Jonathan Bell

Jaden Fiotto-Kaufman, Alexander R. Loftus, Eric Todd et al.

We introduce NNsight and NDIF, technologies that work in tandem to enable scientific study of the representations and computations learned by very large neural networks. NNsight is an open-source system that extends PyTorch to introduce deferred remote execution. The National Deep Inference Fabric (NDIF) is a scalable inference service that executes NNsight requests, allowing users to share GPU resources and pretrained models. These technologies are enabled by the Intervention Graph, an architecture developed to decouple experimental design from model runtime. Together, this framework provides transparent and efficient access to the internals of deep neural networks such as very large language models (LLMs) without imposing the cost or complexity of hosting customized models individually. We conduct a quantitative survey of the machine learning literature that reveals a growing gap in the study of the internals of large-scale AI. We demonstrate the design and use of our framework to address this gap by enabling a range of research methods on huge models. Finally, we conduct benchmarks to compare performance with previous approaches. Code, documentation, and tutorials are available at https://nnsight.net/.

Yangtian Zi, Zixuan Wu, Aleksander Boruch-Gruszecki et al.

Fine-tuning large language models for code editing has typically relied on mining commits and pull requests. The working hypothesis has been that commit messages describe human intent in natural language, and patches to code describe the changes that implement that intent. However, much of the previously collected data is noisy: commit messages are terse, human-written commits commingle several unrelated edits, and many commits come from simple, rule-based bots. The recent adoption of software engineering agents changes this landscape. Code changes co-authored by humans and agents tend to be more narrowly scoped and focused on clearer goals. Their commit messages, generated by LLMs, articulate intent and rationale in much greater detail. Moreover, when these changes land in public repositories, they are implicitly filtered by humans: maintainers discard low-quality commits to their projects. We present AgentPack, a corpus of 1.3M code edits co-authored by Claude Code, OpenAI Codex, and Cursor Agent across public GitHub projects up to mid-August 2025. We describe the identification and curation pipeline, quantify adoption trends of these agents, and analyze the structural properties of the edits. Finally, we show that models fine-tuned on AgentPack can outperform models trained on prior human-only commit corpora, highlighting the potential of using public data from software engineering agents to train future code-editing models.

Jonathan Bell, Hannah M. Dee

We present an approach to leaf level segmentation of images of Arabidopsis thaliana plants based upon detected edges. We introduce a novel approach to edge classification, which forms an important part of a method to both count the leaves and establish the leaf area of a growing plant from images obtained in a high-throughput phenotyping system. Our technique uses a relatively shallow convolutional neural network to classify image edges as background, plant edge, leaf-on-leaf edge or internal leaf noise. The edges themselves were found using the Canny edge detector and the classified edges can be used with simple image processing techniques to generate a region-based segmentation in which the leaves are distinct. This approach is strong at distinguishing occluding pairs of leaves where one leaf is largely hidden, a situation which has proved troublesome for plant image analysis systems in the past. In addition, we introduce the publicly available plant image dataset that was used for this work.

Fang-Hsiang Su, Jonathan Bell, Gail Kaiser et al.

Android applications are usually obfuscated before release, making it difficult to analyze them for malware presence or intellectual property violations. Obfuscators might hide the true intent of code by renaming variables and/or modifying program structures. It is challenging to search for executables relevant to an obfuscated application for developers to analyze efficiently. Prior approaches toward obfuscation resilient search have relied on certain structural parts of apps remaining as landmarks, un-touched by obfuscation. For instance, some prior approaches have assumed that the structural relationships between identifiers are not broken by obfuscators; others have assumed that control flow graphs maintain their structures. Both approaches can be easily defeated by a motivated obfuscator. We present a new approach,Macneto, to search for programs relevant to obfuscated executables leveraging deep learning and principal components on instructions. Macneto makes few assumptions about the kinds of modifications that an obfuscator might perform. We show that it has high search precision for executables obfuscated by a state-of-the-art obfuscator that changes control flow. Further, we also demonstrate the potential of Macneto to help developers understand executables, where Macneto infers keywords (which are from the relevant unobfuscated program) for obfuscated executables.

Jonathan Bell, Hannah M. Dee

We describe a new measure for the evaluation of region level segmentation of objects, as applied to evaluating the accuracy of leaf-level segmentation of plant images. The proposed approach enforces the rule that a region (e.g. a leaf) in either the image being evaluated or the ground truth image evaluated against can be mapped to no more than one region in the other image. We call this measure the subset-matched Jaccard index.