luke Zettlemoyer

Current models for Word Sense Disambiguation (WSD) struggle to disambiguate rare senses, despite reaching human performance on global WSD metrics. This stems from a lack of data for both modeling and evaluating rare senses in existing WSD datasets. In this paper, we introduce FEWS (Few-shot Examples of Word Senses), a new low-shot WSD dataset automatically extracted from example sentences in Wiktionary. FEWS has high sense coverage across different natural language domains and provides: (1) a large training set that covers many more senses than previous datasets and (2) a comprehensive evaluation set containing few- and zero-shot examples of a wide variety of senses. We establish baselines on FEWS with knowledge-based and neural WSD approaches and present transfer learning experiments demonstrating that models additionally trained with FEWS better capture rare senses in existing WSD datasets. Finally, we find humans outperform the best baseline models on FEWS, indicating that FEWS will support significant future work on low-shot WSD.

Arie Cattan, Alon Eirew, Gabriel Stanovsky et al.

Recent evaluation protocols for Cross-document (CD) coreference resolution have often been inconsistent or lenient, leading to incomparable results across works and overestimation of performance. To facilitate proper future research on this task, our primary contribution is proposing a pragmatic evaluation methodology which assumes access to only raw text -- rather than assuming gold mentions, disregards singleton prediction, and addresses typical targeted settings in CD coreference resolution. Aiming to set baseline results for future research that would follow our evaluation methodology, we build the first end-to-end model for this task. Our model adapts and extends recent neural models for within-document coreference resolution to address the CD coreference setting, which outperforms state-of-the-art results by a significant margin.

Bhargavi Paranjape, Mandar Joshi, John Thickstun et al.

Decisions of complex language understanding models can be rationalized by limiting their inputs to a relevant subsequence of the original text. A rationale should be as concise as possible without significantly degrading task performance, but this balance can be difficult to achieve in practice. In this paper, we show that it is possible to better manage this trade-off by optimizing a bound on the Information Bottleneck (IB) objective. Our fully unsupervised approach jointly learns an explainer that predicts sparse binary masks over sentences, and an end-task predictor that considers only the extracted rationale. Using IB, we derive a learning objective that allows direct control of mask sparsity levels through a tunable sparse prior. Experiments on ERASER benchmark tasks demonstrate significant gains over norm-minimization techniques for both task performance and agreement with human rationales. Furthermore, we find that in the semi-supervised setting, a modest amount of gold rationales (25% of training examples) closes the gap with a model that uses the full input.

Mandar Joshi, Kenton Lee, Yi Luan et al.

We present a method to represent input texts by contextualizing them jointly with dynamically retrieved textual encyclopedic background knowledge from multiple documents. We apply our method to reading comprehension tasks by encoding questions and passages together with background sentences about the entities they mention. We show that integrating background knowledge from text is effective for tasks focusing on factual reasoning and allows direct reuse of powerful pretrained BERT-style encoders. Moreover, knowledge integration can be further improved with suitable pretraining via a self-supervised masked language model objective over words in background-augmented input text. On TriviaQA, our approach obtains improvements of 1.6 to 3.1 F1 over comparable RoBERTa models which do not integrate background knowledge dynamically. On MRQA, a large collection of diverse QA datasets, we see consistent gains in-domain along with large improvements out-of-domain on BioASQ (2.1 to 4.2 F1), TextbookQA (1.6 to 2.0 F1), and DuoRC (1.1 to 2.0 F1).

Mandar Joshi, Omer Levy, Daniel S. Weld et al.

We apply BERT to coreference resolution, achieving strong improvements on the OntoNotes (+3.9 F1) and GAP (+11.5 F1) benchmarks. A qualitative analysis of model predictions indicates that, compared to ELMo and BERT-base, BERT-large is particularly better at distinguishing between related but distinct entities (e.g., President and CEO). However, there is still room for improvement in modeling document-level context, conversations, and mention paraphrasing. Our code and models are publicly available.

Yinhan Liu, Myle Ott, Naman Goyal et al.

Language model pretraining has led to significant performance gains but careful comparison between different approaches is challenging. Training is computationally expensive, often done on private datasets of different sizes, and, as we will show, hyperparameter choices have significant impact on the final results. We present a replication study of BERT pretraining (Devlin et al., 2019) that carefully measures the impact of many key hyperparameters and training data size. We find that BERT was significantly undertrained, and can match or exceed the performance of every model published after it. Our best model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results highlight the importance of previously overlooked design choices, and raise questions about the source of recently reported improvements. We release our models and code.

Mandar Joshi, Danqi Chen, Yinhan Liu et al.

We present SpanBERT, a pre-training method that is designed to better represent and predict spans of text. Our approach extends BERT by (1) masking contiguous random spans, rather than random tokens, and (2) training the span boundary representations to predict the entire content of the masked span, without relying on the individual token representations within it. SpanBERT consistently outperforms BERT and our better-tuned baselines, with substantial gains on span selection tasks such as question answering and coreference resolution. In particular, with the same training data and model size as BERT-large, our single model obtains 94.6% and 88.7% F1 on SQuAD 1.1 and 2.0, respectively. We also achieve a new state of the art on the OntoNotes coreference resolution task (79.6\% F1), strong performance on the TACRED relation extraction benchmark, and even show gains on GLUE.

Mandar Joshi, Eunsol Choi, Omer Levy et al.

Reasoning about implied relationships (e.g., paraphrastic, common sense, encyclopedic) between pairs of words is crucial for many cross-sentence inference problems. This paper proposes new methods for learning and using embeddings of word pairs that implicitly represent background knowledge about such relationships. Our pairwise embeddings are computed as a compositional function on word representations, which is learned by maximizing the pointwise mutual information (PMI) with the contexts in which the two words co-occur. We add these representations to the cross-sentence attention layer of existing inference models (e.g. BiDAF for QA, ESIM for NLI), instead of extending or replacing existing word embeddings. Experiments show a gain of 2.7% on the recently released SQuAD2.0 and 1.3% on MultiNLI. Our representations also aid in better generalization with gains of around 6-7% on adversarial SQuAD datasets, and 8.8% on the adversarial entailment test set by Glockner et al. (2018).