人才雷达 · Talent Radar

★★ HexGen / HexiScale / LLMFabric — decentralized LLM training & serving on heterogeneous clusters. Direct fit for JouleBeat token scheduling at infrastructure level. Young AP, OSS-heavy.

主页研究方向白纸黑字写着 'LLM Long-Horizon Optimization'——用 LLM 做复杂多步优化 + code-agent 开发,与 JouleBeat '长时程 agent runtime' 命题正中靶心。CUHK-深圳本科(与 Jim 同根 warm path)+ NUS RA + 第46届 ICPC World Finalist,工程与算法双强。Sky Computing Lab 博一,与已触达的 7 位同门(Shiyi Cao/Tyler Griggs/Zhanghao Wu/Shu Liu/Ziming Mao/Dacheng Li/Nicholas Lee)同 lab,群辐射效应可借力。年轻、availability 高。

Cambridge ML PhD + Schölkopf postdoc, now leads SphereLab at CUHK CSE focusing on 'Scalable Principles for Learning and Reasoning' — foundation models + LLM inductive bias is upstream of JouleBeat's representation work. Young AP, ~17-student lab.

First author of A3: Android Agent Arena — direct mobile GUI agent evaluation benchmark, exactly the long-horizon eval substrate JouleBeat wants. CUHK MMLab same-root.

LMCache 的命名者与核心实现者,把 KV cache 抽出 GPU、跨 engine 共享、与 vLLM 集成实现 15× 吞吐提升。LMCache 已成为 vLLM 生态最活跃的 KV cache layer。**直接命中 JouleBeat 的 token 调度约束架构核心**。

RAP recently appointed — leads agent + RAG research at PolyU NLP. PEToolLLM ACL'25.

Verlog (NeurIPS'25 Workshop) 一作——第一个能训 400+ turn trajectory 的多轮 RL 框架，引入 'Dual-discounting GAE' 解耦 step-level 与 token-level credit assignment。这就是 JouleBeat 命题中 long-horizon credit assignment 的代表性工作。

FinAI + multi-agent systems; international background (Yale + Manchester); just returned to WHU.

UltraHorizon(2509.21766)一作,把 agent 评测推到 35k+ token / 60+ tool call / 最高 200k+ token / 400+ tool call 的'超长时程'范围——**这就是 JouleBeat 要攻的分布**。揭示 in-context locking 和 functional gap 是两类核心失效模式,对后续 runtime 设计有直接方向意义。

Plan-and-Act(ICML 2025)一作,long-horizon planning + executor 分层。Berkeley BAIR。

Long-running MISC Lab — recruiting PhDs for 'trustworthy AI, federated learning, GNN, NLP'. Senior CUHK voice with industry reach; potential talk host or advisor.

SphereLab PhD with focus on reasoning + RL substrate. Active homepage.

DeepServe (ATC'25)、RaaS (ACL'25 reasoning-aware attention sparsity)、EPIC (ICML'25 position-independent prefix caching) 一作——专精 serverless LLM serving + 冷启动 + 长 decode。JouleBeat serverless agent runtime 的关键人。

自我描述 'LLM RL, especially Agentic RL and efficient RL training for LLMs，long-horizon planning + memory-aware training + real-world infrastructure constraints'——这一句话基本就是 JouleBeat slogan。也是 SGLang RL 模块贡献者 + GLM-4.5/GLM-5 agentic 核心贡献者。

AgentScope (1.0) 一作 / lead maintainer——'developer-centric agentic framework',抽象 message / model / memory / tool 四模块 + ReAct 范式 + asynchronous design,正面对位 JouleBeat runtime,有大量可借鉴的工程经验。

Transfer-learning founder, moved from NTU to CUHK CSE — strong RL/representation pedigree. Senior advisor / talk host candidate.

Web search + NLP + recommender systems; long publication record incl. SIGIR.

AgentOrchestra(89.04% GAIA)一作,把 orchestration 抽象为 TEA 协议 + 版本化生命周期,runtime 架构思考清晰。

★ Highest CUHK affinity — Hongsheng inherited Tang Xiaoou's MMLab. Massive output on multimodal LLMs + GUI agents (A3 Android Agent Arena, EasyRef). Tao 同根直系。

★★★ Jim's direct warm node — Tao 是 Jim 的 CUHK IE 同门(汤晓鸥时代)且现在 HKU XLANG 是 GUI agent / OSWorld 顶级 lab. 必须列入。