KAIST AI

Achieving generalizable manipulation is the north star for robotics learning, and while we’ve in the past seen incredible results on specific tasks using fine-tuned VLAs, this north star has remained elusive. Perhaps what is needed is a different approach. DreamZero proposes World Action models (WAMs), which jointly model both action and video in order to achieve state-of-the-art performance on benchmarks like MolmoSpaces and RoboArena. @SeonghyeonYe of @NVIDIARobotics joins us to talk about building a 14B parameter autoregressive diffusion model which achieves state-of-the-art generalization on real world tasks and on the best available benchmarks. Watch episode #68 of RoboPapers, with @micoolcho and @chris_j_paxton, now!

New Paper💡 Have you ever heard of grokking, a sudden transition from memorization to generalization? People have attributed grokking to weight decay, Fourier structure, optimization regimes, phase transitions, numerical effects… These can shape the training dynamics, but they don’t answer the core question: "what determines which representation the model learns, and why it generalizes?" We argue the key is intrinsic task symmetries. Paper: arxiv.org/pdf/2603.01968

Noisy context breaks many LLM reasoning runs, so this paper rewards models for using only relevant text. The authors build NoisyBench and show LLMs get misled by distractors, then train them to filter. Across 11 tasks, irrelevant documents, old chat history, and tricky lookalike passages cut accuracy by up to 80%. Context is the pile of text an LLM sees before answering, and distractors are parts that sound useful but are wrong. NoisyBench makes this problem measurable by taking normal benchmarks and adding random documents, irrelevant chats, and hard negatives, long passages that look relevant but are wrong. When the authors turn models into AI agents that call tools, the agent often trusts noisy tool output too much, and simple prompt tweaks or supervised fine tuning, training on labeled examples, does not fix it. This matters because real apps mix questions with messy search results and chat history, so a wrong clue can steer the whole answer. Their answer is Rationale Aware Reward (RARE), a training method where the model gets rewarded for using helpful context, so longer reasoning stays focused instead of chasing distractors. ---- Paper Link – arxiv. org/abs/2601.07226 Paper Title: "Lost in the Noise: How Reasoning Models Fail with Contextual Distractors"

🌎Real-world knowledge evolves constantly and emerges incrementally. Can LLMs adapt to new information on the fly? 🤯Frontier models and agentic approaches all struggle, missing when to update the fact, or getting distracted by irrelevant information. We introduce ✨OAKS✨, a benchmark for evaluating models’ online adaptation to streaming, continually updating knowledge.

🚀 Code Released! EgoX: Egocentric Video Generation from a Single Exocentric Video ✨ 💻 The code is now available! You can generate egocentric videos just like in the demo, or use EgoX to transform your own exocentric videos into egocentric ones.

Excited to present Directional Textual Inversion (DTI) at #ICLR 2026! Textual Inversion (TI) often fails on complex prompts. We fix this by optimizing strictly on the hypersphere. Work by our team at @KAIST_AI . arXiv: arxiv.org/abs/2512.13672 Project: kunheek.github.io/dti

Hello, world! We welcome you to explore our Seoul World Model

Align LLMs with the preference dataset ONLY with 💡ORPO💡 We introduce ORPO, alignment without reference model & SFT! With awesome dataset from @argilla_io + Mistral(7B) + ORPO, we present 🌟Mistral-ORPO-β🌟 🧵 👉 AlpacaEval 2.0: 12.20% 👉 IFEval: 66.19% 👉 MT-Bench: 7.32

Very proud of the undergraduates from the #kaist #computerscience department who have been studying the behavior of #LLMs in the Korean Language. Congratulations! w/ @HaneulYoo13 @j6mes @aliceoh

If you are interested in medical NLP/AI, our lab is hosting one of the shared tasks at NAACL-ClinicalNLP 2024 on reliable text-to-SQL on electronic health records. #EHRSQL. Please check out the link for more information. sites.google.com/view/ehrsql-20….

Pleased to see Prof Jong C. Park from #KAIST Computer Science department present the opening keynote at #EMNLP2023 this year sharing insights from more than 20 years of development of human centric Natural Language Processing nlpcl.kaist.ac.kr

Our paper "Flooding with Absorption: An Efficient Protocol for Heterogeneous Bandits over Complex Networks" won the Best Student Paper at OPODIS 2023! 🌐📚 Congratulations to Junghyun Lee, Laura Schmid, and Se-Young Yun! 🔗 Read the paper: arxiv.org/abs/2303.05445

Taehyung and Jihoon from Data Mining Lab received the IEEE ICDM Best Student Paper Runner-up Award. Congrats! arxiv.org/abs/2309.10310 icdm.zhonghuapu.com/Awards/BestPap… kaistdata.github.io

🚀 Big news from KAIST AI! We're now on LinkedIn! 🌟 🔗 Follow us: linkedin.com/company/kaist-…

[2/2] Authors: Geewook Kim (@GeewookKim), Hodong Lee, Daehee Kim, Haeji Jung, Sanghee Park, Yoonsik Kim, Sangdoo Yun, Taeho Kil, Bado Lee, Seunghyun Park Url: arxiv.org/abs/2305.15080

[1/2] Visually-Situated Natural Language Understanding with Contrastive Reading Model and Frozen Large Language Models (long paper / KAIST LK Lab & NAVER Cloud AI & Korea University & NAVER AI Lab)

⭐️Congratulations to everyone at KAIST AI who has a total of 11 papers accepted to #EMNLP2023! Check them here 📷🧵 #heading=h.csdjiropd9wf" target="_blank" rel="nofollow noopener">docs.google.com/document/d/1YB…

[2/2] Authors: Seonghyeon Ye (@SeonghyeonYe), Joel Jang (@jang_yoel), Doyoung Kim (@Doe_Young_Kim), Yongrae Jo (@dreamgonfly), Minjoon Seo (@seo_minjoon) Url: arxiv.org/abs/2210.03029

[1/2] Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt (findings / KAIST LK Lab)

[2/2] Authors: Seungone Kim (@seungonekim), Se June Joo (@joocjun), Doyoung Kim (@Doe_Young_Kim), Joel Jang (@jang_yoel), Seonghyeon Ye (@SeonghyeonYe), Jamin Shin (@jay_shin), Minjoon Seo (@seo_minjoon) Url: arxiv.org/abs/2305.14045

[1/2] The CoT Collection: Improving Zero-shot and Few-shot Learning of Language Models via Chain-of-Thought Fine-Tuning (long paper / KAIST LK Lab & NAVER AI Lab & University of Washington)

An Integrated Search System for Korea Weather Data (Industry Track / KAIST LK Lab & Korea University) Authors: Jinkyung Jo (@Jinkyung_Jo), Dayeon Ki (@zoeykii), Soyoung Yoon (@soyoung971030), and Minjoon Seo (@seo_minjoon) Url: drive.google.com/file/d/1uO-Nfi…

ROAST: Robustifying Language Models via Adversarial Perturbation with Selective Training (findings / KAIST ALIN Lab & Meta AI) Authors: Jaehyung Kim (@jhkim940331), Yuning Mao, Rui Hou, Hanchao Yu, Davis Liang, Pascale Fung, Qifan Wang, Fuli Feng, Lifu Huang, Madian Khabsa