Astar: Alibaba & Zhejiang Univ's AI That Guides AI Evolution, Beating Human Experts 100x Faster

Background: The Human Intuition Bottleneck in AI Iteration

Industrial AI model iteration differs fundamentally from ordinary software engineering. Changing a model requires expensive validation cycles — retraining and offline evaluation can take hours to days — so teams cannot brute-force all ideas. Historically, the decision of which direction to pursue has relied on senior algorithm experts' intuition, creating a scarce-expertise bottleneck.

Why General LLMs Fail in Industrial Settings

General LLMs absorb vast public knowledge but struggle to translate it into actionable advice for a specific, highly customized production system. Engineers who follow generic suggestions often wait days for training only to see metrics stagnate or regress, wasting compute and time.

Astar: Learning from the System's Own Evolutionary History

Astar takes a different approach: it mines the AI system's own version-control history. Every Git commit records code changes alongside loss curves and business metric shifts, forming a real-world, feedback-rich "pitfall guide." The team turned these scattered records into training data for a specialized LLM that proposes evolution directions conditioned on the current system state.

Four Challenges and Astar's Solutions

C1 – Data scarcity: Pairwise augmentation. Instead of only adjacent commits, any two historical experiments are paired (e.g., v1 vs v5); the lower-loss version becomes the positive sample. This expands sparse logs into dense supervision and teaches both short-term tweaks and long-term leaps.

C2 – Noisy commits: Two-stage filtering. (1) Execution-logic filtering via call-graph reachability and AST analysis removes logging, dead code, and formatting changes. (2) Evolutionary-intent filtering uses an LLM to semantically classify remaining diffs; only changes with a clear optimization intent are kept.

C3 – Vast search space: Hierarchical hints. Training data is labeled with a three-level taxonomy: high-level direction → sub-module → concrete action. At inference, Astar first picks a direction (e.g., "optimizer"), then a module ("Muon / gradient orthogonalization"), then a specific operation ("gradient denoising"), constraining the huge space into a structured generation process.

C4 – Expensive validation: Offline reward model. A reward model trained on historical positive/negative samples predicts the probability that a proposed change lowers loss in ~1 second. This enables rapid pre-filtering of candidates and provides cheap RL feedback for Astar's own post-training.

Training Pipeline and Benchmark Results

Astar follows a standard three-stage regimen: mid-training on the curated evolution corpus → supervised fine-tuning (SFT) with hierarchical hints → reinforcement learning (RL) guided by the reward model. On a recommendation-system optimization task:

The 0.6B parameter Astar achieves a single-generation success rate (S@1) of 54.35% .

The 8B version reaches 67.86% S@1.

Both significantly outperform GPT-5.5 (30.71%) and human senior algorithm experts (32.29%).

This demonstrates that in vertical engineering decision-making, feeding a model high-quality, system-specific evolutionary history is as critical as scaling generic parameters.

Production Deployment on Lazada Ad Recommendation System

Astar was given full control of Alibaba Lazada's core advertising recommendation system for 20 consecutive automated iterations with zero human intervention. Results:

Offline HitRate@200 increased by 23.6% .

Online A/B test: GMV (gross merchandise value) + 4.86% , ad revenue + 1.82% , CTR and order volume also up significantly.

In a mature, heavily optimized e-commerce ad stack, such gains from purely AI-proposed changes are remarkable.

Case Study: Muon Optimizer with Spectral Denoising

When the team tried introducing the Muon optimizer, its aggressive gradient orthogonalization amplified noise in the recommendation setting. Generic advice would be "tune learning rate" or "revert to Adam." Astar instead proposed: apply spectral denoising before orthogonalization , using the Marchenko–Pastur law from random matrix theory to compute a hard threshold — signals above the edge are kept, those inside the noise bulk are suppressed. This preserved Muon's benefits while eliminating noise amplification, with zero extra training parameters.

Closing the Loop: Toward Meta-AI

The team designed a closed loop: Astar proposes direction → Code Agent implements and runs experiment → results feed back as new training data for Astar. This makes Astar a "living" model that improves alongside the system it guides. The vision shifts from humans manually tuning models to humans designing Meta-AI that guides AI , freeing algorithm engineers to focus on frontier architecture breakthroughs, new business paradigms, and building smarter evolution flywheels.