Ofir Press

Ofir Press is an AI researcher whose collaborative work includes SWE-bench, which evaluates models on real repository issues, SWE-agent, which gives agents an interface for working with code, and ProgramBench, which tests their ability to rebuild programs from their behavior and documentation. His contributions also include language-model embedding methods and self-ask prompting.[4][5][9][7][6]

Press coauthored SWE-bench with Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei and Karthik Narasimhan. First submitted in October 2023, the paper introduced a benchmark built from real GitHub issues and their corresponding pull requests. Its original set contained 2,294 problems across 12 Python repositories: a model receives a codebase and an issue description, then must edit the codebase to address the issue.[4]

In May 2024, Press and collaborators introduced SWE-agent, investigating how interface design affects language-model agents' ability to perform software engineering tasks. Its custom agent-computer interface supports creating and editing files, navigating repositories, and running tests and other programs. The coauthor team comprised John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan and Press.[5]

Press's later work includes coauthoring ProgramBench, first submitted on May 5, 2026. Here, agents receive a program and its documentation and must architect and implement a codebase that reproduces the reference executable's behavior. Agent-driven fuzzing generates end-to-end behavioral tests, allowing evaluation without prescribing how the implementation should be structured.[9]

During this period of software-agent research, Press worked as a postdoctoral researcher with Karthik Narasimhan's group at Princeton. The lab's alumni record lists his appointment as 2023–25.[2][3][4][5]

Before his software-agent work, Press studied language-model representations. In April 2017, he and Lior Wolf published Using the Output Embedding to Improve Language Models. They treated the model's output weight matrix as a word embedding and recommended tying the input and output embeddings, sharing their weights. They reported reduced perplexity across the language models studied and showed that weight tying could reduce the size of their neural translation models to less than half without harming performance.[7]

With Noah A. Smith and Mike Lewis, Press also introduced Attention with Linear Biases, or ALiBi, a position method for language models.[8]

Press's work on prompting includes Measuring and Narrowing the Compositionality Gap in Language Models, published in December 2023 with Muru Zhang, Sewon Min, Ludwig Schmidt, Noah Smith and Mike Lewis. Their self-ask method makes a model explicitly ask and answer follow-up questions before answering the initial question. That structure also permits a search engine to supply answers to the intermediate questions; the researchers reported that this improved accuracy in their experiments.[6]

Press also demonstrated a natural-language plotting interface, according to Latent Space's account. The interface let a user generate a chart through matplotlib and iteratively describe changes in natural language.[1]