
SkyDiscover: Self-Improving AI for the AI Systems Stack
Coding agents can optimize systems for specific workloads and hardware. But better benchmark results don’t always mean better software: agents can exploit gaps in tests, improving performance metrics while violating intended behavior.
In this Nebius Science Paper Club session, Shu Liu, a Computer Science PhD student at UC Berkeley, advised by Ion Stoica, presents SkyDiscover-Synthesize (SkySynth).
SkySynth brings implementation and specification into the same improvement loop. It combines test-driven synthesis, where an auditor turns discovered loopholes into new tests, with proof-driven synthesis for formally specified requirements. We’ll examine examples across key-value stores, inference engines and model routers, the reported performance gains, and the challenges that remain for real-world deployment.
What you’ll learn
- Why workload-specific systems can outperform general-purpose designs.
- How reward hacking can produce misleading performance gains.
- How SkySynth evolves code alongside tests or machine-checked proofs.
- What the case studies reveal about specialization, correctness and remaining limitations.
Who should attend
ML researchers, systems engineers and PhD students working on coding agents, AI infrastructure or automated software optimization. Familiarity with coding agents and basic systems concepts will help you follow the discussion.
Recommended reading: Building Specialized Systems We Can Trust with Agents
About Nebius Science Paper Club
Nebius Science Paper Club is a webinar series led by Nebius researchers. Each session brings together paper authors and practitioners to discuss new ideas, research and discoveries in AI.