Divyanshu Saxena

And a snap of me presenting this work at NSDI 2026 in Seattle last week.

This is joint work w/ Gaurav, Jiaxin, @bobowang2014, @IsilDillig, Sanjay, and @adityaakella Catch this work at NSDI first day, first session, first talk (track II)!

**Performance robustness** matters for business-critical applications as much as safety and correctness -- yet, even learned controllers miss latency targets under changing environments. Next week, I will be at NSDI to present my work that addresses this challenge!

This is joint work with @chenxiyang_ut, Rohit, @kshiteejmahajan, @swarat, and @adityaakella If you are at EuroSys, please check out the talk during Day 2 afternoon session (track B) and reach out if there are any thoughts/questions!

Key idea in Canopy: we construct a quantitative feedback from formal verifiers and integrate this feedback *into the training loop* of learned controllers. Thus, the neural network learns to not just optimize for empirical performance but also: how to meet desired properties!

Learned congestion controllers may not always result in good performance - even SOTA learned controllers can result in bandwidth starvation, and other types of poor behavior! Our work on making learned controllers meet desired properties will appear at EuroSys this week!

Our @ACMSIGOPS blog post lays out the LDOS roadmap for learned OS policies. Would love to hear thoughts from the systems + ML community! Co-written with @divytweet Aneesh Durg, Sujay Yadalam, Jiayi (Jane) Chen, Rohit Dwivedula, and Chris Rossbach. Many thanks to the @NSF CISE Expeditions program for supporting this work.

Thanks for featuring our article! This blog post highlights the motivation behind and the vision for LDOS @ldosexpedition. We are actively working towards the next horizon of intelligent Operating Systems!

New SIGOPS Blog -- "LDOS: Toward A Learning-Directed Operating System" by by Divyanshu Saxena, Aneesh Durg, Sujay Yadalam, Jane Chen, Rohit Dwivedula, Chris Rossbach, and Aditya Akella. sigops.org/2026/ldos-towa…

We look at this "joint learning" problem in the context of hardware caching and prefetching, and propose two techniques: either learning shared embeddings using a joint encoder, or using contrastive learning to align individual embeddings.

If you are attending @NeurIPSConf, do check out our poster at the ML for Systems workshop on a novel approach for improving interdependent systems policies by developing shared representations!

For decades, OSes have relied on hand-tuned heuristics. What if they could write their own? 🔥 Meet PolicySmith, our LLM-driven framework that generates high-performing, interpretable system code. 📄arxiv.org/abs/2510.08803 at #HotNets25 #LDOS #MLforSystems #CodeGeneration

Joint work by Rohit, myself, and our advisors @adityaakella, @daehyeok_kim, and @swarat.

Crucially, we identify what this high-level specification should contain - abstracting the overall search into a small number of components. We apply this to Web Caching and Congestion Control; and discover heuristics that outperform established baselines in our prototype!

Our vision for using LLM-driven search for discovering better systems heuristics is set to appear at HotNets'25! Our thesis is that we can achieve **instance-optimality** using advanced code-generation and reasoning capabilities of LLMs!

@sarsanaee Wow, is that a decade worth of electricity?!

My lovely bill!