arXiv subjects

Rishi Desai

Publications and source records attributed to Rishi Desai.

6 recordsLinked to original sources

arXiv2026

SWE-Marathon: Can Agents Autonomously Complete Ultra-Long-Horizon Software Work?

AI agents are increasingly expected to complete long-horizon workflows that require sustained progress over hours, millions of tokens, and complex environments. Yet current agent benchmarks largely evaluate short-form tasks, such as single pull requests, small tickets, or 5-10 minute exercises, limiting our ability to measure agents' capabilities in planning, long-context understanding, and memory use. We introduce SWE-Marathon, a benchmark of 20 long-horizon tasks spanning software engineering and adjacent technical domains. Each task consists of a unique executable environment, a human-written reference solution, and a multi-layer verification suite. Logged agent attempts average 27.2M total tokens, making SWE-Marathon substantially longer-horizon than existing SWE and command-line agent benchmarks. Current frontier coding agents solve fewer than 30% of tasks. Failures often arise from poor self-verification, self-reported infeasibility, and premature termination. We also observe reward-hacking behavior in 13.8% of rollouts, where agents attempt to exploit the environment or verifier to bypass the intended workflow. SWE-Marathon includes adversarial review of test suites and execution environments, as well as multi-layer checks designed to prevent shortcut solutions. We release SWE-Marathon, evaluation code, and agent trajectories at https://swe-marathon.org/.

cs.SE↗

arXiv2025

Assessing treatment effects in observational data with missing confounders: A comparative study of practical doubly-robust and traditional missing data methods

In pharmacoepidemiology, safety and effectiveness are frequently evaluated using readily available administrative and electronic health records data. In these settings, detailed confounder data are often not available in all data sources and therefore missing on a subset of individuals. Multiple imputation (MI) and inverse-probability weighting (IPW) are go-to analytical methods to handle missing data and are dominant in the biomedical literature. Doubly-robust methods, which are consistent under fewer assumptions, can be more efficient with respect to mean-squared error. We discuss two practical-to-implement doubly-robust estimators, generalized raking and inverse probability-weighted targeted maximum likelihood estimation (TMLE), which are both currently under-utilized in biomedical studies. We compare their performance to IPW and MI in a detailed numerical study for a variety of synthetic data-generating and missingness scenarios, including scenarios with rare outcomes and a high missingness proportion. Further, we consider plasmode simulation studies that emulate the complex data structure of a large electronic health records cohort in order to compare anti-depressant therapies in a rare-outcome setting where a key confounder is prone to more than 50\% missingness. We provide guidance on selecting a missing data analysis approach, based on which methods excelled with respect to the bias-variance trade-off across the different scenarios studied.

stat.ME↗

arXiv2025

A cautionary note for plasmode simulation studies in the setting of causal inference

Plasmode simulation has become an important tool for evaluating the operating characteristics of different statistical methods in complex settings, such as pharmacoepidemiological studies of treatment effectiveness using electronic health records (EHR) data. These studies provide insight into how estimator performance is impacted by challenges including rare events, small sample size, etc., that can indicate which among a set of methods performs best in a real-world dataset. Plasmode simulation combines data resampled from a real-world dataset with synthetic data to generate a known truth for an estimand in realistic data. There are different potential plasmode strategies currently in use. We compare two popular plasmode simulation frameworks. We provide numerical evidence and a theoretical result, which shows that one of these frameworks can cause certain estimators to incorrectly appear overly biased with lower than nominal confidence interval coverage. Detailed simulation studies using both synthetic and real-world EHR data demonstrate that these pitfalls remain at large sample sizes and when analyzing data from a randomized controlled trial. We conclude with guidance for the choice of a plasmode simulation approach that maintains good theoretical properties to allow a fair evaluation of statistical methods while also maintaining the desired similarity to real data.

stat.ME↗

arXiv2024

Privacy-Constrained Policies via Mutual Information Regularized Policy Gradients

As reinforcement learning techniques are increasingly applied to real-world decision problems, attention has turned to how these algorithms use potentially sensitive information. We consider the task of training a policy that maximizes reward while minimizing disclosure of certain sensitive state variables through the actions. We give examples of how this setting covers real-world problems in privacy for sequential decision-making. We solve this problem in the policy gradients framework by introducing a regularizer based on the mutual information (MI) between the sensitive state and the actions. We develop a model-based stochastic gradient estimator for optimization of privacy-constrained policies. We also discuss an alternative MI regularizer that serves as an upper bound to our main MI regularizer and can be optimized in a model-free setting, and a powerful direct estimator that can be used in an environment with differentiable dynamics. We contrast previous work in differentially-private RL to our mutual-information formulation of information disclosure. Experimental results show that our training method results in policies that hide the sensitive state, even in challenging high-dimensional tasks.

cs.LG↗

arXiv2021

Home Action Genome: Cooperative Compositional Action Understanding

Existing research on action recognition treats activities as monolithic events occurring in videos. Recently, the benefits of formulating actions as a combination of atomic-actions have shown promise in improving action understanding with the emergence of datasets containing such annotations, allowing us to learn representations capturing this information. However, there remains a lack of studies that extend action composition and leverage multiple viewpoints and multiple modalities of data for representation learning. To promote research in this direction, we introduce Home Action Genome (HOMAGE): a multi-view action dataset with multiple modalities and view-points supplemented with hierarchical activity and atomic action labels together with dense scene composition labels. Leveraging rich multi-modal and multi-view settings, we propose Cooperative Compositional Action Understanding (CCAU), a cooperative learning framework for hierarchical action recognition that is aware of compositional action elements. CCAU shows consistent performance improvements across all modalities. Furthermore, we demonstrate the utility of co-learning compositions in few-shot action recognition by achieving 28.6% mAP with just a single sample.

cs.CV↗

arXiv2021

Exploring Extended Reality with ILLIXR: A New Playground for Architecture Research

As we enter the era of domain-specific architectures, systems researchers must understand the requirements of emerging application domains. Augmented and virtual reality (AR/VR) or extended reality (XR) is one such important domain. This paper presents ILLIXR, the first open source end-to-end XR system (1) with state-of-the-art components, (2) integrated with a modular and extensible multithreaded runtime, (3) providing an OpenXR compliant interface to XR applications (e.g., game engines), and (4) with the ability to report (and trade off) several quality of experience (QoE) metrics. We analyze performance, power, and QoE metrics for the complete ILLIXR system and for its individual components. Our analysis reveals several properties with implications for architecture and systems research. These include demanding performance, power, and QoE requirements, a large diversity of critical tasks, inter-dependent execution pipelines with challenges in scheduling and resource management, and a large tradeoff space between performance/power and human perception related QoE metrics. ILLIXR and our analysis have the potential to propel new directions in architecture and systems research in general, and impact XR in particular. ILLIXR is open-source and available at https://illixr.github.io

cs.DC↗