Auto-evolved heuristics for capacity-constrained retired-battery reassembly. Combines a multi-attribute stochastic MDP with an AlphaEvolve outer loop that mutates heuristic operators (allocation, neighborhood search, feasibility filtering). Reaches <0.52% gap vs. the exact solver.
Distributed cascaded coordination between traffic-signal control and CAV platoon trajectory planning in mixed-autonomy urban networks. Produced one IEEE ITSC paper and one CSSCI-core journal article.
A 7-Agent architecture for creator-oriented long-form writing: hot-topic tracking (Zhihu / Toutiao / Weibo), 4-channel search (Google / Zhihu / BoCha / Tavily), adversarial critic review and multi-platform style adaptation. −70% topic-research time; per-article cost under ¥0.5.
Qwen3-0.6B + LoRA SFT with 3,120 seed → 7,612 turn-split samples, turn-by-turn loss weighting to fix tool-call over-triggering. Argument JSON validity 58% → 96%; end-to-end success 27% → 94%.
A domain-specialized multi-agent system for semiconductor research reports. Six LangGraph agents (plan / retrieve / analyze / chart / write / audit) with hybrid retrieval, credibility scoring and expert-critic feedback loops.