Shubham Agarwal
Hi! I am a Ph.D. candidate in Computer Science at UC Berkeley, in the Sky Computing Lab and the EPIC Data Lab. I am advised by Prof. Ion Stoica and Prof. Aditya Parameswaran. I am also a Student Researcher at Google Cloud AI Research, co-hosted by Chun-Liang Li, Rui Meng and Violet Peng.
My research is on self-improving AI and systems. I build agents that continuously optimize systems such as LLM serving engines and GPU kernels, and agents that build new systems backed by tests and machine-checked proofs. I am a core developer and maintainer of SkyDiscover, our open-source framework for self-improving systems, which has 800+ GitHub stars and is used at Google, Uber and AMD. Within it, I lead DuoDiscover, now integrated into Google Antigravity.
Before Berkeley, I was a Research Associate at Adobe Research, working with Subrata Mitra and Shiv Kumar Saini on generative-AI serving and cloud reliability systems that shipped in three Adobe products. I received my Bachelor's in Computer Science from BITS Pilani in 2022.
The best way to reach me is email: shubham3@berkeley.edu.
More about my research
Continuous optimization
Agents that keep optimizing a running system, reaching 2.17× over AMD's expert-tuned SGLang.
Agents write fast evaluators for slow benchmarks, then recheck on the real system. First on 121 of 235 SOL-Exec GPU kernels.
Building trustworthy systems
Agents that build new systems backed by tests and machine-checked proofs.
Agents write code with its Rocq proof, for distributed key-value stores. The code runs up to 3× faster than hand-verified code.
Agents build a system per workload and turn each caught shortcut into a test. Key-value stores beat FASTER on 14 of 18 workloads.
News
| Nov 2026 | Advanced to Ph.D. candidacy at the beginning of my second year. |
|---|---|
| Nov 2026 | Talks on SkyDiscover at the SGLang Summit, Microsoft Azure Research, AMD and Inferact. |
| Oct 2026 | DuoDiscover is out on arXiv. |
| Sep 2026 | Inductive Deductive Synthesis accepted as an Oral at NeurIPS 2026; RCStat also accepted. |
| Aug 2026 | Arming Data Agents with Tribal Knowledge accepted at VLDB 2026. |
| Jul 2026 | Best Paper Award at the ICML 2026 FAGEN Workshop for Fantastic Adaptive Taxonomies. |
| Jul 2026 | EvoX accepted at COLM 2026. |
| Jun 2026 | Joined Google Cloud AI Research as a Student Researcher. |
| Apr 2026 | SkyDiscover accepted as a Spotlight demo at ACM CAIS 2026, where I co-organized the AI Agents for Discovery in the Wild workshop. |
| Mar 2026 | Released SkyDiscover, our open-source framework for self-improving systems. |
| Aug 2025 | Started my Ph.D. at UC Berkeley. |
| Jul 2022 | Joined Adobe Research as a Research Associate. |
Publications 17
* equal contribution. Full list on Google Scholar.
-
NeurIPS'26RCStat: A Statistical Framework of Relative Contextualization in TransformersThe 40th Conference on Neural Information Processing Systems (NeurIPS) 2026
-
VLDB'26
-
ICLR'26FlowCast: Trajectory Forecasting for Scalable Zero-Cost Speculative Flow MatchingThe 14th International Conference on Learning Representations (ICLR) 2026
-
CIDR'26Supporting Our AI Overlords: Redesigning Data Systems to be Agent-FirstThe 16th Conference on Innovative Data Systems Research (CIDR) 2026
-
UAI'25Root Cause Analysis of Failures from Partial Causal StructuresThe 41st Conference on Uncertainty in Artificial Intelligence (UAI) 2025
Preprints 7
-
arXiv'26DuoDiscover: Scaling Autonomous Discovery Beyond the Evaluation BottleneckarXiv preprint 2026
-
arXiv'26The Time is Here for Just-in-Time Systems: Challenges and OpportunitiesarXiv preprint 2026; NeurIPS 2026 ML for Systems Workshop
-
arXiv'26Fantastic Adaptive Taxonomies and How to Use ThemarXiv preprint 2026; ICML 2026 FAGEN Workshop Best Paper Oral
-
arXiv'26Dialogue to Discovery: Attribute-Aware Preference Elicitation for Conversational Product Search AssistantsarXiv preprint 2026
Patents 8
-
USPTOUsing Reinforcement Learning to Recommend Data VisualizationsUS Patent 12,443,621 (granted)
-
USPTOData Exploration using Natural Language with Data SamplingUS Patent 12,664,160 (granted)
-
USPTOMulti-Modal Retrieval using an Intermediate Noise StateUS Patent 12,694,578 (granted)
-
USPTORelative Contextualization using Pre-Softmax Logits in TransformersUS Patent Application 19/285,322
-
USPTOManaging Chunk Caches for Efficient Retrieval-Augmented Generation in LLMsUS Patent Application 19/074,061
-
USPTOReCon: Acceleration for Text-to-Image Synthesis with Concept TrajectoriesUS Patent Application 63/698,535
-
USPTOMicromanaging Prompts for High-Throughput Text-to-Image InferenceUS Patent Application 18/808,654
-
USPTOA System and Method for Outage ForecastingUS Patent Application 17/656,263
Talks 15
† joint talk with co-authors.
| Dec 2026 | Inductive Deductive SynthesisOral presentation, NeurIPS 2026, Sydney (upcoming) |
|---|---|
| Nov 2026 | Perfmaxxing SGLang for Next-Gen Model Serving with SkyDiscoverSGLang Summit |
| Nov 2026 | Scaling Swarms of AgentsMicrosoft Azure Research (Systems) |
| Nov 2026 | Continuous Optimization of Inference Engines with Self-Evolving AgentsAMD × UC Berkeley |
| Nov 2026 | Towards Continuous Optimization of LLM Inference Systems with SkyDiscoverInferact† |
| Nov 2026 | SkyDiscover-SynthesizeBroadcom × Sky Computing Lab Research Review |
| Oct 2026 | SkyDiscover-SynthesizeEPIC Data Lab Retreat, Berkeley |
| Oct 2026 | SkyDiscover-OptimizeAMD, San Jose† |
| Oct 2026 | Supporting Our AI Overlords: Redesigning Data Systems to be Agent-FirstFrontier Research Club† |
| Sep 2026 | Agents for OptimizationGoogle |
| Jun 2026 | SkyDiscover: A Flexible Framework for AI-Driven Scientific and Algorithmic DiscoveryAIDDA 2026† |
| Apr 2026 | Arming Data Agents with Tribal KnowledgeUC Berkeley Database Seminar |
| Nov 2024 | Approximate Caching for Efficiently Serving Text-to-Image Diffusion ModelsUW Systems Group |
| Apr 2024 | Approximate Caching for Efficiently Serving Text-to-Image Diffusion ModelsNSDI 2024, Santa Clara |
| Sep 2023 | ESRO: Experience Assisted Service Reliability against OutagesASE 2023, Luxembourg |
Honors and Service
- Best Paper Award, ICML 2026 FAGEN Workshop, for Fantastic Adaptive Taxonomies and How to Use Them
- UC Berkeley EECS Ph.D. Fellowship (2025)
- Laude Institute Slingshots grant (SkyDiscover team, 2026)
- Co-organizer and Area Chair, AI Agents for Discovery in the Wild workshop, CAIS 2026
- External reviewer for NeurIPS, ICLR, ATC, AISTATS and AAAI
- Officer, Computer Science Graduate Entrepreneurs (CSGE), UC Berkeley, helping turn research into startups
- Industry mentor to a UMass master's student team researching LLM-based RAG systems