Alibaba and Zhejiang University unveil Astar, an LLM that learns from a system's own evolution history
jiqizhixin · x · 2026-09-21
Alibaba and Zhejiang University present Astar, an LLM that guides AI system evolution by deep-learning the system's own history — code commits and experiment results. It filters noise via code call-graph reachability analysis and AST (stripping logging, dead code, formatting changes), internalizes real evolution experience, and adaptively explores future directions, closing the loop from idea to code to training, evaluation and deployment. Reported result: 10–100x boost in algorithm engineers' new-idea efficiency.
More from Companies & People
- India's biggest AI build session announced: free entry with $500 in prizes — pritisinghhhh · 2026-09-21
- Anthropic, OpenAI, xAI and Google Sued Over Coordinated AI Slowdown — DavidLinthicum · 2026-09-21
- Modal x Gray Area exhibition to showcase 13 commissioned AI art works with $20k compute grants — charles_irl · 2026-09-21
- Senior engineers offering referrals but big companies have almost no entry-level AI roles — cneuralnetwork · 2026-09-21
- 66% of leaders won't hire without AI skills — here are 7 you can prove with projects — thisdudelikesAI · 2026-09-21
- Codex Community Hosts ~30 Global Meetups This Week, From Osaka to Nairobi — paw_lean · 2026-09-21