AI-powered web agents have the potential to automate repetitive tasks, such as form filling, information retrieval, and scheduling, but they struggle to reliably execute these tasks without human intervention, requiring users to provide detailed guidance during every run. We address this limitation by automatically synthesizing reusable workflows from an agent’s successful and failed attempts. These workflows incorporate execution guards that help agents detect and fix errors while keeping users informed of progress and issues. Our approach enables agents to successfully complete repetitive tasks of the same type with minimal user intervention, increasing the success rates from 24.2% to 70.1% across fifteen tasks. To evaluate this approach, we invited nine users and found that our agent helped them complete web tasks with a higher success rate and less guidance compared to two baseline methods, as well as allowed users to easily monitor agent behavior and understand its failures.

Users often rely on GUIs to edit and interact with visualizations — a daunting task due to the large space of editing options. As a result, users are either overwhelmed by a complex UI or constrained by a custom UI with a tailored, fixed subset of options with limited editing flexibility. Natural Language Interfaces (NLIs) are emerging as a feasible alternative for users to specify edits. However, NLIs forgo the advantages of traditional GUI: the ability to explore and repeat edits and see instant visual feedback.

Large language models have shown remarkable aptitude in code generation, but still struggle to perform complex tasks. Self-repair---in which the model debugs and repairs its own code---has recently become a popular way to boost performance in these settings. However, despite its increasing popularity, existing studies of self-repair have been limited in scope; in many settings, its efficacy thus remains poorly understood. In this paper, we analyze Code Llama, GPT-3.5 and GPT-4's ability to perform self-repair on problems taken from HumanEval and APPS. We find that when the cost of carrying out repair is taken into account, performance gains are often modest, vary a lot between subsets of the data, and are sometimes not present at all. We hypothesize that this is because self-repair is bottlenecked by the model's ability to provide feedback on its own code; using a stronger model to artificially boost the quality of the feedback, we observe substantially larger performance gains. Similarly, a small-scale study in which we provide GPT-4 with feedback from human participants suggests that even for the strongest models, self-repair still lags far behind what can be achieved with human-level debugging.

Data wrangling, the process of preparing raw data for further analysis in computational notebooks, is a crucial yet time-consuming step in data science. Code generation has the potential to automate the data wrangling process to reduce analysts' overhead by translating user intents into executable code. Precisely generating data wrangling code necessitates a comprehensive consideration of the rich context present in notebooks, including textual context, code context and data context. However, notebooks often interleave multiple non-linear analysis tasks into linear sequence of code blocks, where the contextual dependencies are not clearly reflected. Directly training models with source code blocks fails to fully exploit the contexts for accurate wrangling code generation.

Search-based software testing (SBST) generates high-coverage test cases for programs under test with a combination of test case generation and mutation. SBST's performance relies on there being a reasonable probability of generating test cases that exercise the core logic of the program under test. Given such test cases, SBST can then explore the space around them to exercise various parts of the program. This paper explores whether Large Language Models (LLMs) of code, such as OpenAI's Codex, can be used to help SBST's exploration. Our proposed algorithm, CodaMosa, conducts SBST until its coverage improvements stall, then asks Codex to provide example test cases for under-covered functions. These examples help SBST redirect its search to more useful areas of the search space. On an evaluation over 486 benchmarks, CodaMosa achieves statistically significantly higher coverage on many more benchmarks (173 and 279) than it reduces coverage on (10 and 4), compared to SBST and LLM-only baselines.

Large language models (LLMs) have demonstrated an impressive ability to generate code for various programming tasks. In many instances, LLMs can generate a correct program for a task when given numerous trials. Consequently, a recent trend is to do large scale sampling of programs using a model and then filtering/ranking the programs based on the program execution on a small number of known unit tests to select one candidate solution. However, these approaches assume that the unit tests are given and assume the ability to safely execute the generated programs (which can do arbitrary dangerous operations such as file manipulations). Both of the above assumptions are impractical in real-world software development. In this paper, we propose CodeRanker, a neural ranker that can predict the correctness of a sampled program without executing it. Our CodeRanker is fault-aware i.e., it is trained to predict different kinds of execution information such as predicting the exact compile/runtime error type (e.g., an IndexError or a TypeError). We show that CodeRanker can significantly increase the pass@1 accuracy of various code generation models (including Codex, GPT-Neo, GPT-J) on APPS, HumanEval and MBPP datasets.

Forecasting complex vehicle and pedestrian multi-modal distributions requires powerful probabilistic approaches. Normalizing flows (NF) have recently emerged as an attractive tool to model such distributions. However, a key drawback is that independent samples drawn from a flow model often do not adequately capture all the modes in the underlying distribution. We propose Likelihood-Based Diverse Sampling (LDS), a method for improving the quality and the diversity of trajectory samples from a pre-trained flow model. Rather than producing individual samples, LDS produces a set of trajectories in one shot. Given a pre-trained forecasting flow model, we train LDS using gradients from the model, to optimize an objective function that rewards high likelihood for individual trajectories in the predicted set, together with high spatial separation among trajectories. LDS outperforms state-of-art post-hoc neural diverse forecasting methods for various pre-trained flow models as well as conditional variational autoencoder (CVAE) models. Crucially, it can also be used for transductive trajectory forecasting, where the diverse forecasts are trained on-the-fly on unlabeled test examples. LDS is easy to implement, and we show that it offers a simple plug-in improvement over baselines on two challenging benchmarks. Code is at: https://github.com/JasonMa2016/LDS

A key challenge for reinforcement learning is solving long-horizon planning problems. Recent work has leveraged programs to guide reinforcement learning in these settings. However, these approaches impose a high manual burden on the user since they must provide a guiding program for every new task. Partially observed environments further complicate the programming task because the program must implement a strategy that correctly, and ideally optimally, handles every possible configuration of the hidden regions of the environment. We propose a new approach, model predictive program synthesis (MPPS), that uses program synthesis to automatically generate the guiding programs. It trains a generative model to predict the unobserved portions of the world, and then synthesizes a program based on samples from this model in a way that is robust to its uncertainty. In our experiments, we show that our approach significantly outperforms non-program-guided approaches on a set of challenging benchmarks, including a 2D Minecraft-inspired environment where the agent must complete a complex sequence of subtasks to achieve its goal, and achieves a similar performance as using handcrafted programs to guide the agent. Our results demonstrate that our approach can obtain the benefits of program-guided reinforcement learning without requiring the user to provide a new guiding program for every new task.

We study the problem of inferring communication structures that can solve cooperative multi-agent planning problems while minimizing the amount of communication. We quantify the amount of communication as the maximum degree of the communication graph; this metric captures settings where agents have limited bandwidth. Minimizing communication is challenging due to the combinatorial nature of both the decision space and the objective; for instance, we cannot solve this problem by training neural networks using gradient descent. We propose a novel algorithm that synthesizes a control policy that combines a programmatic communication policy used to generate the communication graph with a transformer policy network used to choose actions. Our algorithm first trains the transformer policy, which implicitly generates a "soft" communication graph; then, it synthesizes a programmatic communication policy that "hardens" this graph, forming a neurosymbolic transformer. Our experiments demonstrate how our approach can synthesize policies that generate low-degree communication graphs while maintaining near-optimal performance.

The great promise of modular robots is the ability to create on demand robots; however, choosing the “right” design based on a task is still a challenging problem. In this paper, we present an approach to automatically synthesize both the design and control for modular robots from a task description. In particular, we focus on manipulators composed of one degree-of-freedom (DoF) modules. Our approach is able to handle partially infeasible tasks by either identifying the infeasible part and finding a design that satisfies the feasible part or searching for multiple designs that together satisfy the entire task. We compare our approach to a baseline genetic algorithm in a series of increasingly complex environments.

Data integration between web sources and relational data is a key challenge faced by data scientists and spreadsheet users. There are two main challenges in programmatically joining web data with relational data. First, most websites do not expose a direct interface to obtain tabular data, so the user needs to formulate a logic to get to different webpages for each input row in the relational table. Second, after reaching the desired webpage, the user needs to write complex scripts to extract the relevant data, which is often conditioned on the input data. Since many data scientists and end-users come from diverse backgrounds, writing such complex regular-expression based logical scripts to perform data integration tasks is unfortunately often beyond their programming expertise. We present WebRelate, a system that allows users to join semi-structured web data with relational data in spreadsheets using input-output examples. WebRelate decomposes the web data integration task into two sub-tasks of i) URL learning and ii) input-dependent web extraction. We introduce a novel synthesis paradigm called "Output-constrained Programming By Examples", which allows us to use the finite set of possible outputs for the new inputs to efficiently constrain the search in the synthesis algorithm. We instantiate this paradigm for the two sub-tasks in WebRelate. The first sub-task generates the URLs for the webpages containing the desired data for all rows in the relational table. WebRelate achieves this by learning a string transformation program using a few example URLs. The second sub-task uses examples of desired data to be extracted from the corresponding webpages and learns a program to extract the data for the other rows. We design expressive domain-specific languages for URL generation and web data extraction, and present efficient synthesis algorithms for learning programs in these DSLs from few input-output examples. We evaluate WebRelate on 88 real-world web data integration tasks taken from online help forums and Excel product team, and show that WebRelate can learn the desired programs within few seconds using only 1 example for the majority of the tasks.

While computer-aided design is a major part of many modern manufacturing pipelines, the design files typically generated describe raw geometry. Lost in this representation is the procedure by which these designs were generated. In this paper, we present a method for reverse-engineering the process by which 3D models may have been generated, in the language of constructive solid geometry (CSG). Observing that CSG is a formal grammar, we formulate this inverse CSG problem as a program synthesis problem. Our solution is an algorithm that couples geometric processing with state-of-the-art program synthesis techniques. In this scheme, geometric processing is used to convert the mixed discrete and continuous domain of CSG trees to a pure discrete domain where modern program synthesizers excel. We demonstrate the efficiency and scalability of our algorithm on several different examples, including those with over 100 primitive parts. We show that our algorithm is able to find simple programs which are close to the ground truth, and demonstrate our method's applicability in mesh re-editing. Finally, we compare our method to prior state-of-the-art. We demonstrate that our algorithm dominates previous methods in terms of resulting CSG compactness and runtime, and can handle far more complex input meshes than any previous method.

It is often possible to improve a concurrent system's performance by leveraging the semantics of its datatypes. We build a new software transactional memory (STM) around this observation. A conventional STM tracks read- and write-sets of memory words; even simple operations can generate large sets. Our STM, which we call STO, tracks abstract operations on transactional datatypes instead. Parts of the transactional commit protocol are delegated to these datatypes' implementations, which can use datatype semantics, and new commit protocol features, to reduce bookkeeping, limit false conflicts, and implement efficient concurrency control. We test these ideas on the STAMP benchmark suite for STM applications and on our own prior work, the Silo high-performance in-memory database, observing large performance improvements in both systems.