Ph.D. student @ HKUST(GZ)
Efficient and reliable LLMs: inference, long context, KV cache, retrieval, and agentic workflows.
|
400+ stars across 50 original repositories |
3.6k commits · 560+ PRs past year, public activity |
benchmark → method → artifact how I like research to ship |
|
Inference efficiency KV-cache compression, token-efficient reasoning, energy-to-token evaluation, serving bottlenecks. |
Long-context evaluation Generation-focused benchmarks, dense reasoning integrity, multi-turn coherence. |
|
Agent systems Tool use, post-training, harness design, local-first agent workflow infrastructure. |
Research infrastructure Reproducible artifacts, project pages, scholar tracking, figure and report tooling. |
|
Long-context generation benchmark for coherent, context-aware long-form responses — the gap left open by extractive needle-in-a-haystack tests.
|
Policy-conditioned live-market evaluation for LLM trading agents. Benchmark the policy, not just the model.
|
|
Position paper project page: LLM inference should be evaluated as energy-to-token production, not tokens per second alone.
|
Local-first agent task hub: SQLite queueing, dependency-aware dispatch, templates, pipelines, and a thin dashboard.
|
|
Adapters between XML-like tool calls and OpenAI-style structured tool-call histories — for post-training and eval harnesses.
|
Tooling for publication-grade scientific figures — reproducible, paper-ready plots without the last-minute scramble.
|
Merged PRs into other people's projects — 29 outside my own repositories.
LMFlow · 8.5k★15 merged PRs — LISA (reasoning-aware fine-tuning), DoRA and Hymba support, LoRA target-module fixes, eval-during-training, large-data preprocessing. |
kvpress · 1.2k★Merged |
|
Merged into awesome-vibe-coding (5.3k★), posterly (411★), alpaca_eval (2k★), Prompt-Engineering-Guide. |
The tooling layer around agent research — mostly reusable skills and curated indexes.
awesome-deepseek-harness · 353★A curated list of plugins, skills, MCP servers, patch/profile layers, orchestrators & UIs for DeepSeek Harness (DSH).
|
auto-research ·
auto-kernel-research ·
EvoHarness.skills Autonomous research loops, kernel optimization, academic review, spec-driven development, model-specific delegators. |
|
llm_benchmark_eval ·
llm-benchmark-research |
ai-semiconductor-bottleneck-atlas Web decks, reading notes, and project pages for papers. |
long-context generation ──┬── LongGenBench
├── semantic integrity under KV compression
└── multi-turn coherence / FlowKV
agent capability eval ────┬── QuantArena
├── tool-use adapters
└── local-first agent workflow runtime
efficient inference ──────┬── ChunkKV / KV compression
├── token-efficient reasoning
└── energy-to-token production
research tooling ─────────┬── benchmark & dataset builders
├── figure / report / poster pipelines
└── agent skill ecosystem (DSH, Codex, Claude)
