FOCAL Lab@UIUC

FOrmally Certified Automation and Learning (FOCAL) Lab at the University of Illinois Urbana-Champaign led by Prof. Gagandeep Singh!

[FOCAL Lab@UIUC @uiuc_focal_lab]

Formal guarantees for LLMs are finally practical 🔥

[Nalin Wadhwa @ ICLR 2026 @nalin_wadhwa]

🚨 Can we truly prove that LLMs won't leak your data or generate unsafe code?

Almost all LLM safety evals today only give estimates.
Sampling. Benchmarks. Leaderboards.
Zero guarantees ❌

We present BEAVER, the first system to provably bound LLM constraint satisfaction.

🧵👇

[Isha Chaudhary @Ish_cha_]

#ICLR2025 was a blast!
Grateful for the opportunity to present our work on formally certifying LLMs. For more details of our work, please check out our project page: https://t.co/ZN6s4yNLvU

[Shubham Ugare @shubhamdugare]

Meet us at #ICLR2025 next week to hear about IterGen! We introduce grammar-guided navigation for LLM generation --- precise forward and backward control to improve semantic accuracy.

[Isha Chaudhary @Ish_cha_]

UPDATE: QuaCer-B is accepted at #ICLR2025!

[Isha Chaudhary @Ish_cha_]

Happy to see support for our setup and methods in our LLM certification frameworks (https://t.co/OE1I25UJ2O, https://t.co/ZzOLBVbaKo) from @AnthropicAI.

[FOCAL Lab@UIUC @uiuc_focal_lab]

@TarunSures41845 will be in Vancouver at NeurIPS this week! Reach out to chat about deep learning, robustness, and verification 🙌

N/N

[FOCAL Lab@UIUC @uiuc_focal_lab]

🚀 Excited to present our paper "Relational DNN Verification Leaps Forward With RABBit" at #NeurIPS2024 on December 11th!

Authored by @TarunSures41845 and @debangshuban18!

1/N

[FOCAL Lab@UIUC @uiuc_focal_lab]

🔑 Key takeaway:
DNNs are safer against practical adversarial attacks than traditional input-specific verifiers suggest.

📄 Read our paper: https://t.co/OWQqZZ8WXw

5/N

[FOCAL Lab@UIUC @uiuc_focal_lab]

To address this, this paper introduces RABBit, a SOTA relational verifier that tracks dependencies between perturbed inputs under UAP attack, improving verified robustness accuracy by up to 6%.

N/N

[FOCAL Lab@UIUC @uiuc_focal_lab]

🎉 Exciting news! We're thrilled to announce that two of our papers have been accepted at #NeurIPS2024!

Details in the thread!

1/N

[FOCAL Lab@UIUC @uiuc_focal_lab]

Most DNN verifiers focus on input-specific adversarial attacks, but practical attacks often follow the UAP model, applying the same perturbation across inputs. This leads to overly pessimistic robustness guarantees.

6/N

[Siebel School of Computing and Data Science @siebelschool]

#ILLINOIS computer science professor Gagandeep Singh and PhD student Isha Chaudhary have determined that large language models have biased output that can spread misinformation and widen social gaps between various demographic groups. Read more!
▶️https://t.co/g8MDbQtJkX

[Isha Chaudhary @Ish_cha_]

Thanks @IllinoisCS for writing about our work here https://t.co/dvRpwdo3DZ!