ADRS

🎯 Can we make AI-driven algorithm discovery dramatically cheaper without sacrificing performance? In our latest AI-Driven Research for Systems (ADRS) case study, we feature LEVI: an LLM-based evolutionary framework built around a simple idea — invest in the search harness, not just the model. Today’s ADRS frameworks can produce strong algorithms, but they are often too expensive for the future we actually want: continuous, bespoke optimization tailored to each deployment’s exact workload, hardware, and SLOs. LEVI is designed to lower that barrier. Instead of relying on expensive frontier models for every mutation, LEVI uses smaller, cheaper models for most refinements and reserves larger models for rarer paradigm shifts. It combines this with a stronger diversity mechanism that maintains variation across both code structure and behavior. ⚡ The results: Best score on every ADRS benchmark where improvement is possible ~3–7× cheaper than baselines in the main benchmark comparison Roughly $4.50 per problem on most tasks, versus $15–$30 for baselines This is the direction ADRS should go next: not one expensive run for everyone, but cheap, repeatable optimization for each real deployment. ✍️ Read the blog: https://lnkd.in/gP3mk-Fk 📚 ADRS Blog Series: https://lnkd.in/gqtJ-GJ7 📄 ADRS Paper: https://lnkd.in/gdfjA26V 👩💻 LEVI Code: https://lnkd.in/g76VMTyM 🚀 Follow: https://lnkd.in/gCA_mFGG for updates! Special thanks to Temoor Tanveer for this work! 💫

Can AI learn how to optimize itself? Researchers spend hours and hours hand-crafting the strategies behind LLM-driven optimization systems like AlphaEvolve: deciding which ideas to reuse, when to explore vs exploit, and what mutations to try. 🤖But what if AI could evolve its own evolution process? We introduce EvoX, a meta-evolution pipeline that lets AI evolve the strategy guiding optimization itself. EvoX treats the evolution strategy as something that can improve over time. The system runs two coupled loops: (1) Solution evolution – generating and evaluating candidate solutions (2) Strategy evolution – adapting how new candidates are generated Across ~200 optimization tasks, EvoX delivers strong results: • 🥇 Best open-source performance on Frontier-CS, improving median scores by ~34% across 172 programming problems • 🎯 Matches or surpasses AlphaEvolve and prior human SOTA on 6/8 math benchmarks and all 7 systems optimization tasks • ⚙️ Real systems improvements discovered automatically, including 41% lower cross-cloud transfer cost, 14% better GPU load balance for MoE serving, and 29% lower KV-cache pressure • 💸 Highly cost-efficient discovery, breaking optimization plateaus for <$5 compute on tasks where existing frameworks spend 3x more and still stagnate EvoX is fully open-source and built on top of SkyDiscover. If you're interested in AI-driven discovery, evolutionary search, or automated algorithm design, we’d love for you to try it out! 👉 Read the Blog: https://lnkd.in/gJV-7Szy  💻 Code: https://lnkd.in/gx2BFur5 📄 Paper: https://lnkd.in/gexX8z2x Huge thanks to my incredible collaborators: Shubham Agarwal, Monishwaran Maheswaran, Mert Cemri, Zhifei Li, Qiuyang Mang, Ashwin Naren, Ethan Boneh, Audrey Cheng, Melissa Pan, Alexander Du, Kurt Keutzer, Alvin Cheung, Alex Dimakis, Koushik Sen, Matei Zaharia, and Ion Stoica, and many others. We also thanks Jiarong Xing, Asankhaya Sharma, Joseph Gonzalez, and Alex Krentsel for their useful feedback!

AlphaEvolve showed LLMs can discover novel algorithms, but it's closed-source, and open alternatives rely on static search policies. We introduce AdaEvolve, a fully adaptive evolutionary algorithm for LLM-driven discovery. It dynamically adjusts exploration rates, resource allocation, and search strategies based on observed progress and no manual tuning required. The problem: Existing frameworks (OpenEvolve, GEPA, ShinkaEvolve) use fixed controllers, such as static exploration rates, rigid prompts, uniform compute allocation, which are all decided before the run begins. The LLM mutation operator is sophisticated; the algorithm governing it is blind to whether search is progressing or stuck. The idea: Fitness improvement trajectories act as a gradient analogue for gradient-free optimization. A single accumulated improvement signal coordinates three adaptation levels: → Local: dynamically modulates exploration vs. exploitation within each subpopulation  → Global: bandit-based compute routing across subpopulations, with dynamic spawning when all stagnate  → Meta-Guidance: when numerical adaptation is insufficient, a separate LLM generates new high-level solution tactics Results across 185 problems (same models, same budgets): Math (6 tasks): Best open-source on all 6. Matches or exceeds AlphaEvolve on 4/6, including new best-known results on Circle Packing. Systems (7 ADRS tasks): Wins all 7 across GPT-5 and Gemini-3-Pro. 41% lower cross-cloud transfer cost, 14% better MoE load balance. Frontier-CS (172 algorithmic tasks): +34% median score over the strongest open-source baseline. Single-call GPT-5 gets a median of 0 on these problems. This is a clear indication that evolutionary scaffolding matters. AdaEvolve is built on SkyDiscover, our open-source framework for AI-driven scientific and algorithmic discovery, and requires only ~2,500 lines of code on top of the framework's modular architecture. Joint work with Shubham Agarwal, Akshat Gupta, Shu Liu, Audrey Cheng, Ashwin Naren, Qiuyang Mang, Lutfi Eren Erdogan, Koushik Sen, Matei Zaharia, Alex Dimakis and Ion Stoica. Also many thanks to Alex Krentsel and Joseph Gonzalez for fruitful discussions. 📄 Paper: arxiv.org/abs/2602.20133  💻 Code: https://lnkd.in/euWyEyki  📝 Blog: https://lnkd.in/gYrfUjEh

AI systems are now discovering novel algorithms that surpass human experts. However, the frameworks powering these breakthroughs, such as AlphaEvolve, are often closed-source. Meanwhile, existing open-source alternatives remain rigidly coupled and monolithic, making it incredibly difficult to iterate or test new ideas. Introducing 🌟SkyDiscover🌟: a flexible framework for AI-driven scientific and algorithmic discovery. Using SkyDiscover, we built two new algorithms that match or exceed AlphaEvolve and outperform every other open-source alternative across 200+ benchmarks across math, systems, competitive programming, and even image generation (multi-modal) tasks. Research in AI-driven discovery has been severely bottlenecked. If you want to test a new selection strategy, you usually have to rip apart a core system. SkyDiscover fixes this by cleanly decoupling the discovery loop into four reusable primitives: Context Builder, Solution Generator, Evaluator, and Solution Selector. To demonstrate its power, we used SkyDiscover's modular playground to implement two novel evolutionary algorithms, AdaEvolve and EvoX, in just ~2.5K lines of code. By dynamically adapting their search strategies based on real-time progress, and even using LLMs to continuously rewrite their own discovery code on the fly, these algorithms achieved state-of-the-art results across 200+ benchmarks: 🚀 Real-world systems impact: Discovered a data-routing policy that cuts cross-cloud transfer costs by 41%, an MoE load-balancing strategy that is 14% more balanced and 9.5× faster, and a GPU placement algorithm that reduces KV-cache pressure by 29% (beating the published human SOTA). 🏆 Best open-source performance: Improved median scores by ~34% across 172 Frontier-CS programming problems over OpenEvolve, GEPA, and ShinkaEvolve. 💪 Matching AlphaEvolve: Matched or exceeded AlphaEvolve and human SOTA baselines on 8 math and 6 systems optimization tasks. SkyDiscover provides a unified interface for running and comparing methods fairly across math, systems, code, and even multi-modal optimization. We want the community to try it, break it, and build the next generation of algorithms on it. 🌐 Blog: https://lnkd.in/ewgBR5ci 🔗 Code: https://lnkd.in/euWyEyki 📄 AdaEvolve Paper: arxiv.org/abs/2602.20133 📄 EvoX Paper: arxiv.org/abs/2602.23413 Huge thanks to my incredible collaborators: Shu Liu, Shubham Agarwal, Alex Krentsel, Ashwin Naren, Qiuyang Mang, Zhifei Li, Akshat Gupta, Monishwaran Maheswaran, Audrey Cheng, Melissa Pan, Ethan Boneh, Kannan Ramchandran, Koushik Sen, Alex Dimakis, Matei Zaharia, and Ion Stoica. Also, huge thanks to Jiarong Xing, Asankhaya Sharma, and Joseph Gonzalez for their valuable feedback! #MachineLearning #ArtificialIntelligence #LLMs #OpenSource #Optimization #Systems

AI systems are now discovering novel algorithms that surpass human experts. However, the frameworks powering these breakthroughs, such as AlphaEvolve, are often closed-source. Meanwhile, existing open-source alternatives remain rigidly coupled and monolithic, making it incredibly difficult to iterate or test new ideas. Introducing 🌟SkyDiscover🌟: a flexible framework for AI-driven scientific and algorithmic discovery. Using SkyDiscover, we built two new algorithms that match or exceed AlphaEvolve and outperform every other open-source alternative across 200+ benchmarks across math, systems, competitive programming, and even image generation (multi-modal) tasks. Research in AI-driven discovery has been severely bottlenecked. If you want to test a new selection strategy, you usually have to rip apart a core system. SkyDiscover fixes this by cleanly decoupling the discovery loop into four reusable primitives: Context Builder, Solution Generator, Evaluator, and Solution Selector. To demonstrate its power, we used SkyDiscover's modular playground to implement two novel evolutionary algorithms, AdaEvolve and EvoX, in just ~2.5K lines of code. By dynamically adapting their search strategies based on real-time progress, and even using LLMs to continuously rewrite their own discovery code on the fly, these algorithms achieved state-of-the-art results across 200+ benchmarks: 🚀 Real-world systems impact: Discovered a data-routing policy that cuts cross-cloud transfer costs by 41%, an MoE load-balancing strategy that is 14% more balanced and 9.5× faster, and a GPU placement algorithm that reduces KV-cache pressure by 29% (beating the published human SOTA). 🏆 Best open-source performance: Improved median scores by ~34% across 172 Frontier-CS programming problems over OpenEvolve, GEPA, and ShinkaEvolve. 💪 Matching AlphaEvolve: Matched or exceeded AlphaEvolve and human SOTA baselines on 8 math and 6 systems optimization tasks. SkyDiscover provides a unified interface for running and comparing methods fairly across math, systems, code, and even multi-modal optimization. We want the community to try it, break it, and build the next generation of algorithms on it. 🌐 Blog: https://lnkd.in/ewgBR5ci 🔗 Code: https://lnkd.in/euWyEyki 📄 AdaEvolve Paper: arxiv.org/abs/2602.20133 📄 EvoX Paper: arxiv.org/abs/2602.23413 Huge thanks to my incredible collaborators: Shu Liu, Shubham Agarwal, Alex Krentsel, Ashwin Naren, Qiuyang Mang, Zhifei Li, Akshat Gupta, Monishwaran Maheswaran, Audrey Cheng, Melissa Pan, Ethan Boneh, Kannan Ramchandran, Koushik Sen, Alex Dimakis, Matei Zaharia, and Ion Stoica. Also, huge thanks to Jiarong Xing, Asankhaya Sharma, and Joseph Gonzalez for their valuable feedback! #MachineLearning #ArtificialIntelligence #LLMs #OpenSource #Optimization #Systems

As we get better and better autonomous translations of existing software, I wrote a short bit about how they work. (https://lnkd.in/euza_udh)