Excited to share that our paper titled "Evolving Abstract Transformers for Gradient-Guided, Adaptable Abstract Interpretation" has been accepted at PLDI 2026! #PLDI2026 Huge thanks to my collaborators @debangshuban18 and @ggn_dp_sngh! Details in 🧵 [1/N]

There’s a lot of discussion around using LLMs to generate compiler code via prompting, but this on-the-fly approach can be unreliable and brittle. We propose 𝐑𝐮𝐥𝐞𝐅𝐥𝐨𝐰, a more reliable way to use LLMs for optimizing data-analysis programs (e.g, Pandas). (1/5)

Excited to share our @NeurIPSConf paper "Sketching for Distributed Deep Learning: A Sharper Analysis": openreview.net/pdf?id=0G0VpM… We provide a significantly improved convergence analysis for sketching-based distributed learning frameworks by exploiting the properties of the deep learning losses, such as restricted strong smoothness. With that, we break the dimension dependence in the convergence error and communication cost -- showing the promise of sketching for larger models. Unfortunately, I am not attending NeurIPS this year but visit our poster on Thursday 11-2 PST and ask Mayank any questions you may have. w/ Mayank Shrivastava, Qiaobo Li, @sanmikoyejo, Arindam Banerjee

Are you sure that nothing can drive your LLM towards boldly discriminating against protected demographic groups in nearly every other prompt? Can you guarantee that!? Well … now you can!! 📢 Introducing QuaCer-B, the first certification framework for bias in LLM responses. A🧵