Peter Shaw

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.

Linlu Qiu, Peter Shaw, Panupong Pasupat et al. · mit

Despite their strong performance on many tasks, pre-trained language models have been shown to struggle on out-of-distribution compositional generalization. Meanwhile, recent work has shown considerable improvements on many NLP tasks from model scaling. Can scaling up model size also improve compositional generalization in semantic parsing? We evaluate encoder-decoder models up to 11B parameters and decoder-only models up to 540B parameters, and compare model scaling curves for three different methods for applying a pre-trained language model to a new task: fine-tuning all parameters, prompt tuning, and in-context learning. We observe that fine-tuning generally has flat or negative scaling curves on out-of-distribution compositional generalization in semantic parsing evaluations. In-context learning has positive scaling curves, but is generally outperformed by much smaller fine-tuned models. Prompt-tuning can outperform fine-tuning, suggesting further potential improvements from scaling as it exhibits a more positive scaling curve. Additionally, we identify several error trends that vary with model scale. For example, larger models are generally better at modeling the syntax of the output space, but are also more prone to certain types of overfitting. Overall, our study highlights limitations of current techniques for effectively leveraging model scale for compositional generalization, while our analysis also suggests promising directions for future work.

Yury Zemlyanskiy, Michiel de Jong, Joshua Ainslie et al. · mit

A common recent approach to semantic parsing augments sequence-to-sequence models by retrieving and appending a set of training samples, called exemplars. The effectiveness of this recipe is limited by the ability to retrieve informative exemplars that help produce the correct parse, which is especially challenging in low-resource settings. Existing retrieval is commonly based on similarity of query and exemplar inputs. We propose GandR, a retrieval procedure that retrieves exemplars for which outputs are also similar. GandRfirst generates a preliminary prediction with input-based retrieval. Then, it retrieves exemplars with outputs similar to the preliminary prediction which are used to generate a final prediction. GandR sets the state of the art on multiple low-resource semantic parsing tasks.

Jacob Eisenstein, Chirag Nagpal, Alekh Agarwal et al. · deepmind

Reward models play a key role in aligning language model applications towards human preferences. However, this setup creates an incentive for the language model to exploit errors in the reward model to achieve high estimated reward, a phenomenon often termed \emph{reward hacking}. A natural mitigation is to train an ensemble of reward models, aggregating over model outputs to obtain a more robust reward estimate. We explore the application of reward ensembles to alignment at both training time (through reinforcement learning) and inference time (through reranking). First, we show that reward models are \emph{underspecified}: reward models that perform similarly in-distribution can yield very different rewards when used in alignment, due to distribution shift. Second, underspecification results in overoptimization, where alignment to one reward model does not improve reward as measured by another reward model trained on the same data. Third, overoptimization is mitigated by the use of reward ensembles, and ensembles that vary by their \emph{pretraining} seeds lead to better generalization than ensembles that differ only by their \emph{fine-tuning} seeds, with both outperforming individual reward models. However, even pretrain reward ensembles do not eliminate reward hacking: we show several qualitative reward hacking phenomena that are not mitigated by ensembling because all reward models in the ensemble exhibit similar error patterns.

Xing Han Lù, Amirhossein Kazemnejad, Nicholas Meade et al. · eth-zurich, mila

Web agents enable users to perform tasks on web browsers through natural language interaction. Evaluating web agents trajectories is an important problem, since it helps us determine whether the agent successfully completed the tasks. Rule-based methods are widely used for this purpose, but they are challenging to extend to new tasks and may not always recognize successful trajectories. We may achieve higher accuracy through human evaluation, but the process would be substantially slower and more expensive. Automatic evaluations with LLMs may avoid the challenges of designing new rules and manually annotating trajectories, enabling faster and cost-effective evaluation. However, it is unclear how effective they are at evaluating web agents. To this end, we propose AgentRewardBench, the first benchmark to assess the effectiveness of LLM judges for evaluating web agents. AgentRewardBench contains 1302 trajectories across 5 benchmarks and 4 LLMs. Each trajectory in AgentRewardBench is reviewed by an expert, who answers questions pertaining to the success, side effects, and repetitiveness of the agent. Using our benchmark, we evaluate 12 LLM judges and find that no single LLM excels across all benchmarks. We also find that the rule-based evaluation used by common benchmarks tends to underreport the success rate of web agents, highlighting a key weakness of rule-based evaluation and the need to develop more flexible automatic evaluations. We release the benchmark at: https://agent-reward-bench.github.io

Peter Shaw, James Cohan, Jacob Eisenstein et al. · deepmind

We propose a new programming language called ALTA and a compiler that can map ALTA programs to Transformer weights. ALTA is inspired by RASP, a language proposed by Weiss et al. (2021), and Tracr (Lindner et al., 2023), a compiler from RASP programs to Transformer weights. ALTA complements and extends this prior work, offering the ability to express loops and to compile programs to Universal Transformers, among other advantages. ALTA allows us to constructively show how Transformers can represent length-invariant algorithms for computing parity and addition, as well as a solution to the SCAN benchmark of compositional generalization tasks, without requiring intermediate scratchpad decoding steps. We also propose tools to analyze cases where the expressibility of an algorithm is established, but end-to-end training on a given training set fails to induce behavior consistent with the desired algorithm. To this end, we explore training from ALTA execution traces as a more fine-grained supervision signal. This enables additional experiments and theoretical analyses relating the learnability of various algorithms to data availability and modeling decisions, such as positional encodings. We make the ALTA framework -- language specification, symbolic interpreter, and weight compiler -- available to the community to enable further applications and insights.

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, James Cohan, Jacob Eisenstein et al.

The Minimum Description Length (MDL) principle offers a formal framework for applying Occam's razor in machine learning. However, its application to neural networks such as Transformers is challenging due to the lack of a principled, universal measure for model complexity. This paper introduces the theoretical notion of asymptotically optimal description length objectives, grounded in the theory of Kolmogorov complexity. We establish that a minimizer of such an objective achieves optimal compression, for any dataset, up to an additive constant, in the limit as model resource bounds increase. We prove that asymptotically optimal objectives exist for Transformers, building on a new demonstration of their computational universality. We further show that such objectives can be tractable and differentiable by constructing and analyzing a variational objective based on an adaptive Gaussian mixture prior. Our empirical analysis shows that this variational objective selects for a low-complexity solution with strong generalization on an algorithmic task, but standard optimizers fail to find such solutions from a random initialization, highlighting key optimization challenges. More broadly, by providing a theoretical framework for identifying description length objectives with strong asymptotic guarantees, we outline a potential path towards training neural networks that achieve greater compression and generalization.

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.

Chaitanya Malaviya, Peter Shaw, Ming-Wei Chang et al.

Formulating selective information needs results in queries that implicitly specify set operations, such as intersection, union, and difference. For instance, one might search for "shorebirds that are not sandpipers" or "science-fiction films shot in England". To study the ability of retrieval systems to meet such information needs, we construct QUEST, a dataset of 3357 natural language queries with implicit set operations, that map to a set of entities corresponding to Wikipedia documents. The dataset challenges models to match multiple constraints mentioned in queries with corresponding evidence in documents and correctly perform various set operations. The dataset is constructed semi-automatically using Wikipedia category names. Queries are automatically composed from individual categories, then paraphrased and further validated for naturalness and fluency by crowdworkers. Crowdworkers also assess the relevance of entities based on their documents and highlight attribution of query constraints to spans of document text. We analyze several modern retrieval systems, finding that they often struggle on such queries. Queries involving negation and conjunction are particularly challenging and systems are further challenged with combinations of these operations.

Joseph R. Barr, Peter Shaw, Tyler Thatcher

We describe a workflow used to analyze the source code of the {\sc Android OS kernel} and rate for a particular kind of bugginess that exposes a program to hacking. The workflow represents a novel approach for components' vulnerability rating. The approach is inspired by recent work on embedding source code functions. The workflow combines deep learning with heuristics and machine learning. Deep learning is used to embed function/method labels into a Euclidean space. Because the corpus of Android kernel source code is rather limited (containing approximately 2 million C/C++ functions \& Java methods), a straightforward embedding is untenable. To overcome the challenge of the dearth of data, it's necessary to go through an intermediate step of the \textit{Byte-Pair Encoding}. Subsequently, we embed the tokens from which we assemble an embedding of function/method labels. Long short-term memory networks (LSTM) are used to embed tokens into vectors in $\mathbb{R}^d$ from which we form a \textit{cosine matrix} consisting of the cosine between every pair of vectors. The cosine matrix may be interpreted as a (combinatorial) `weighted' graph whose vertices represent functions/methods and `weighted' edges correspond to matrix entries. Features that include function vectors plus those defined heuristically are used to score for risk of bugginess.

Linlu Qiu, Peter Shaw, Panupong Pasupat et al.

Generic unstructured neural networks have been shown to struggle on out-of-distribution compositional generalization. Compositional data augmentation via example recombination has transferred some prior knowledge about compositionality to such black-box neural models for several semantic parsing tasks, but this often required task-specific engineering or provided limited gains. We present a more powerful data recombination method using a model called Compositional Structure Learner (CSL). CSL is a generative model with a quasi-synchronous context-free grammar backbone, which we induce from the training data. We sample recombined examples from CSL and add them to the fine-tuning data of a pre-trained sequence-to-sequence model (T5). This procedure effectively transfers most of CSL's compositional bias to T5 for diagnostic tasks, and results in a model even stronger than a T5-CSL ensemble on two real world compositional generalization tasks. This results in new state-of-the-art performance for these challenging semantic parsing tasks requiring generalization to both natural language variation and novel compositions of elements.

Wang Zhu, Peter Shaw, Tal Linzen et al.

Neural network models often generalize poorly to mismatched domains or distributions. In NLP, this issue arises in particular when models are expected to generalize compositionally, that is, to novel combinations of familiar words and constructions. We investigate learning representations that facilitate transfer learning from one compositional task to another: the representation and the task-specific layers of the models are strategically trained differently on a pre-finetuning task such that they generalize well on mismatched splits that require compositionality. We apply this method to semantic parsing, using three very different datasets, COGS, GeoQuery and SCAN, used alternately as the pre-finetuning and target task. Our method significantly improves compositional generalization over baselines on the test set of the target task, which is held out during fine-tuning. Ablation studies characterize the utility of the major steps in the proposed algorithm and support our hypothesis.

Joseph R. Barr, Peter Shaw, Marcus Sobel

It's regarded as an axiom that a good model is one that compromises between bias and variance. The bias is measured in training cost, while the variance of a (say, regression) model is measure by the cost associated with a validation set. If reducing bias is the goal, one will strive to fetch as complex a model as necessary, but complexity is invariably coupled with variance: greater complexity implies greater variance. In practice, driving training cost to near zero does not pose a fundamental problem; in fact, a sufficiently complex decision tree is perfectly capable of driving training cost to zero; however, the problem is often with controlling the model's variance. We investigate various regression model frameworks, including generalized linear models, Cox proportional hazard models, ARMA, and illustrate how misspecifying a model affects the variance.

Bowen Zhang, Hexiang Hu, Linlu Qiu et al.

We investigate ways to compose complex concepts in texts from primitive ones while grounding them in images. We propose Concept and Relation Graph (CRG), which builds on top of constituency analysis and consists of recursively combined concepts with predicate functions. Meanwhile, we propose a concept composition neural network called Composer to leverage the CRG for visually grounded concept learning. Specifically, we learn the grounding of both primitive and all composed concepts by aligning them to images and show that learning to compose leads to more robust grounding results, measured in text-to-image matching accuracy. Notably, our model can model grounded concepts forming at both the finer-grained sentence level and the coarser-grained intermediate level (or word-level). Composer leads to pronounced improvement in matching accuracy when the evaluation data has significant compound divergence from the training data.

Linlu Qiu, Hexiang Hu, Bowen Zhang et al.

We analyze the grounded SCAN (gSCAN) benchmark, which was recently proposed to study systematic generalization for grounded language understanding. First, we study which aspects of the original benchmark can be solved by commonly used methods in multi-modal research. We find that a general-purpose Transformer-based model with cross-modal attention achieves strong performance on a majority of the gSCAN splits, surprisingly outperforming more specialized approaches from prior work. Furthermore, our analysis suggests that many of the remaining errors reveal the same fundamental challenge in systematic generalization of linguistic constructs regardless of visual context. Second, inspired by this finding, we propose challenging new tasks for gSCAN by generating data to incorporate relations between objects in the visual environment. Finally, we find that current models are surprisingly data inefficient given the narrow scope of commands in gSCAN, suggesting another challenge for future work.

Jeremy R. Cole, Nanjiang Jiang, Panupong Pasupat et al.

The dominant paradigm for semantic parsing in recent years is to formulate parsing as a sequence-to-sequence task, generating predictions with auto-regressive sequence decoders. In this work, we explore an alternative paradigm. We formulate semantic parsing as a dependency parsing task, applying graph-based decoding techniques developed for syntactic parsing. We compare various decoding techniques given the same pre-trained Transformer encoder on the TOP dataset, including settings where training data is limited or contains only partially-annotated examples. We find that our graph-based approach is competitive with sequence decoders on the standard setting, and offers significant improvements in data efficiency and settings where partially-annotated data is available.

Jonathan Herzig, Peter Shaw, Ming-Wei Chang et al.

Sequence-to-sequence (seq2seq) models are prevalent in semantic parsing, but have been found to struggle at out-of-distribution compositional generalization. While specialized model architectures and pre-training of seq2seq models have been proposed to address this issue, the former often comes at the cost of generality and the latter only shows limited success. In this paper, we study the impact of intermediate representations on compositional generalization in pre-trained seq2seq models, without changing the model architecture at all, and identify key aspects for designing effective representations. Instead of training to directly map natural language to an executable form, we map to a reversible or lossy intermediate representation that has stronger structural correspondence with natural language. The combination of our proposed intermediate representations and pre-trained models is surprisingly effective, where the best combinations obtain a new state-of-the-art on CFQ (+14.8 accuracy points) and on the template-splits of three text-to-SQL datasets (+15.0 to +19.4 accuracy points). This work highlights that intermediate representations provide an important and potentially overlooked degree of freedom for improving the compositional generalization abilities of pre-trained seq2seq models.

Peter Shaw, Ming-Wei Chang, Panupong Pasupat et al.

Sequence-to-sequence models excel at handling natural language variation, but have been shown to struggle with out-of-distribution compositional generalization. This has motivated new specialized architectures with stronger compositional biases, but most of these approaches have only been evaluated on synthetically-generated datasets, which are not representative of natural language variation. In this work we ask: can we develop a semantic parsing approach that handles both natural language variation and compositional generalization? To better assess this capability, we propose new train and test splits of non-synthetic datasets. We demonstrate that strong existing approaches do not perform well across a broad set of evaluations. We also propose NQG-T5, a hybrid model that combines a high-precision grammar-based approach with a pre-trained sequence-to-sequence model. It outperforms existing approaches across several compositional generalization challenges on non-synthetic data, while also being competitive with the state-of-the-art on standard evaluations. While still far from solving this problem, our study highlights the importance of diverse evaluations and the open challenge of handling both compositional generalization and natural language variation in semantic parsing.

Zhen Zeng, Yuefeng Lu, Judong Shen et al.

Precision medicine is becoming a focus in medical research recently, as its implementation brings values to all stakeholders in the healthcare system. Various statistical methodologies have been developed tackling problems in different aspects of this field, e.g., assessing treatment heterogeneity, identifying patient subgroups, or building treatment decision models. However, there is a lack of new tools devoted to selecting and prioritizing predictive biomarkers. We propose a novel tree-based ensemble method, random interaction forest (RIF), to generate predictive importance scores and prioritize candidate biomarkers for constructing refined treatment decision models. RIF was evaluated by comparing with the conventional random forest and univariable regression methods and showed favorable properties under various simulation scenarios. We applied the proposed RIF method to a biomarker dataset from two phase III clinical trials of bezlotoxumab on $\textit{Clostridium difficile}$ infection recurrence and obtained biologically meaningful results.

Thomas Müller, Francesco Piccinno, Massimo Nicosia et al.

We present a novel approach to answering sequential questions based on structured objects such as knowledge bases or tables without using a logical form as an intermediate representation. We encode tables as graphs using a graph neural network model based on the Transformer architecture. The answers are then selected from the encoded graph using a pointer network. This model is appropriate for processing conversations around structured data, where the attention mechanism that selects the answers to a question can also be used to resolve conversational references. We demonstrate the validity of this approach with competitive results on the Sequential Question Answering (SQA) task (Iyyer et al., 2017).

Peter Shaw, Philip Massey, Angelica Chen et al.

Structured information about entities is critical for many semantic parsing tasks. We present an approach that uses a Graph Neural Network (GNN) architecture to incorporate information about relevant entities and their relations during parsing. Combined with a decoder copy mechanism, this approach provides a conceptually simple mechanism to generate logical forms with entities. We demonstrate that this approach is competitive with the state-of-the-art across several tasks without pre-training, and outperforms existing approaches when combined with BERT pre-training.

Peter Shaw, Jakob Uszkoreit, Ashish Vaswani

Relying entirely on an attention mechanism, the Transformer introduced by Vaswani et al. (2017) achieves state-of-the-art results for machine translation. In contrast to recurrent and convolutional neural networks, it does not explicitly model relative or absolute position information in its structure. Instead, it requires adding representations of absolute positions to its inputs. In this work we present an alternative approach, extending the self-attention mechanism to efficiently consider representations of the relative positions, or distances between sequence elements. On the WMT 2014 English-to-German and English-to-French translation tasks, this approach yields improvements of 1.3 BLEU and 0.3 BLEU over absolute position representations, respectively. Notably, we observe that combining relative and absolute position representations yields no further improvement in translation quality. We describe an efficient implementation of our method and cast it as an instance of relation-aware self-attention mechanisms that can generalize to arbitrary graph-labeled inputs.