🚀 Introducing Structured LLM, a new framework for making large language models more aligned, useful, and efficient.

👉 Check it out here: structuredllm.com

[5/n] 💡 Why it matters
Structured LLM enables outputs that are both syntactically valid and semantically sound across multiple domains, including programming, math reasoning, and trustworthy AI.

It opens the door to reliable and trustworthy LLM applications.

[6/n]
Contributors: Sasa Misailovic, Gagandeep Singh, Shubham Ugare, Debangshu Banerjee, Tarun Suresh, Rohan Gumaste, Hangoo Kang, Preston Firestone, Madhav Kanda, Adharsh Kamath

#AI #LLM #StructuredOutput #NLP #MachineLearning #CodeGeneration #TrustworthyAI

1/4
𝐍𝐞𝐰 𝐏𝐚𝐩𝐞𝐫 accepted @ 𝐈𝐂𝐌𝐋 𝟐𝟎𝟐𝟓!

@debangshuban18 , @TarunSures41845 , @shubhamdugare, Sasa Misailovic, @ggn_dp_sngh

introduce CRANE — a decoding algorithm that combines LLM reasoning with formal guarantees.
📄 arxiv.org/abs/2502.09061

3/4
The paper theoretically proves that constrained decoding can still preserve performance, even when the grammar is augmented with Chain-of-Thought steps.

This bridges free-form reasoning with formal output guarantees.

4/4
On symbolic reasoning benchmarks like GSM-symbolic and FOLIO, CRANE achieves:
📈 Up to +10% accuracy over prior constrained decoding methods

#ICML2025 #LLM #ConstrainedDecoding #FormalMethods

Looks like my AMS Early Career Notice on proof assistants is published! Check it out here: ams.org/journals/notic…

#ILLINOIS CS professor Darko Marinov & former students Qingzhou Luo, Farah Hariri & Lamyaa Eloussi were awarded 2024 Test of Time Honorable Mention by the ACM ICF of Software Engineering for their 2014 paper "An empirical analysis of flaky tests." 🔸🔹
siebelschool.illinois.edu/news/marinov-f…

Yay, our paper on the abilities of transformer models to learn to emulate structural recursion was accepted to TMLR, a wonderful machine learning research journal! With Dylan Zhang, Curt Tigges, Zory Zhang, Stella Biderman, and Maxim Raginsky. Will update the arXiv version soon!

Our paper "Correctly Compiling Proofs About Programs Without Proving Compilers Correct" by Audrey Seo*, (@audreyleeseo), Christopher Lam* (@gracelessfaller), Dan Grossman (@djg98115), and Talia Ringer will appear in ITP 2024! Preprint: dependenttyp.es/pdf/potpie.pdf

* co-first authors

Congratulations to Qingzhou Luo, Farah Hariri, Lamyaa Eloussi, and Darko Marinov for the FSE 2024 Test of Time Award Runner-Up Honorable Mention from FSE 2014! @plfmse @illinoiscs @FSEconf

Our alumni are awesome! Congratulations, @adithyamuralism. What a big win for Wisconsin 🎉

@plfmse students rock 🎉

@yinlin_deng is presenting her FuzzGPT work at #icse24. FuzzGPT aims to guide LLMs (which were trained to generate more natural and common info) to generate exceptional and edge-case tests for more effective fuzzing.

📝arxiv.org/abs/2304.02014

@steven_xia_ is presenting fuzz4all, the first universal fuzzer via large language models, at #icse24! Joint work with Matteo Paltenghi, Jiale Tian, and Michael Pradel.

🧑‍💻 fuzz4all.github.io
📝 arxiv.org/abs/2308.04748