Lianhui Qin (@Lianhuiq) / X

Lianhui Qin

802 posts

Lianhui Qin

@Lianhuiq

Assistant Professor at UCSD CSE. NLP, ML, AI. I’m recruiting PhD students.

San Diego, CA

Joined October 2018

Pinned
Lianhui Qin
@Lianhuiq
May 20
Scaling embodied AI starts with automating the environments. Introducing SimWorld Studio: a self-evolving factory for endless interactive 3D environments where agents act, fail, and learn. With coding-agent + embodied-agent co-evolution, navigation success improves from 50%
00:00
58K
Lianhui Qin
@Lianhuiq
Jul 10, 2023
[Personal news] 📢So excited to share that I’ve just graduated @uwcse and will be joining @UCSanDiego @ucsd_cse 🌊☀️as an Assistant Professor in Fall 2024. Meanwhile I’m doing a postdoc @allen_ai. Look forward to working with students and colleagues on NLP, ML, etc.
226K
Lianhui Qin
@Lianhuiq
Oct 9, 2025
💡Our new paradigm: Latent diffusion for text reasoning. LLMs’ chain of thought (CoT) is linear and brittle. One token at a time. Irreversible. What if reasoning could self-refine globally at the semantic level? 🤔 We introduced LaDiR, which moves reasoning into latent space,
Murray Kang
@haoqik322
Oct 9, 2025
🧵1/ Latent diffusion shines in image generation for its abstraction, iterative-refinement, and parallel exploration. Yet, applying it to text reasoning is hard — language is discrete. 💡 Our work LaDiR (Latent Diffusion Reasoner) makes it possible — using VAE + block-wise
79K
Lianhui Qin
@Lianhuiq
Oct 7, 2025
🧠How can LLMs self-evolve over time? They need memory. LLMs burn huge compute on each query and forget everything afterward. ArcMemo introduces abstraction memory, which stores reusable reasoning patterns and recombines them to strengthen compositional reasoning. 📈On
Matt Ho
@matt_seb_ho
Oct 7, 2025
ArcMemo yields +7.5% relative on ARC-AGI vs o4-mini (same backbone). It extends the LLM idea of “compressing knowledge for generalization” into a lightweight, continually learnable abstract memory—model-agnostic and text-based. Preprint: Lifelong LM Learning via Abstract Memory
88K
Lianhui Qin
@Lianhuiq
Mar 29, 2022
We're thrilled to introduce 🧊COLD decoding, a general constrained text generation approach via energy based modeling🌞. We can plug in any differentiable constraints into an energy function, and apply Langevin Dynamics for efficient sampling. 🥳 paper arxiv.org/pdf/2202.11705…
Lianhui Qin
@Lianhuiq
Nov 29, 2023
🚀 PhD Opportunities in AI @ucsd_cse for Fall '24🌞 I'm recruiting PhD students with a passion for Large Language Models (LLMs), reasoning, generation, and AI for science. I'll be attending #NeurIPS2023 and #AAAI2024. Happy to catch up there. ☕️ #AIResearch #PhD #UCSD #LLMs
Lianhui Qin
@Lianhuiq
Jul 10, 2023
[Personal news] 📢So excited to share that I’ve just graduated @uwcse and will be joining @UCSanDiego @ucsd_cse 🌊☀️as an Assistant Professor in Fall 2024. Meanwhile I’m doing a postdoc @allen_ai. Look forward to working with students and colleagues on NLP, ML, etc.
126K
Lianhui Qin
@Lianhuiq
Dec 5, 2024
🎺My lab has several PhD openings in AI/ML/NLP @UCSanDiego @ucsd_cse 🌟Join us to explore machine reasoning & generation with multi-modality, at scale of multi- to massive-agents, using arbitrary domain constraints, for everyday and scientific applications. Plus, live the San
33K
Lianhui Qin
@Lianhuiq
Jun 24, 2024
💡Divergence thinking💡 is a hallmark of human creativity and problem-solving 🤖Can LLMs also do divergent reasoning to generate diverse solutions🤔? Introducing Flow-of-Reasoning (FoR) 🌊, a data-efficient way of training LLM policy to generate diverse, high-quality reasoning
00:00
50K
Lianhui Qin
@Lianhuiq
Mar 6, 2024
📢Introducing ❄️COLD-Attack⚔️, a unified framework for controllable jailbreaking of LLMs. Thanks to the controllability, COLD-Attack enables new jailbreak scenarios that are hard to detect🧐: 1⃣revising a user query adversarially with minimal paraphrasing 2⃣inserting stealthy
25K
Lianhui Qin
@Lianhuiq
Oct 13, 2020
🤔How can a pre-trained left-to-right LM do nonmonotonic reasoning, that requires conditioning on a future constraint⏲️? Our #emnlp2020 paper introduces DELOREAN🚘: an unsupervised backpropagation-based decoding strategy that considers both past context and future constraints.
Lianhui Qin
@Lianhuiq
Aug 21, 2025
Been quiet here for a bit, but had to share what I just tried with Mirage 2!! 📣📣📣 I dropped in a random ❄️Game of Thrones image and suddenly I could walk around inside it with just my keyboard, then typed 🧟‍♂️🧟‍♀️ “Night King + White Walkers” and they appeared 😱 The controls
00:00
20K
Lianhui Qin
@Lianhuiq
Mar 5, 2024
🧑‍🔬LLMs for complex Chemistry reasoning!🧪 Interestingly, we found LLMs (GPT-4) have already encoded lots of ⚗️Chemistry knowledge. 🤔What is really missing is a structured process to elicit the right knowledge, and use the knowledge to perform grounded reasoning. A very
Siru Ouyang
@Siru_Ouyang
Mar 5, 2024
🚀Announcing StructChem: A simple yet effective prompting strategy, unlocking the power of LLMs for complex chemistry reasoning. This task requires: - Extensive domain knowledge - Precise scientific computing - Compositional step-by-step reasoning Paper: arxiv.org/abs/2311.09656
25K
Lianhui Qin
@Lianhuiq
Apr 28, 2025
🚨💡New finding: LLMs struggle at multi-agent collaboration—not due to lack of knowledge, but because communicating plans via natural language across 👨‍👩‍👧‍👦agents is hard. 👩‍🍳🔨🛖In MINDcraft + MineCollab, agents must coordinate to cook, craft, and build in Minecraft. 📉Even small
00:26
Isadora White
@isadorcw
Apr 28, 2025
The real world is an embodied multi-agent system with natural language communication. What if we had a benchmark and platform to study those challenges? ⛏️Introducing MINDcraft and MineCollab, the 1st platform and benchmark for studying embodied multi-agent LLM collaboration!
16K
Lianhui Qin
@Lianhuiq
Oct 17, 2022
COLD decoding accepted by #NeurIPS2022 It enables generating arbitrarily constrained text with pretrained LMs, through continuous text approximation, energy-based modeling, and Langevin dynamics Checkout the latest version arxiv.org/abs/2202.11705 Code github.com/qkaren/COLD_de…
arxiv.org
COLD Decoding: Energy-based Constrained Text Generation with...
Many applications of text generation require incorporating different constraints to control the semantics or style of generated text. These constraints can be hard (e.g., ensuring certain keywords...