xiao9905 Goto Github PK

I am Xiao Liu, a third-year PhD student in Tsinghua University since 2021.

• 🔭 Interested in Machine Learning, Natural Language Processing, and Foundation Models.

• 🌱 Find my up-to-date publication list in Google Scholar! Some of my proud leading works:

  Large Language Model (LLM) Training and Prompt Learning

  • P-tuning and P-tuning v2 (ACL'22): pioneer works on prompt tuning
  • GLM-130B (ICLR'23): an open bilingual (Enligsh & Chinese) pre-trained model with 130 billion parameters based on GLM (ACL'22); better than GPT-3 175B on LAMBADA and MMLU.
  • ChatGLM-6B & ChatGLM2-6B & ChatGLM3-6B: an open bilingual dialogue language model that requires only 6GB to run. Receiving , , and GitHub Stars!
  • WebGLM (KDD'23): an efficient web-enhanced question answering system based on GLM-10B, outperforming WebGPT-13B and approaching WebGPT-175B performance in human evaluation.
  • ChatGLM-Math: employing self-critique with RFT and DPO to enable SOTA mathematical capabilities wihtouth compromising language abilities.
  Foundational Agents For Real-world Challenging Missions

  • AgentBench (ICLR'24): the first systematic multi-dimensional benchmark to evaluate LLMs as Agents in 8 distinct environments deriving from real-world practical missions. Find LLM-as-Agent demos at llmbench.ai/agent!
  • AutoWebGLM (KDD'24): a strong web navigating agent constructed upon ChatGLM-3-6B, outperforming prompted GPT-4 on Mind2Web, WebArena, and our constructed new dataset AutoWebBench.
  Alignment and Scalable Oversights over LLMs and Diffusers

  • ImageReward (NeurIPS'23): the first general-purpose text-to-image human preference reward model (RM) for RLHF, outperforming CLIP/BLIP/Aesthetic by 30% in terms of human preference prediction.
  • BPO (Black-box Prompt Optimization, ACL'24): a novel direction to align LLMs via preference-aware prompt optimization. Improving ChatGPT, Claude, LLaMA on human preference's win rates by 20%+ without training them.
  • AlignBench (ACL'24): the first comprehensive benchmark on evaluating LLMs' Chinese alignment, deriving from ChatGLM's online real scenarios. Submit your LLMs to acquire CritiqueLLM's judgement on AlignBench on llmbench.ai/align!
  • CritiqueLLM (ACL'24): scaling LLM-as-Critic for scalable oversights on LLM alignment. A series of strong critqiue LLMs ranging from 6B to 66B.
  Self-supervised Learning and Reasoning

  • Self-supervised Learning: Generative or Contrastive (TKDE'21): one of the most cited survey on self-supervised learning
  • SelfKG (WWW'22): self-supervised alignment can be comparable to supervised ones, Best Paper Nominee in WWW 2022.
  • kgTransformer (KDD'22): pre-training knowledge graph transformers with mixture-of-experts (MoE) for complex logical reasoning

• 🤔 Dedicated to building next-generation of AI systems via both Large Pre-trained Model and Symbolic Agent Reasoning.

• 💬 Feel free to drop me an email for:

  • Any form of collaboration
  • Any issue about my works or code
  • Interesting ideas to discuss or just chatting