AI routinely fixes software bugs today. Flashback to 2017. In 2017, the Repairnator expedition aimed to be the first to have a real world bug fixed by a machine, and merged by a human. Paper: "Repairnator patches programs automatically" (ACM Ubiquity 2019)

Excited to join UCL UCL Computer Science as a Lecturer (aka Assistant Professor) starting in Spring 2025. I'm looking forward to collaborating with amazing colleagues Federica Sarro 🏳️‍🌈, and looking for 25Fall PhD students to work on multimodal agentic coding. If you're interested, drop me an email!

We invite you to the PhD defense of Khashayar Etemadi on "Efficient Exploration and Analysis of Program Repair Search Spaces". When: Friday 22 November 2024, 10h CET Where: KTH, Room F3 Zoom: kth-se.zoom.us/j/64148823838

Do Automated Fixes Truly Mitigate Smart Contract Exploits? - 👍 the answer is "yes", for some vulnerability types (Reentrancy, Arithmetic, Bad Randomness, Unchecked Low-Level Calls and Access Control) - 🧂 this is modulo dataset Full paper: arxiv.org/pdf/2501.04600

Software diversity is an old idea in reliability. What if we blend it with cryptoeconomic thinking? Economic incentives would nudge diversity which would then result in the highest systemic reliability. Paper: "Proving and Rewarding Client Diversity" arxiv.org/pdf/2411.18401

Chinese AI 🇨🇳 takes 1st place on RepairBench. RepairBench is a benchmark of LLMs for program repair. DeepSeek-R1 is the latest reasoning model from DeepSeek AI (and open-source!). repairbench.github.io

OpenAI strikes back and reclaims first place 🥇 on the RepairBench leaderboard for automated bug fixing. repairbench.github.io

Large language models (LLMs) are great at repairing programs, but they can be pricey! CigaR has a single goal: generating good patches while using fewer tokens and fewer $$$. Paper: CigaR: Cost-efficient Program Repair with LLMs arxiv.org/pdf/2402.06598

The DISL dataset contains one of the largest ever collection of 514,506 unique Solidity files. It is a useful resource for researchers, developers, and enthusiasts in the realm of Ethereum smart contracts. paper: arxiv.org/pdf/2403.16861

Developers don't like to read XML. Introducing "plain code serialization" with ProDJ. Simple idea: turn runtime Java objects into plain, readable code! 🖥️💡 Serializing Java Objects in Plain Code arxiv.org/pdf/2405.11294

Just now The program equivalence problem is one of the hardest in CS. We explore the use of LLMs for this. Mokav is able to successfully identifying functional differences in 81.7% of program pairs! Mokav: Execution-driven Differential Testing with LLMs arxiv.org/pdf/2406.10375

SIGBOVIK. Don't miss the extendeded submission deadline. Call for papers SIGBOVIK 2025: lists.andrew.cmu.edu/pipermail/sigb…

The UZH Blockchain Center is pleased to invite you to our upcoming lecture: “Software Supply Chain Attacks and Defenses for Web3” by Prof. Dr Martin Monperrus Martin Monperrus 🗓 Date: 27 March 2025 ⏰ Time: 18:15–20:00 CET

AI research overlooks the most pressing issues. Our team at KTH finally found a solution to the cat image generation problem: UPPERCASE IS ALL YOU NEED Read the full SIGBOVIK paper: monperrus.net/martin/SIGBOVI…

[software supply chain security] My new GPG key for signing commits and releases is 9D91C939710507D65F2ADC785BE694BB2D59D0B0 monperrus.net/martin/protoco…

Onchain truth about smart contracts on Ethereum: 🔍 11 deployers created 20.5 million contracts Our recent research papers uncovers the hidden complexities of smart contract dependencies. On-Chain Analysis of Smart Contract Dependency Risks on Ethereum arxiv.org/pdf/2503.19548

🔒 Reproducible builds alert. Our new study tackles the tricky problem of unreproducible builds in Java, a common headache for developers. Canonicalization for Unreproducible Builds in Java arxiv.org/pdf/2504.21679

If you supervise PhD students, there is a magic skill. Humor for graduate training In ACM Inroads, 2025. publisher: dl.acm.org/doi/10.1145/37… open-access: kth.diva-portal.org/smash/get/diva… by Benoit Baudry Martin Monperrus

AI fixes bugs by thousands every day with next token prediction. What if we could optimize programs towards correctness, instead of blindly output tokens one at a time? See "Gradient-Based Program Repair: Fixing Bugs in Continuous Program Spaces" Arxiv: arxiv.org/pdf/2505.17703

Missed my talk on generative AI for synthetic test data at SAST'30? Catchup with the metapaper "Generative AI to Generate Test Data Generators". Generative AI to Generate Test Data Generators IEEE Software, 2024 Open-access: doi.org/10.1109/MS.202…