To mitigate this drop, we propose and evaluate the use of contrastive learning, which naturally ranks similarity between objects, thus enabling more effective semantic clone detection in-the-wild.

🙏 Huge thanks to my amazing co-authors, Francesco Sovrano, Earl T. Barr, and @sback.it. [4/4]

By evaluating six models on clones of unseen functionality, we observe a significant performance drop for models explicitly trained for clone detection. For general-purpose LLMs, the drop is lower, but still exists. [3/4]

SOTA clone detection models are trained on clones of specific functionalities and tested on different clones of the same functionalities. But in practice, developers need to identify clones of functionalities the models have not been trained on. How well do models perform in such scenarios? [2/4]

🙏 Many thanks to Marco Castelluccio for the great collaboration, and to @sback.it for his invaluable mentorship during this work.

📄 Preprint: arxiv.org/abs/2509.01616
💻 Code: github.com/kitsiosk/blast

#ASE2025

5/5

We deployed BLAST in three open-source repositories from @mozilla.org, where it proposed 11 fail-to-pass tests to the developers, 6 of which were confirmed to reproduce the designated issue. This calls for scrutiny towards the widely used fail-to-pass metric, which we discuss in detail. 4/5

BLAST generates such fail-to-pass tests in 151 out of 426 (35.4%) issue-patch pairs from a widely used benchmark, outperforming state-of-the-art approaches while requiring only 2 LLM queries and 1 minute of lightweight SBST generation. 3/5

We introduce BLAST, a tool that combines LLMs and Search-Based Software Testing (SBST) to generate tests that fail before a patch and pass after. 2/5

Already mentioned by someone else here, but is there any practical way a dev could support 🇺🇦?

Congrats!

Wow, these are indeed top-notch researchers in the field, totally worth reading some of their work.

Here is also a VM from scratch in Python instead of C, by Greg Wilson: third-bit.com/sdxpy/vm/