Pengcheng Yin

Shuyan Zhou, Li Zhang, Yue Yang et al. · allen-ai, cmu

Procedures are inherently hierarchical. To "make videos", one may need to "purchase a camera", which in turn may require one to "set a budget". While such hierarchical knowledge is critical for reasoning about complex procedures, most existing work has treated procedures as shallow structures without modeling the parent-child relation. In this work, we attempt to construct an open-domain hierarchical knowledge-base (KB) of procedures based on wikiHow, a website containing more than 110k instructional articles, each documenting the steps to carry out a complex procedure. To this end, we develop a simple and efficient method that links steps (e.g., "purchase a camera") in an article to other articles with similar goals (e.g., "how to choose a camera"), recursively constructing the KB. Our method significantly outperforms several strong baselines according to automatic evaluation, human judgment, and application to downstream tasks such as instructional video retrieval. A demo with partial data can be found at https://wikihow-hierarchy.github.io. The code and the data are at https://github.com/shuyanzhou/wikihow_hierarchy.

Aakanksha Chowdhery, Sharan Narang, Jacob Devlin et al. · deepmind, stanford

Large language models have been shown to achieve remarkable performance across a variety of natural language tasks using few-shot learning, which drastically reduces the number of task-specific training examples needed to adapt the model to a particular application. To further our understanding of the impact of scale on few-shot learning, we trained a 540-billion parameter, densely activated, Transformer language model, which we call Pathways Language Model PaLM. We trained PaLM on 6144 TPU v4 chips using Pathways, a new ML system which enables highly efficient training across multiple TPU Pods. We demonstrate continued benefits of scaling by achieving state-of-the-art few-shot learning results on hundreds of language understanding and generation benchmarks. On a number of these tasks, PaLM 540B achieves breakthrough performance, outperforming the finetuned state-of-the-art on a suite of multi-step reasoning tasks, and outperforming average human performance on the recently released BIG-bench benchmark. A significant number of BIG-bench tasks showed discontinuous improvements from model scale, meaning that performance steeply increased as we scaled to our largest model. PaLM also has strong capabilities in multilingual tasks and source code generation, which we demonstrate on a wide array of benchmarks. We additionally provide a comprehensive analysis on bias and toxicity, and study the extent of training data memorization with respect to model scale. Finally, we discuss the ethical considerations related to large language models and discuss potential mitigation strategies.

Pengcheng Yin, Wen-Ding Li, Kefan Xiao et al. · stanford

Computational notebooks, such as Jupyter notebooks, are interactive computing environments that are ubiquitous among data scientists to perform data wrangling and analytic tasks. To measure the performance of AI pair programmers that automatically synthesize programs for those tasks given natural language (NL) intents from users, we build ARCADE, a benchmark of 1082 code generation problems using the pandas data analysis framework in data science notebooks. ARCADE features multiple rounds of NL-to-code problems from the same notebook. It requires a model to understand rich multi-modal contexts, such as existing notebook cells and their execution states as well as previous turns of interaction. To establish a strong baseline on this challenging task, we develop PaChiNCo, a 62B code language model (LM) for Python computational notebooks, which significantly outperforms public code LMs. Finally, we explore few-shot prompting strategies to elicit better code with step-by-step decomposition and NL explanation, showing the potential to improve the diversity and explainability of model predictions.

Ansong Ni, Pengcheng Yin, Yilun Zhao et al. · salesforce

Recently, large language models (LLMs), especially those that are pretrained on code, have demonstrated strong capabilities in generating programs from natural language inputs in a few-shot or even zero-shot manner. Despite promising results, there is a notable lack of a comprehensive evaluation of these models language-to-code generation capabilities. Existing studies often focus on specific tasks, model architectures, or learning paradigms, leading to a fragmented understanding of the overall landscape. In this work, we present L2CEval, a systematic evaluation of the language-to-code generation capabilities of LLMs on 7 tasks across the domain spectrum of semantic parsing, math reasoning and Python programming, analyzing the factors that potentially affect their performance, such as model size, pretraining data, instruction tuning, and different prompting methods. In addition to assessing model performance, we measure confidence calibration for the models and conduct human evaluations of the output programs. This enables us to identify and analyze the typical failure modes across various tasks and models. L2CEval offers a comprehensive understanding of the capabilities and limitations of LLMs in language-to-code generation. We also release the evaluation framework and all model outputs, hoping to lay the groundwork for further future research in this domain.

Xinyun Chen, Renat Aksitov, Uri Alon et al. · deepmind

Self-consistency with chain-of-thought prompting (CoT) has demonstrated remarkable performance gains on various challenging tasks, by utilizing multiple reasoning paths sampled from large language models (LLMs). However, self-consistency relies on the answer extraction process to aggregate multiple solutions, which is not applicable to free-form answers. In this work, we propose Universal Self-Consistency (USC), which leverages LLMs themselves to select the most consistent answer among multiple candidates. We evaluate USC on a variety of benchmarks, including mathematical reasoning, code generation, long-context summarization, and open-ended question answering. On open-ended generation tasks where the original self-consistency method is not applicable, USC effectively utilizes multiple samples and improves the performance. For mathematical reasoning, USC matches the standard self-consistency performance without requiring the answer formats to be similar. Finally, without access to execution results, USC also matches the execution-based voting performance on code generation.

Ruisheng Cao, Fangyu Lei, Haoyuan Wu et al. · tsinghua

Data science and engineering workflows often span multiple stages, from warehousing to orchestration, using tools like BigQuery, dbt, and Airbyte. As vision language models (VLMs) advance in multimodal understanding and code generation, VLM-based agents could potentially automate these workflows by generating SQL queries, Python code, and GUI operations. This automation can improve the productivity of experts while democratizing access to large-scale data analysis. In this paper, we introduce Spider2-V, the first multimodal agent benchmark focusing on professional data science and engineering workflows, featuring 494 real-world tasks in authentic computer environments and incorporating 20 enterprise-level professional applications. These tasks, derived from real-world use cases, evaluate the ability of a multimodal agent to perform data-related tasks by writing code and managing the GUI in enterprise data software systems. To balance realistic simulation with evaluation simplicity, we devote significant effort to developing automatic configurations for task setup and carefully crafting evaluation metrics for each task. Furthermore, we supplement multimodal agents with comprehensive documents of these enterprise data software systems. Our empirical evaluation reveals that existing state-of-the-art LLM/VLM-based agents do not reliably automate full data workflows (14.0% success). Even with step-by-step guidance, these agents still underperform in tasks that require fine-grained, knowledge-intensive GUI actions (16.2%) and involve remote cloud-hosted workspaces (10.6%). We hope that Spider2-V paves the way for autonomous multimodal agents to transform the automation of data science and engineering workflow. Our code and data are available at https://spider2-v.github.io.

Ruoxi Sun, Sercan Ö. Arik, Rajarishi Sinha et al.

Text-to-SQL aims to automate the process of generating SQL queries on a database from natural language text. In this work, we propose "SQLPrompt", tailored to improve the few-shot prompting capabilities of Text-to-SQL for Large Language Models (LLMs). Our methods include innovative prompt design, execution-based consistency decoding strategy which selects the SQL with the most consistent execution outcome among other SQL proposals, and a method that aims to improve performance by diversifying the SQL proposals during consistency selection with different prompt designs ("MixPrompt") and foundation models ("MixLLMs"). We show that \emph{SQLPrompt} outperforms previous approaches for in-context learning with few labeled data by a large margin, closing the gap with finetuning state-of-the-art with thousands of labeled data.

Kensen Shi, Joey Hong, Yinlin Deng et al.

When writing programs, people have the ability to tackle a new complex task by decomposing it into smaller and more familiar subtasks. While it is difficult to measure whether neural program synthesis methods have similar capabilities, we can measure whether they compositionally generalize, that is, whether a model that has been trained on the simpler subtasks is subsequently able to solve more complex tasks. In this paper, we characterize several different forms of compositional generalization that are desirable in program synthesis, forming a meta-benchmark which we use to create generalization tasks for two popular datasets, RobustFill and DeepCoder. We then propose ExeDec, a novel decomposition-based synthesis strategy that predicts execution subgoals to solve problems step-by-step informed by program execution at each step. When used with Transformer models trained from scratch, ExeDec has better synthesis performance and greatly improved compositional generalization ability compared to baselines. Finally, we use our benchmarks to demonstrate that LLMs struggle to compositionally generalize when asked to do programming-by-example in a few-shot setting, but an ExeDec-style prompting approach can improve the generalization ability and overall performance.

Kensen Shi, Deniz Altınbüken, Saswat Anand et al.

We propose using natural language outlines as a novel modality and interaction surface for providing AI assistance to developers throughout the software development process. An NL outline for a code function comprises multiple statements written in concise prose, which partition the code and summarize its main ideas in the style of literate programming. Crucially, we find that modern LLMs can generate accurate and high-quality NL outlines in practice. Moreover, NL outlines enable a bidirectional sync between code and NL, where a developer can change either code or NL and have the LLM automatically update the other. We discuss many use cases for NL outlines: they can accelerate understanding and navigation of code and diffs, simplify code maintenance, augment code search, steer code generation, and more. We then propose and compare multiple LLM prompting techniques for generating outlines and ask professional developers to judge outline quality. Finally, we present two case studies applying NL outlines toward code review and malware detection.

Ningyi Liao, Dingheng Mo, Siqiang Luo et al.

Recent advances in data processing have stimulated the demand for learning graphs of very large scales. Graph Neural Networks (GNNs), being an emerging and powerful approach in solving graph learning tasks, are known to be difficult to scale up. Most scalable models apply node-based techniques in simplifying the expensive graph message-passing propagation procedure of GNN. However, we find such acceleration insufficient when applied to million- or even billion-scale graphs. In this work, we propose SCARA, a scalable GNN with feature-oriented optimization for graph computation. SCARA efficiently computes graph embedding from node features, and further selects and reuses feature computation results to reduce overhead. Theoretical analysis indicates that our model achieves sub-linear time complexity with a guaranteed precision in propagation process as well as GNN training and inference. We conduct extensive experiments on various datasets to evaluate the efficacy and efficiency of SCARA. Performance comparison with baselines shows that SCARA can reach up to 100x graph propagation acceleration than current state-of-the-art methods with fast convergence and comparable accuracy. Most notably, it is efficient to process precomputation on the largest available billion-scale GNN dataset Papers100M (111M nodes, 1.6B edges) in 100 seconds.

Kensen Shi, Joey Hong, Manzil Zaheer et al.

When writing programs, people have the ability to tackle a new complex task by decomposing it into smaller and more familiar subtasks. While it is difficult to measure whether neural program synthesis methods have similar capabilities, what we can measure is whether they compositionally generalize, that is, whether a model that has been trained on the simpler subtasks is subsequently able to solve more complex tasks. In this paper, we focus on measuring the ability of learned program synthesizers to compositionally generalize. We first characterize several different axes along which program synthesis methods would be desired to generalize, e.g., length generalization, or the ability to combine known subroutines in new ways that do not occur in the training data. Based on this characterization, we introduce a benchmark suite of tasks to assess these abilities based on two popular existing datasets, SCAN and RobustFill. Finally, we make first attempts to improve the compositional generalization ability of Transformer models along these axes through novel attention mechanisms that draw inspiration from a human-like decomposition strategy. Empirically, we find our modified Transformer models generally perform better than natural baselines, but the tasks remain challenging.

Fangyu Lei, Jixuan Chen, Yuxiao Ye et al. · tsinghua

Real-world enterprise text-to-SQL workflows often involve complex cloud or local data across various database systems, multiple SQL queries in various dialects, and diverse operations from data transformation to analytics. We introduce Spider 2.0, an evaluation framework comprising 632 real-world text-to-SQL workflow problems derived from enterprise-level database use cases. The databases in Spider 2.0 are sourced from real data applications, often containing over 1,000 columns and stored in local or cloud database systems such as BigQuery and Snowflake. We show that solving problems in Spider 2.0 frequently requires understanding and searching through database metadata, dialect documentation, and even project-level codebases. This challenge calls for models to interact with complex SQL workflow environments, process extremely long contexts, perform intricate reasoning, and generate multiple SQL queries with diverse operations, often exceeding 100 lines, which goes far beyond traditional text-to-SQL challenges. Our evaluations indicate that based on o1-preview, our code agent framework successfully solves only 21.3% of the tasks, compared with 91.2% on Spider 1.0 and 73.0% on BIRD. Our results on Spider 2.0 show that while language models have demonstrated remarkable performance in code generation -- especially in prior text-to-SQL benchmarks -- they require significant improvement in order to achieve adequate performance for real-world enterprise usage. Progress on Spider 2.0 represents crucial steps towards developing intelligent, autonomous, code agents for real-world enterprise settings. Our code, baseline models, and data are available at https://spider2-sql.github.io

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.

Celine Lee, Jing Nathan Yan, Chen Liang et al.

Modern language models have historically relied on two dominant design choices: subword tokenization and autoregressive (AR) ordering. These design decisions bake in priors that dictate a model's learning. Recently, two alternative paradigms have challenged this: byte-level modeling, which bypasses static statistically-derived token vocabularies, and masked diffusion modeling (MDM), which conducts parallel, non-sequential generation. Their intersection represents a fully end-to-end modality-agnostic generative prototype; however, removing these structural priors incurs a significant computational cost. In this work, we investigate this cost through a compute-matched scaling study. Our results reveal that the performance penalty of byte modeling is not uniform; across scale, the scaling overhead of byte modeling is worse for MDM than for AR. We hypothesize that this disparity stems from context fragility: while AR's stable causal history allows models to naturally rediscover subword patterns, the MDM objective destroys the local contiguity required to efficiently resolve semantics from raw bytes. Our findings from controlled permutation experiments suggest that future modality-agnostic designs must incorporate alternative structural biases to maintain viable scaling trajectories in the byte regime.

John Yang, Kilian Lieret, Jeffrey Ma et al.

Turning ideas into full software projects from scratch has become a popular use case for language models. Agents are being deployed to seed, maintain, and grow codebases over extended periods with minimal human oversight. Such settings require models to make high-level software architecture decisions. However, existing benchmarks measure focused, limited tasks such as fixing a single bug or developing a single, specified feature. We therefore introduce ProgramBench to measure the ability of software engineering agents to develop software holisitically. In ProgramBench, given only a program and its documentation, agents must architect and implement a codebase that matches the reference executable's behavior. End-to-end behavioral tests are generated via agent-driven fuzzing, enabling evaluation without prescribing implementation structure. Our 200 tasks range from compact CLI tools to widely used software such as FFmpeg, SQLite, and the PHP interpreter. We evaluate 9 LMs and find that none fully resolve any task, with the best model passing 95\% of tests on only 3\% of tasks. Models favor monolithic, single-file implementations that diverge sharply from human-written code.

Qiushi Sun, Zhirui Chen, Fangzhi Xu et al.

Neural Code Intelligence -- leveraging deep learning to understand, generate, and optimize code -- holds immense potential for transformative impacts on the whole society. Bridging the gap between Natural Language and Programming Language, this domain has drawn significant attention from researchers in both research communities over the past few years. This survey presents a systematic and chronological review of the advancements in code intelligence, encompassing over 50 representative models and their variants, more than 20 categories of tasks, and an extensive coverage of over 680 related works. We follow the historical progression to trace the paradigm shifts across different research phases (e.g., from modeling code with recurrent neural networks to the era of Large Language Models). Concurrently, we highlight the major technical transitions in models, tasks, and evaluations spanning through different stages. For applications, we also observe a co-evolving shift. It spans from initial endeavors to tackling specific scenarios, through exploring a diverse array of tasks during its rapid expansion, to currently focusing on tackling increasingly complex and varied real-world challenges. Building on our examination of the developmental trajectories, we further investigate the emerging synergies between code intelligence and broader machine intelligence, uncovering new cross-domain opportunities and illustrating the substantial influence of code intelligence across various domains. Finally, we delve into both the opportunities and challenges associated with this field, alongside elucidating our insights on the most promising research directions. An ongoing, dynamically updated project and resources associated with this survey have been released at https://github.com/QiushiSun/Awesome-Code-Intelligence.

Tianbao Xie, Chen Henry Wu, Peng Shi et al.

Structured knowledge grounding (SKG) leverages structured knowledge to complete user requests, such as semantic parsing over databases and question answering over knowledge bases. Since the inputs and outputs of SKG tasks are heterogeneous, they have been studied separately by different communities, which limits systematic and compatible research on SKG. In this paper, we overcome this limitation by proposing the UnifiedSKG framework, which unifies 21 SKG tasks into a text-to-text format, aiming to promote systematic SKG research, instead of being exclusive to a single task, domain, or dataset. We use UnifiedSKG to benchmark T5 with different sizes and show that T5, with simple modifications when necessary, achieves state-of-the-art performance on almost all of the 21 tasks. We further demonstrate that multi-task prefix-tuning improves the performance on most tasks, largely improving the overall performance. UnifiedSKG also facilitates the investigation of zero-shot and few-shot learning, and we show that T0, GPT-3, and Codex struggle in zero-shot and few-shot learning for SKG. We also use UnifiedSKG to conduct a series of controlled experiments on structured knowledge encoding variants across SKG tasks. UnifiedSKG is easily extensible to more tasks, and it is open-sourced at https://github.com/hkunlp/unifiedskg.

Alex Shypula, Pengcheng Yin, Jeremy Lacomis et al.

Program optimization is the process of modifying software to execute more efficiently. Superoptimizers attempt to find the optimal program by employing significantly more expensive search and constraint solving techniques. Generally, these methods do not scale well to programs in real development scenarios, and as a result, superoptimization has largely been confined to small-scale, domain-specific, and/or synthetic program benchmarks. In this paper, we propose a framework to learn to superoptimize real-world programs by using neural sequence-to-sequence models. We created a dataset consisting of over 25K real-world x86-64 assembly functions mined from open-source projects and propose an approach, Self Imitation Learning for Optimization (SILO) that is easy to implement and outperforms a standard policy gradient learning approach on our dataset. Our method, SILO, superoptimizes 5.9% of our test set when compared with the gcc version 10.3 compiler's aggressive optimization level -O3. We also report that SILO's rate of superoptimization on our test set is over five times that of a standard policy gradient approach and a model pre-trained on compiler optimization demonstration.

Frank F. Xu, Zhengbao Jiang, Pengcheng Yin et al.

Open-domain code generation aims to generate code in a general-purpose programming language (such as Python) from natural language (NL) intents. Motivated by the intuition that developers usually retrieve resources on the web when writing code, we explore the effectiveness of incorporating two varieties of external knowledge into NL-to-code generation: automatically mined NL-code pairs from the online programming QA forum StackOverflow and programming language API documentation. Our evaluations show that combining the two sources with data augmentation and retrieval-based data re-sampling improves the current state-of-the-art by up to 2.2% absolute BLEU score on the code generation testbed CoNaLa. The code and resources are available at https://github.com/neulab/external-knowledge-codegen.

Zhengbao Jiang, Pengcheng Yin, Graham Neubig

Open information extraction (IE) is the task of extracting open-domain assertions from natural language sentences. A key step in open IE is confidence modeling, ranking the extractions based on their estimated quality to adjust precision and recall of extracted assertions. We found that the extraction likelihood, a confidence measure used by current supervised open IE systems, is not well calibrated when comparing the quality of assertions extracted from different sentences. We propose an additional binary classification loss to calibrate the likelihood to make it more globally comparable, and an iterative learning process, where extractions generated by the open IE model are incrementally included as training samples to help the model learn from trial and error. Experiments on OIE2016 demonstrate the effectiveness of our method. Code and data are available at https://github.com/jzbjyb/oie_rank.

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.

Hongjin Su, Ruoxi Sun, Jinsung Yoon et al.

Autonomous agents powered by large language models (LLMs) have the potential to enhance human capabilities, assisting with digital tasks from sending emails to performing data analysis. The abilities of existing LLMs at such tasks are often hindered by the lack of high-quality agent data from the corresponding environments they interact with. We propose Learn-by-interact, a data-centric framework to adapt LLM agents to any given environments without human annotations. Learn-by-interact synthesizes trajectories of agent-environment interactions based on documentations, and constructs instructions by summarizing or abstracting the interaction histories, a process called backward construction. We assess the quality of our synthetic data by using them in both training-based scenarios and training-free in-context learning (ICL), where we craft innovative retrieval approaches optimized for agents. Extensive experiments on SWE-bench, WebArena, OSWorld and Spider2-V spanning across realistic coding, web, and desktop environments show the effectiveness of Learn-by-interact in various downstream agentic tasks -- baseline results are improved by up to 12.2\% for ICL with Claude-3.5 and 19.5\% for training with Codestral-22B. We further demonstrate the critical role of backward construction, which provides up to 14.0\% improvement for training. Our ablation studies demonstrate the efficiency provided by our synthesized data in ICL and the superiority of our retrieval pipeline over alternative approaches like conventional retrieval-augmented generation (RAG). We expect that Learn-by-interact will serve as a foundation for agent data synthesis as LLMs are increasingly deployed at real-world environments.

Ansong Ni, Miltiadis Allamanis, Arman Cohan et al. · cambridge, microsoft-research

A fundamental skill among human developers is the ability to understand and reason about program execution. As an example, a programmer can mentally simulate code execution in natural language to debug and repair code (aka. rubber duck debugging). However, large language models (LLMs) of code are typically trained on the surface textual form of programs, thus may lack a semantic understanding of how programs execute at run-time. To address this issue, we propose NExT, a method to teach LLMs to inspect the execution traces of programs (variable states of executed lines) and reason about their run-time behavior through chain-of-thought (CoT) rationales. Specifically, NExT uses self-training to bootstrap a synthetic training set of execution-aware rationales that lead to correct task solutions (e.g., fixed programs) without laborious manual annotation. Experiments on program repair tasks based on MBPP and HumanEval demonstrate that NExT improves the fix rate of a PaLM 2 model, by 26.1% and 14.3% absolute, respectively, with significantly improved rationale quality as verified by automated metrics and human raters. Our model can also generalize to scenarios where program traces are absent at test-time.

Yeming Wen, Pengcheng Yin, Kensen Shi et al.

Large language models (LLMs) have recently demonstrated a remarkable ability to generate code from natural language (NL) prompts. However, in the real world, NL is often too ambiguous to capture the true intent behind programming problems, requiring additional input-output (I/O) specifications. Unfortunately, LLMs can have difficulty aligning their outputs with both the NL prompt and the I/O specification. In this paper, we give a way to mitigate this issue in the context of data science programming, where tasks require explicit I/O specifications for clarity. Specifically, we propose GIFT4Code, a novel approach for the instruction fine-tuning of LLMs with respect to I/O specifications. Our method leverages synthetic data produced by the LLM itself and utilizes execution-derived feedback as a key learning signal. This feedback, in the form of program I/O specifications, is provided to the LLM to facilitate instruction fine-tuning. We evaluated our approach on two challenging data science benchmarks, Arcade and DS-1000. The results demonstrate a significant improvement in the LLM's ability to generate code that is not only executable but also accurately aligned with user specifications, substantially improving the quality of code generation for complex data science tasks.

Miltiadis Allamanis, Sheena Panthaplackel, Pengcheng Yin · cambridge, microsoft-research

To evaluate code large language models (LLMs), research has relied on a few small manually curated benchmarks, such as HumanEval and MBPP, which represent a narrow part of the real-world software domains. In this work, we introduce round-trip correctness (RTC) as an alternative evaluation method. RTC allows Code LLM evaluation on a broader spectrum of real-world software domains without the need for costly human curation. RTC rests on the idea that we can ask a model to make a prediction (e.g., describe some code using natural language), feed that prediction back (e.g., synthesize code from the predicted description), and check if this round-trip leads to code that is semantically equivalent to the original input. We show how to employ RTC to evaluate code synthesis and editing. We find that RTC strongly correlates with model performance on existing narrow-domain code synthesis benchmarks while allowing us to expand to a much broader set of domains and tasks which was not previously possible without costly human annotations.

Miltiadis Allamanis, Pengcheng Yin · cambridge, microsoft-research

To evaluate large language models (LLMs) for code, research has used manually created unit test-based benchmarks. However, these tests are often inadequate, missing corner cases and other implementation-specific oddities. This work introduces ProbeGen, a whitebox method that takes two or more executable pieces of code and searches for counterexamples to their equivalence. Comparing code semantics requires a deep understanding of code. We demonstrate that LLMs with execution feedback perform well at this task. In a common code synthesis benchmark, ProbeGen disproves 18% of samples considered equivalent to the ground truth by the benchmark-provided unit tests. Additionally, using ProbeGen, we can semantically cluster LLM samples for semantic self-consistency, improving pass@1 by 10% by unifying syntactically distinct but semantically similar samples.

Hanjun Dai, Bethany Yixin Wang, Xingchen Wan et al.

Analytics on structured data is a mature field with many successful methods. However, most real world data exists in unstructured form, such as images and conversations. We investigate the potential of Large Language Models (LLMs) to enable unstructured data analytics. In particular, we propose a new Universal Query Engine (UQE) that directly interrogates and draws insights from unstructured data collections. This engine accepts queries in a Universal Query Language (UQL), a dialect of SQL that provides full natural language flexibility in specifying conditions and operators. The new engine leverages the ability of LLMs to conduct analysis of unstructured data, while also allowing us to exploit advances in sampling and optimization techniques to achieve efficient and accurate query execution. In addition, we borrow techniques from classical compiler theory to better orchestrate the workflow between sampling methods and foundation model calls. We demonstrate the efficiency of UQE on data analytics across different modalities, including images, dialogs and reviews, across a range of useful query types, including conditional aggregation, semantic retrieval and abstraction aggregation.

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Ruoxi Sun, Sercan Ö. Arik, Alex Muzio et al.

Text-to-SQL, the process of translating natural language into Structured Query Language (SQL), represents a transformative application of large language models (LLMs), potentially revolutionizing how humans interact with data. This paper introduces the SQL-PaLM framework, a comprehensive solution for understanding and enhancing Text-to-SQL using LLMs, using in the learning regimes of few-shot prompting and instruction fine-tuning. With few-shot prompting, we explore the effectiveness of consistency decoding with execution-based error filtering. With instruction fine-tuning, we delve deep in understanding the critical paradigms that influence the performance of tuned LLMs. In particular, we investigate how performance can be improved through expanded training data coverage and diversity, synthetic data augmentation, and integrating query-specific database content. We propose a test-time selection method to further refine accuracy by integrating SQL outputs from multiple paradigms with execution feedback as guidance. Additionally, we tackle the practical challenge of navigating intricate databases with a significant number of tables and columns, proposing efficient techniques for accurately selecting relevant database elements to enhance Text-to-SQL performance. Our holistic approach yields substantial advancements in Text-to-SQL, as demonstrated on two key public benchmarks, Spider and BIRD. Through comprehensive ablations and error analyses, we shed light on the strengths and weaknesses of our framework, offering valuable insights into Text-to-SQL's future work.

Rohan Anil, Andrew M. Dai, Orhan Firat et al.

We introduce PaLM 2, a new state-of-the-art language model that has better multilingual and reasoning capabilities and is more compute-efficient than its predecessor PaLM. PaLM 2 is a Transformer-based model trained using a mixture of objectives. Through extensive evaluations on English and multilingual language, and reasoning tasks, we demonstrate that PaLM 2 has significantly improved quality on downstream tasks across different model sizes, while simultaneously exhibiting faster and more efficient inference compared to PaLM. This improved efficiency enables broader deployment while also allowing the model to respond faster, for a more natural pace of interaction. PaLM 2 demonstrates robust reasoning capabilities exemplified by large improvements over PaLM on BIG-Bench and other reasoning tasks. PaLM 2 exhibits stable performance on a suite of responsible AI evaluations, and enables inference-time control over toxicity without additional overhead or impact on other capabilities. Overall, PaLM 2 achieves state-of-the-art performance across a diverse set of tasks and capabilities. When discussing the PaLM 2 family, it is important to distinguish between pre-trained models (of various sizes), fine-tuned variants of these models, and the user-facing products that use these models. In particular, user-facing products typically include additional pre- and post-processing steps. Additionally, the underlying models may evolve over time. Therefore, one should not expect the performance of user-facing products to exactly match the results reported in this report.

Pengcheng Yin, John Wieting, Avirup Sil et al.

Semantic parsers map natural language utterances into meaning representations (e.g., programs). Such models are typically bottlenecked by the paucity of training data due to the required laborious annotation efforts. Recent studies have performed zero-shot learning by synthesizing training examples of canonical utterances and programs from a grammar, and further paraphrasing these utterances to improve linguistic diversity. However, such synthetic examples cannot fully capture patterns in real data. In this paper we analyze zero-shot parsers through the lenses of the language and logical gaps (Herzig and Berant, 2019), which quantify the discrepancy of language and programmatic patterns between the canonical examples and real-world user-issued ones. We propose bridging these gaps using improved grammars, stronger paraphrasers, and efficient learning methods using canonical examples that most likely reflect real user intents. Our model achieves strong performance on two semantic parsing benchmarks (Scholar, Geo) with zero labeled data.

Shuyan Zhou, Pengcheng Yin, Graham Neubig

When humans conceive how to perform a particular task, they do so hierarchically: splitting higher-level tasks into smaller sub-tasks. However, in the literature on natural language (NL) command of situated agents, most works have treated the procedures to be executed as flat sequences of simple actions, or any hierarchies of procedures have been shallow at best. In this paper, we propose a formalism of procedures as programs, a powerful yet intuitive method of representing hierarchical procedural knowledge for agent command and control. We further propose a modeling paradigm of hierarchical modular networks, which consist of a planner and reactors that convert NL intents to predictions of executable programs and probe the environment for information necessary to complete the program execution. We instantiate this framework on the IQA and ALFRED datasets for NL instruction following. Our model outperforms reactive baselines by a large margin on both datasets. We also demonstrate that our framework is more data-efficient, and that it allows for fast iterative development.

Ziyu Yao, Frank F. Xu, Pengcheng Yin et al.

While most neural generative models generate outputs in a single pass, the human creative process is usually one of iterative building and refinement. Recent work has proposed models of editing processes, but these mostly focus on editing sequential data and/or only model a single editing pass. In this paper, we present a generic model for incremental editing of structured data (i.e., "structural edits"). Particularly, we focus on tree-structured data, taking abstract syntax trees of computer programs as our canonical example. Our editor learns to iteratively generate tree edits (e.g., deleting or adding a subtree) and applies them to the partially edited data, thereby the entire editing process can be formulated as consecutive, incremental tree transformations. To show the unique benefits of modeling tree edits directly, we further propose a novel edit encoder for learning to represent edits, as well as an imitation learning method that allows the editor to be more robust. We evaluate our proposed editor on two source code edit datasets, where results show that, with the proposed edit encoder, our editor significantly improves accuracy over previous approaches that generate the edited program directly in one pass. Finally, we demonstrate that training our editor to imitate experts and correct its mistakes dynamically can further improve its performance.

Pengcheng Yin, Graham Neubig, Wen-tau Yih et al.

Recent years have witnessed the burgeoning of pretrained language models (LMs) for text-based natural language (NL) understanding tasks. Such models are typically trained on free-form NL text, hence may not be suitable for tasks like semantic parsing over structured data, which require reasoning over both free-form NL questions and structured tabular data (e.g., database tables). In this paper we present TaBERT, a pretrained LM that jointly learns representations for NL sentences and (semi-)structured tables. TaBERT is trained on a large corpus of 26 million tables and their English contexts. In experiments, neural semantic parsers using TaBERT as feature representation layers achieve new best results on the challenging weakly-supervised semantic parsing benchmark WikiTableQuestions, while performing competitively on the text-to-SQL dataset Spider. Implementation of the model will be available at http://fburl.com/TaBERT .

Ansong Ni, Pengcheng Yin, Graham Neubig

A semantic parser maps natural language commands (NLs) from the users to executable meaning representations (MRs), which are later executed in certain environment to obtain user-desired results. The fully-supervised training of such parser requires NL/MR pairs, annotated by domain experts, which makes them expensive to collect. However, weakly-supervised semantic parsers are learnt only from pairs of NL and expected execution results, leaving the MRs latent. While weak supervision is cheaper to acquire, learning from this input poses difficulties. It demands that parsers search a large space with a very weak learning signal and it is hard to avoid spurious MRs that achieve the correct answer in the wrong way. These factors lead to a performance gap between parsers trained in weakly- and fully-supervised setting. To bridge this gap, we examine the intersection between weak supervision and active learning, which allows the learner to actively select examples and query for manual annotations as extra supervision to improve the model trained under weak supervision. We study different active learning heuristics for selecting examples to query, and various forms of extra supervision for such queries. We evaluate the effectiveness of our method on two different datasets. Experiments on the WikiSQL show that by annotating only 1.8% of examples, we improve over a state-of-the-art weakly-supervised baseline by 6.4%, achieving an accuracy of 79.0%, which is only 1.3% away from the model trained with full supervision. Experiments on WikiTableQuestions with human annotators show that our method can improve the performance with only 100 active queries, especially for weakly-supervised parsers learnt from a cold start.

Jeremy Lacomis, Pengcheng Yin, Edward J. Schwartz et al.

The decompiler is one of the most common tools for examining binaries without corresponding source code. It transforms binaries into high-level code, reversing the compilation process. Decompilers can reconstruct much of the information that is lost during the compilation process (e.g., structure and type information). Unfortunately, they do not reconstruct semantically meaningful variable names, which are known to increase code understandability. We propose the Decompiled Identifier Renaming Engine (DIRE), a novel probabilistic technique for variable name recovery that uses both lexical and structural information recovered by the decompiler. We also present a technique for generating corpora suitable for training and evaluating models of decompiled code renaming, which we use to create a corpus of 164,632 unique x86-64 binaries generated from C projects mined from GitHub. Our results show that on this corpus DIRE can predict variable names identical to the names in the original source code up to 74.3% of the time.

Pengcheng Yin, Graham Neubig, Miltiadis Allamanis et al.

We introduce the problem of learning distributed representations of edits. By combining a "neural editor" with an "edit encoder", our models learn to represent the salient information of an edit and can be used to apply edits to new inputs. We experiment on natural language and source code edit data. Our evaluation yields promising results that suggest that our neural network models learn to capture the structure and semantics of edits. We hope that this interesting task and data source will inspire other researchers to work further on this problem.

Pengcheng Yin, Graham Neubig

We present TRANX, a transition-based neural semantic parser that maps natural language (NL) utterances into formal meaning representations (MRs). TRANX uses a transition system based on the abstract syntax description language for the target MR, which gives it two major advantages: (1) it is highly accurate, using information from the syntax of the target MR to constrain the output space and model the information flow, and (2) it is highly generalizable, and can easily be applied to new types of MR by just writing a new abstract syntax description corresponding to the allowable structures in the MR. Experiments on four different semantic parsing and code generation tasks show that our system is generalizable, extensible, and effective, registering strong results compared to existing neural semantic parsers.

Shirley Anugrah Hayati, Raphael Olivier, Pravalika Avvaru et al.

In models to generate program source code from natural language, representing this code in a tree structure has been a common approach. However, existing methods often fail to generate complex code correctly due to a lack of ability to memorize large and complex structures. We introduce ReCode, a method based on subtree retrieval that makes it possible to explicitly reference existing code examples within a neural code generation model. First, we retrieve sentences that are similar to input sentences using a dynamic-programming-based sentence similarity scoring method. Next, we extract n-grams of action sequences that build the associated abstract syntax tree. Finally, we increase the probability of actions that cause the retrieved n-gram action subtree to be in the predicted code. We show that our approach improves the performance on two code generation tasks by up to +2.6 BLEU.

Xinyi Wang, Hieu Pham, Pengcheng Yin et al.

Recent advances in Neural Machine Translation (NMT) show that adding syntactic information to NMT systems can improve the quality of their translations. Most existing work utilizes some specific types of linguistically-inspired tree structures, like constituency and dependency parse trees. This is often done via a standard RNN decoder that operates on a linearized target tree structure. However, it is an open question of what specific linguistic formalism, if any, is the best structural representation for NMT. In this paper, we (1) propose an NMT model that can naturally generate the topology of an arbitrary tree structure on the target side, and (2) experiment with various target tree structures. Our experiments show the surprising result that our model delivers the best improvements with balanced binary trees constructed without any linguistic knowledge; this model outperforms standard seq2seq models by up to 2.1 BLEU points, and other methods for incorporating target-side syntax by up to 0.7 BLEU.

Pengcheng Yin, Chunting Zhou, Junxian He et al.

Semantic parsing is the task of transducing natural language (NL) utterances into formal meaning representations (MRs), commonly represented as tree structures. Annotating NL utterances with their corresponding MRs is expensive and time-consuming, and thus the limited availability of labeled data often becomes the bottleneck of data-driven, supervised models. We introduce StructVAE, a variational auto-encoding model for semisupervised semantic parsing, which learns both from limited amounts of parallel data, and readily-available unlabeled NL utterances. StructVAE models latent MRs not observed in the unlabeled data as tree-structured latent variables. Experiments on semantic parsing on the ATIS domain and Python code generation show that with extra unlabeled data, StructVAE outperforms strong supervised models.

Pengcheng Yin, Bowen Deng, Edgar Chen et al.

For tasks like code synthesis from natural language, code retrieval, and code summarization, data-driven models have shown great promise. However, creating these models require parallel data between natural language (NL) and code with fine-grained alignments. Stack Overflow (SO) is a promising source to create such a data set: the questions are diverse and most of them have corresponding answers with high-quality code snippets. However, existing heuristic methods (e.g., pairing the title of a post with the code in the accepted answer) are limited both in their coverage and the correctness of the NL-code pairs obtained. In this paper, we propose a novel method to mine high-quality aligned data from SO using two sets of features: hand-crafted features considering the structure of the extracted snippets, and correspondence features obtained by training a probabilistic model to capture the correlation between NL and code using neural networks. These features are fed into a classifier that determines the quality of mined NL-code pairs. Experiments using Python and Java as test beds show that the proposed method greatly expands coverage and accuracy over existing mining methods, even when using only a small number of labeled examples. Further, we find that reasonable results are achieved even when training the classifier on one language and testing on another, showing promise for scaling NL-code mining to a wide variety of programming languages beyond those for which we are able to annotate data.

Xuezhe Ma, Pengcheng Yin, Jingzhou Liu et al.

Reward augmented maximum likelihood (RAML), a simple and effective learning framework to directly optimize towards the reward function in structured prediction tasks, has led to a number of impressive empirical successes. RAML incorporates task-specific reward by performing maximum-likelihood updates on candidate outputs sampled according to an exponentiated payoff distribution, which gives higher probabilities to candidates that are close to the reference output. While RAML is notable for its simplicity, efficiency, and its impressive empirical successes, the theoretical properties of RAML, especially the behavior of the exponentiated payoff distribution, has not been examined thoroughly. In this work, we introduce softmax Q-distribution estimation, a novel theoretical interpretation of RAML, which reveals the relation between RAML and Bayesian decision theory. The softmax Q-distribution can be regarded as a smooth approximation of the Bayes decision boundary, and the Bayes decision rule is achieved by decoding with this Q-distribution. We further show that RAML is equivalent to approximately estimating the softmax Q-distribution, with the temperature $τ$ controlling approximation error. We perform two experiments, one on synthetic data of multi-class classification and one on real data of image captioning, to demonstrate the relationship between RAML and the proposed softmax Q-distribution estimation method, verifying our theoretical analysis. Additional experiments on three structured prediction tasks with rewards defined on sequential (named entity recognition), tree-based (dependency parsing) and irregular (machine translation) structures show notable improvements over maximum likelihood baselines.

Pengcheng Yin, Graham Neubig

We consider the problem of parsing natural language descriptions into source code written in a general-purpose programming language like Python. Existing data-driven methods treat this problem as a language generation task without considering the underlying syntax of the target programming language. Informed by previous work in semantic parsing, in this paper we propose a novel neural architecture powered by a grammar model to explicitly capture the target syntax as prior knowledge. Experiments find this an effective way to scale up to generation of complex programs from natural language descriptions, achieving state-of-the-art results that well outperform previous code generation and semantic parsing approaches.

Pengcheng Yin, Zhengdong Lu, Hang Li et al.

We proposed Neural Enquirer as a neural network architecture to execute a natural language (NL) query on a knowledge-base (KB) for answers. Basically, Neural Enquirer finds the distributed representation of a query and then executes it on knowledge-base tables to obtain the answer as one of the values in the tables. Unlike similar efforts in end-to-end training of semantic parsers, Neural Enquirer is fully "neuralized": it not only gives distributional representation of the query and the knowledge-base, but also realizes the execution of compositional queries as a series of differentiable operations, with intermediate results (consisting of annotations of the tables at different levels) saved on multiple layers of memory. Neural Enquirer can be trained with gradient descent, with which not only the parameters of the controlling components and semantic parsing component, but also the embeddings of the tables and query words can be learned from scratch. The training can be done in an end-to-end fashion, but it can take stronger guidance, e.g., the step-by-step supervision for complicated queries, and benefit from it. Neural Enquirer is one step towards building neural network systems which seek to understand language by executing it on real-world. Our experiments show that Neural Enquirer can learn to execute fairly complicated NL queries on tables with rich structures.