🦁 Lion, new optimizer discovered by Google Brain using genetic algorithms that is purportedly better than Adam(w), in Pytorch
-
Updated
Jul 9, 2026 - Python
🦁 Lion, new optimizer discovered by Google Brain using genetic algorithms that is purportedly better than Adam(w), in Pytorch
Efficiently discovering algorithms via LLMs with evolutionary search and reinforcement learning.
Official Pytorch Implementation of Length-Adaptive Transformer (ACL 2021)
Official code, models, and dataset for "Evolution Fine-Tuning (EFT): Learning to Discover Across 371 Optimization Tasks"
resemble is an R package for similarity-based modelling and local learning in spectroscopy. It provides tools for dissimilarity computation, nearest-neighbour search, memory-based learning, and spectral library optimisation (methods designed for large, heterogeneous spectral datasets where global models underperform)
Lightweight, harness-agnostic scaffold for autonomous self-improving loops (Agentic Variation Operators for any agent). bash + git + jq.
A collection of agent skills that give AI agents structured, reusable workflows for tackling complex problems.
METAL-SCI: a scientific compute benchmark for evolutionary LLM kernel search on Apple Silicon Metal
[NeurIPS'22] Does Self-supervised Learning Really Improve Reinforcement Learning from Pixels?
Fork of Karpathy's autoresearch with novelty-search-inspired autonomous search, run memory, and agentic review.
Official implementation of FormalEvolve, accepted to Findings of EMNLP 2026. Neuro-symbolic Lean 4 autoformalization.
Native RNN substrate program for structured recurrence and gate-state causal propagation.
Adversarial Testing Lab for Agentic Safeguards (ATLAS). A synthetic multi-agent eval environment for adversarial fraud decisioning inspired by Anthropic's Project Deal. Measures how model quality, tool access, and agent orchestration affect attack discovery & defensive recovery, with deterministic evals and realistic customer-friction limits
Claude Code plugin for evolutionary ML architecture discovery. Reconstructs mathematical results from first principles without retrieving them from training data.
Searches Particle-Lenia interaction laws. A law is a flat vector of floats: the engine runs it as a simulation, the tuner measures what emerged and scores it, and the search proposes the next one. Rust, CPU and GPU backends.
A verifier-first runtime for independently verifying, comparing, and searching AI-generated executable solutions.
Reusable GEPA optimizer framework for typed candidate generation, evaluation, persistence posture, tracing contracts, and deterministic prompt optimization infrastructure.
Model-driven evolutionary search engine for unresolved problems.
AI autoresearch as a git DAG: bandit-scheduled lineages, agent-written semantic merges, cross-lineage replication as a reward-hacking detector
To associate your repository with the evolutionary-search topic, visit your repo's landing page and select "manage topics."