I’m happy to share our work “AlphaTrans: A Neuro-Symbolic Compositional Approach for Repository-Level Code Translation and Validation” is accepted to.@FSEconf 2025. Joint work b/w @IllinoisCS & @IBMResearch (🧵). 📄 Paper: 📷 Code:

I would like to thank all wonderful authors for their contributions:. Kaiyao Ke, @mrigankpawagi, Salman Abid, @rangeetpan, Saurabh Sinha and @Reyhaneh. Also big shoutout to Raju Pavuluri and Darko Marinov. Looking forward to seeing everyone this June in Trondheim, Norway 🇳🇴.

Our artifacts, including generated partial Python translations, manually verified translations, and automation scripts for reproducing AlphaTrans results, are publicly available. 💿 Code Repository:

Finally, we show AlphaTrans is model-agnostic. A stronger model (GPT-4o) improves the translation quality — functional equivalence increases by 2.81%. We also observed a huge overlap between successful translations and the unique benefits each LLM provides in code translation.

Moreover, we augment existing test suites with EvoSuite. Test augmentation can further validate the correctness of 2.11% of fragments not executed by developer tests. The generated tests are more focused and, on average, invoke 48% fewer methods than the developer-written tests.

AlphaTrans decomposes source unit tests to address the "test translation coupling effect" problem. Test decomposition unburdens validation of fragments from incorrect translations. 62.41% of test fragments for unit tests that would have been marked as failed achieves a test pass.

Although AlphaTrans cannot validate all translations, it provides partial translations and artifacts that developers can use to complete translations. On average, it takes approximately 20.1 hours for developers to achieve green tests, which could otherwise take weeks.

AlphaTrans combines the strengths of static analysis and LLMs for translation. Our study reveals AlphaTrans is effective, achieving 96.40% and 25.14% syntactic and functional correctness when evaluated using an open-source LLM (e.g., DeepSeek-Coder).

In this paper, we propose AlphaTrans, a neuro-symbolic compositional approach for translating and validating repository-level code from Java to Python. AlphaTrans is the first end-to-end framework and methodology that enables translating repository-level code at scale.

life is hard.

RT @RosuGrigore: @lorisdanto The entire publishing process requires a major overhaul in my view. Currently we spend months polishing the p….

RT @Reyhaneh: Introducing Alphatrans, a neuro-symbolic approach for translation and validation of "whole repository" of the large-scale rea….

academic people are too ambitious about open-source llms.

working with a lot of people sucks.

RT @10ardaguler: Türk’e durmak yaraşmaz. ❤️🤍🇹🇷

RT @gdb: Introducing GPT-4o, our new model which can reason across text, audio, and video in real time. It's extremely versatile, fun to p….

Special thanks to @MultiMichele, @Reyhaneh, and all other IBM collaborators for making this leaderboard happen.

Joint work b/w @IllinoisCS and @IBMResearch. Paper: Code: Video: Slides:

Please feel free to open an issue on GitHub if you want a specific LLM to be evaluated on our code translation benchmark:.

Interested in LLM-based Code Translation 🧐?. Check our CodeLingua leaderboard (. We have updated the leaderboard with newly released Granite code LLMs from @IBMResearch. Granite models outperform Claude-3 and GPT-4 in C -> C++ translation 🔥.