Runtao Liu

a Ph.D student in HKUST

[Runtao Liu @rleobest]

🎬 LongVideoAgent — multi-agent reasoning for hour-long video QA

🤝 Multi-agent pipeline: 🧠 Master LLM plans ≤K steps, 🎯 GroundingAgent maps subtitles ➜ relevant clips, 👀 VisionAgent queries those clips ➜ fine-grained evidence.
🎓 GRPO RL improves efficiency & accuracy
📺 New benchmarks: LongTVQA / LongTVQA+

🔗 arXiv https://t.co/97nnCzbaoB
🌐 Project https://t.co/M9YXc7BtZJ
💻 Code https://t.co/Dj0KDbsAdM
🤗 HF https://t.co/yyw3pX0DsD

Happy to introduce our project 🎉 Dataset will be released in 1 week ⏳ and code in 4 weeks 💻🚀

#VideoQA #Agents #LLM

[Letian Wang @Letian_Wang_6]

Can video generation models do for vision what LLMs did for language?

Introducing GenCeption from @GoogleDeepMind: one feed-forward video model for various vision tasks — SOTA, data-efficient, and emerging behaviors (ECCV 2026)

🌐 https://t.co/3rwVgTolZz

(1/8)

[Yueqi Xie @XieYueqi]

🎊Excited to share that our research on validating AI-generated social science data is now online at @PNASNews !

Check it out: https://t.co/ljqoTBVBxB
Great thanks to my collaborators!

💡In this work, we highlight population-level statistical realism as a core criterion.

[Yi R. (May) Fung @ICML✈️ @May_F1_]

Our research has been well-represented at ICLR'26 🇧🇷, with the following papers on reasoning and agentic paradigm improvements. 🎉

1️⃣Dancing in Chains: Strategic Persuasion in Academic Rebuttal via Theory of Mind
2️⃣Diversity-enhanced reasoning for subjective questions
3️⃣Lean4Physics: Comprehensive Reasoning Framework for College-level Physics in Lean4
4️⃣Webwatcher: Breaking new frontier of vision-language deep research agent
5️⃣Towards Self-Evolving Benchmarks: Synthesizing Agent Trajectories via Test-Time Exploration under Validate-by-Reproduce Paradigm

Check it out further offline in case you missed the poster session earlier~ 🍻

[renjie pi @RenjiePi]

Introducing Nemotron-Terminal: a systematic data engineering pipeline for scaling LLM Terminal Agents.

We bridge the gap between open models and proprietary models with a fully open synthetic-to-real trajectory pipeline.

🤯The payoff: SFT on our Nemotron-Terminal-Corpus boosts Qwen3-32B from 3.4% → 27.4% on Terminal-Bench 2.0 (+24.0), rivaling models multiple its size.

What makes it work?

🌟Terminal-Task-Gen: A lightweight data curation pipeline that seamlessly combines the adaptation of existing datasets with robust synthetic task construction.
🌟Nemotron-Terminal-Corpus: A massive, open-source dataset covering diverse terminal interactions, which contains explicit planning and execution traces for complex long-horizon tasks.

And we’re releasing everything:

📦 Nemotron-Terminal-Corpus (Large-scale dataset)
🤖 Nemotron-Terminal models (8B, 14B, 32B)

Paper: https://t.co/ORIZ01sav1
HF Daily: https://t.co/nSH4hu7I5D
Models & Data: https://t.co/J1Zc22M95r

Our tech report just hit the #1 spot on Hugging Face Daily Papers!
We're also incredibly excited to see the open-source community putting our work to the test, with the Nemotron-Terminal-Corpus dataset currently trending at over 1,800 downloads and counting.
We can't wait to see what the community build with it!

[Zhenxing Mi @Mifucius1]

Excited to share our new paper "One4D: Unified 4D Generation and Reconstruction via Decoupled LoRA Control" on arXiv. With our novel designs of Unified Masked Conditioning (UMC) and Decoupled LoRA Control (DLC), One4D can seamlessly handle single-image-to-4D, sparse-frame-to-4D, and full-video-to-4D tasks in a unified model by generating high-quality RGB frames and accurate pointmaps.

Project page: https://t.co/QMGkHd7xWK
arXiv: https://t.co/v1urLAMkPw
Code to be released at: https://t.co/th4P7gdmdM
Huggingface Paper: https://t.co/bZYtt83OGY

Many thanks to our collaborators Prof. Dan Xu @danxuhk and Yuxin Wang.

[Runtao Liu @rleobest]

非常开心介绍我们最近的工作🎬 LongVideoAgent：面向“超长视频问答”的多智能体推理框架，在小时级超长视频中高效找证据并回答问题，并且应用了GRPO 强化学习进一步提升性能。

🔗 arXiv https://t.co/97nnCzbaoB
🌐 Project https://t.co/O44RpLMH1U
💻 Code https://t.co/Dj0KDbsAdM
🤗 HF https://t.co/yyw3pX0DsD

🤝 多智能体协作流程：🧠 主 LLM 负责规划（最多 ≤K 步）
🎯 GroundingAgent 先用字幕把问题定位到相关片段
👀 VisionAgent 再去“点查”这些片段的关键画面/细节，补足字幕缺失的证据
🎓 训练方式：用 GRPO 做强化学习，对齐“更省步骤、更准答案”的协作策略（同时提升效率与准确率）
📺 新基准：LongTVQA / LongTVQA+

[Zhenxing Mi @Mifucius1]

The source code and checkpoints of ThinkDiff are finally released! We also released a modified vLLM for embedding. Feel free to use them and raise an issue if you have any problem!

Github: https://t.co/PgvHdytanj
Checkpoints: https://t.co/NtPBk9Pkpb
vLLM: https://t.co/jbyiIfVhWH

[Teachable Machine @TeachableAI]

This new system called AlignGuard helps make sure AI image generators create pictures that are safe and appropriate. It uses a team of smaller models working together and suggests better training methods to keep things respectful. Tests show it works well, making images safer without sacrificing quality. The creators recommend using it after training an AI model as a final safety check before releasing it to the public.

#AI #ImageGeneration #Ethics

https://t.co/TOOTrzZ2SJ

#ArtificialIntelligence

[Zhenxing Mi @Mifucius1]

🔥🚀 Native image generation of Gemini is just on fire!

Our #ThinkDiff paper bridges vision-language models with diffusion models to unlock native image generation from VLMs. While not yet as perfect as #Gemini, it offers a glimpse of the groundbreaking potential!

🔥 Check out:
🔗 Project: https://t.co/QKjEHeiIBy
📄 Paper: https://t.co/M6JxdgcS3H | https://t.co/u5VvAP1k44
💻 Code to be released at: https://t.co/NVMi3x02gV

Check out our awesome test cases:

1. Generate a style consistent and logic-correct image by in-context reasoning:

[Qingping Yang @huie_qaqp]

🌟 Amazing work! SuperGPQA spans 285 graduate-level disciplines with 26,529 multiple-choice questions, each averaging 9 options per question. 📚💡 Congrats to @GeZhang86038849

🎉 Huge congrats to UTMath for being selected as one of the 15 reliable resources by SuperGPQA experts! From our UTMath test set, 401 questions were included, with 290 high-difficulty questions—that's nearly 1/5 of the toughest math problems in the dataset. 🎓✨ An evidence to the quality of UTMath! 🔢👏

Is verifying a correct answer through multiple-choice options truly sufficient for measuring graduate-level knowledge? 🤔 What about the risk of oversimplifying complexities, or the challenges of designing high-quality distractors? These are questions worth exploring. 📚🌟
#AI #Education #SuperGPQA #Mathematics #UTMath

[Runtao Liu @rleobest]

Great work! 🎉 Our UTMath also identified this but it skips manual transformation. Each problem has input-output pairs forming unit tests ✅ for correctness.
Tested on UTMath GPT4o scores 26% 📉, o1-mini 29% 📊, proving the test's difficulty. 🔥Pls check: https://t.co/EeFW9PThvl

[elvis @omarsar0]

o1-preview shows a whopping 30% reduction in accuracy when Putnam math problems are slightly variated.

Not sure about the reliability of the results but I think it's a good paper worth checking out to understand LLM robustness on complex math problems better.

A hot research topic right now is to understand better whether these models can "reason" robustly or if they are just relying on memorization.

Robustness is important as it's key to model reliability.

[Runtao Liu @rleobest]

@omarsar0 Great work! 🎉 Our UTMath also identified this but it skips manual transformation. Each problem has input-output pairs forming unit tests ✅ for correctness.
Tested on UTMath GPT4o scores 26% 📉, o1-mini 29% 📊, proving the test's difficulty. 🔥 Check: https://t.co/EeFW9PThvl

[Qingping Yang @huie_qaqp]

🚀Excited to introduce our work UTMath: Math Evaluation with UNIT TEST via RCoT🧠💡

🧠With an average of 68 test cases per problem, UTMath ensures that the model truly solves the problem.

💡We release our benchmark and training dataset (~70k).

Explore: https://t.co/Ga3QyzI487

[Runtao Liu @rleobest]

🔥Introducing Latent Guard from Oxford University and HKUST, a customizable T2I safety framework at test time. We release the largest T2I safety dataset with 50k~170k prompts of 723 unsafe concepts in 7 categories.
🌟Code&Data:https://t.co/4VTfui7Yu2
🌟Project: https://t.co/SOxw4rbQ51
🌟Paper:https://t.co/p6mtq49316