New algorithmic frontiers.
PIKE (PyTorch Inference Kernel Evolution) is an LLM-based, multi-agent optimization system built with OpenEvolve that achieves up to 2.88× speedup over PyTorch eager and ~2× over torch.compile on frontier workloads. We present a logical framework for comparing multi-agent evolutionary optimization systems, and explore the configuration space for PyTorch and GPU performance optimization.
Bijective maps are a key tool in combinatorics. We explore whether evolutionary program synthesis frameworks like OpenEvolve can discover novel bijections, testing on challenging problems involving Dyck paths.
SLDAgent is an evolution-based coding agent built on OpenEvolve that automatically discovers scaling laws for large language models. On a new benchmark called SLDBench (5,000+ experiments across seven tasks), SLDAgent consistently discovers laws that are more accurate in extrapolation than human-designed laws—ranging from pre-training and fine-tuning to Mixture-of-Experts (MoE) scaling.
UC Berkeley's Sky Computing Lab used OpenEvolve as an open-source engine for AI-Driven Research for Systems (ADRS) to automatically discover and refine systems algorithms across multiple domains (for instance - MoE expert parallelism load balancer, multi-region spot scheduling, LLM-SQL preprocessing, transaction scheduling, and more). Reported results include up to 5× runtime speedups and double-digit percentage point cost reductions, often found within hours and sub-$20 evaluation budgets. Berkeley's team has documented their experience of using OpenEvolve and the results in their paper and blog.
Researchers from MIT and Stanford built GeoEvolve, a two-loop approach that combines an OpenEvolve-style evolutionary inner loop with an outer loop that retrieves and encodes geospatial domain knowledge ("GeoKnowRAG"). Applied to ordinary kriging (spatial interpolation) and geospatial conformal prediction (uncertainty quantification), GeoEvolve reports interpolation RMSE reductions of roughly 13–21% and interval-score improvements of about 17% on the evaluated datasets.
DeepEvolve builds on the OpenEvolve family of evolutionary coding agents - adding a research layer that plans questions, searches recent literature, synthesizes proposals, then implements and debugs changes across files. Instead of only mutating existing code, the system periodically injects new architectural seeds informed by papers and domain context, and evaluates them in the same reproducible loop as OpenEvolve.
Cerebras has applied CePO, or Cerebras Enhanced Planning and Optimization, to the GPT-OSS-120B model through their inference endpoint. As an OptiLLM technique, CePO leverages test-time computation for iterative planning and refinement, all without retraining the model. CePO works as an inference-time pipeline tailored for Cerebras hardware, allowing the model to plan, iterate on solutions, and refine outputs in real time. As part of the OptiLLM framework, it breaks down complex tasks like code generation into steps: outlining a plan, generating multiple attempts, analyzing for consistency, and picking the strongest result. This uses more tokens overall but turns hardware speed into an advantage for better reasoning, something that is tough on standard setups due to memory limits. The approach builds on earlier work with Llama models and has been extended in updates to models like DeepSeek R1 and Qwen QwQ 32B.
How do we teach machines to discover algorithms? Traditional approaches rely on hand-crafted heuristics, exhaustive search, or gradient-based optimization. But what if we could harness the creative potential of large language models (LLMs) within an evolutionary framework?