KAIST Machine Learning and Artificial Intelligence Group
Daejeon, South Korea
Joined June 2020
- Tweets139
- Following19
- Followers458
- Likes21
MLAI_KAIST retweeted
🚀Excited to share that our Mol-LLaMA is accepted at #NeurIPS2025!!
📄Paper: arxiv.org/abs/2502.13449
🖥️Project Page: mol-llama.github.io/
See you in San Diego!
🚨Excited to introduce Mol-LLaMA, a large molecular language model that grasps the general knowledge centered on molecules, positioning it as a general-purpose assistant for molecular analysis.
Paper: arxiv.org/abs/2502.13449
Project Page: mol-llama.github.io/
🧵below
MLAI_KAIST retweeted
🤗Our paper, "Delta Attention," has been accepted to NIPS 2025. The delta correction is beneficial for anyone interested in improving the accuracy of sparse attention. Wonderful work, @TheOneJeffrey
Noice! Our paper "Delta Attention: Fast and Accurate Sparse Attention Inference by Delta Correction" has been accepted to NeurIPS 2025! See you in San Diego (See part 2 of post for breakdown of our work)
arxiv.org/pdf/2505.11254
MLAI_KAIST retweeted
Our Agent Distillation paper is accepted at #NeurIPS2025 Spotlight! 🚀
Turn your small LM into a strong agent 💪
Code: github.com/Nardien/agent-dis…
🚨 New preprint!
Can small language models (sLMs) solve complex problems like LLMs?
We show how to go beyond cloning reasoning—to distill tool-using agent behavior into sLMs as tiny as 0.5B.
Meet Agent Distillation:
📄 huggingface.co/papers/2505.1…
Here's the details 🧵👇:
MLAI_KAIST retweeted
✨Excited to share our new work "Robust Molecular Property Prediction via Densifying Scarce Labeled Data", accepted by GenBio ICML Workshop 2025!
arxiv : arxiv.org/abs/2506.11877
Special thanks to all the collaborators @TheOneJeffrey @andreisbru and our advisor, @SungJuHwang1
MLAI_KAIST retweeted
🚨 New preprint!
Can small language models (sLMs) solve complex problems like LLMs?
We show how to go beyond cloning reasoning—to distill tool-using agent behavior into sLMs as tiny as 0.5B.
Meet Agent Distillation:
📄 huggingface.co/papers/2505.1…
Here's the details 🧵👇:
MLAI_KAIST retweeted
So excited to share that five papers have been accepted to #ACL2025 🎉
Huge thanks to all my amazing collaborators. I am especially grateful that all the internship projects (which I have worked on during my PhD journey) have all found their way to publication. Its rewarding. 😊
MLAI_KAIST retweeted
Introducing 🌌UniversalRAG — a novel RAG framework that adaptively retrieves from multimodal corpora (📕text, 🖼image, 🎥video) at the right level of granularity, from short snippets to long content.
Huge thanks to @omarsar0 for sharing our work!
Paper: arxiv.org/abs/2504.20734
MLAI_KAIST retweeted
There is a HUGE problem with sparse attention! We found it causes a misalignment of queries and keys, so even if you add a dense decode/generation phase, the sparsely encoded context can be forgotten. But wait... We can fix it with a simple correction! huggingface.co/papers/2505.1…
MLAI_KAIST retweeted
🤔Can you really trust what your AI creates? Meet Silent Branding Attack—a sneaky, no-trigger-needed method to secretly poison text-to-image models!
👀 Test yourself: Can you spot which images below have been silently branded?
MLAI_KAIST retweeted
📣 Excited to introduce Sketch-of-Thought (SoT)
We've drawn inspiration from human cognition to make #LLM reasoning more efficient, reducing token usage by an avg. of 75% while having comparable accuracy to traditional CoT approaches.
🔗Paper: arxiv.org/abs/2503.05179
MLAI_KAIST retweeted
🚨Excited to introduce Mol-LLaMA, a large molecular language model that grasps the general knowledge centered on molecules, positioning it as a general-purpose assistant for molecular analysis.
Paper: arxiv.org/abs/2502.13449
Project Page: mol-llama.github.io/
🧵below
MLAI_KAIST retweeted
Excited to introduce VideoRAG📽️, which extends the current landscape of RAG systems (that largely focus on text) by holistically leveraging the video corpus, utilizing both textual and visual features.
Paper: arxiv.org/abs/2501.05874
@SoyeongJeong97 @kangsan_kim_ @SungJuHwang1
MLAI_KAIST retweeted
🚨 I am on the 2025 faculty job market! 🚨(jaehong31.github.io)
I develop reliable and lifelong embodied AI systems 🔥 that continually evolve capabilities through safe and robust interactions with an ever-changing multimodal world, focusing on: 👇
▶️ Scalable and Multimodal Continual Learning
▶️ OOD Adaptation with Post-Training
▶️ Trustworthy Reasoning + Generation
I’m currently a postdoc with @mohitban47 at @uncnlp and did my Ph.D. with @SungJuHwang1 at #KAIST (@MLAI_KAIST @kaist_ai).
Also, I’ll be in Vancouver to attend #NeurIPS2024. Please reach out in person or via email!
MLAI_KAIST retweeted
🚨Introducing MuDI: model-agnostic Multi-Subject Personalization Method!
Single-subject personalization methods (a.k.a. DreamBooth) usually work well, but they struggle to generate multiple subjects simultaneously.
▶️ We introduce MuDI, a simple framework for multi-subject personalization with identity decoupling, without any attention map control or inversion method!
proj: mudi-t2i.github.io/
🧵👇
MLAI_KAIST retweeted
🚨 New paper! We propose HarmAug, a data augmentation method that distills large safety guard models into a 435M-parameter model. It detects harmful prompts with LLMs, cutting computational costs by 75% while matching performance of 7B+ models!
MLAI_KAIST retweeted
📢😎Introducing our novel attention mechanism, HiP!
Enjoy 3x+ large-scale, long-context LLM decoding throughput just by replacing a single function⚡
HiP is a sub-quadratic, training-free sparse attention mechanism with hierarchical attention pruning.
📝arxiv.org/abs/2406.09827
MLAI_KAIST retweeted
We are excited to share that our efficient linear attention method, SEA: Sparse Linear Attention with Estimated Attention Mask, has been accepted to #ICLR2024.
We are going to present the poster on May 8th in Vienna. See you at #ICLR2024!
Paper: arxiv.org/abs/2310.01777v2
MLAI_KAIST retweeted
Happy to share that our paper is accepted to @naaclmeeting. This is done during my internship at Apple with Jianpeng Cheng, Joris Driesen, Alexandru Coca and @andersjo. arxiv.org/abs/2402.13043
MLAI_KAIST retweeted
Exhausted to collect well-aligned text-image pairs for T2I generation?🤔
📌Check out SELMA - a novel paradigm to improve the faithfulness of T2I models with automatically generated skill-specific text-image pairs.
We leverage skill-specific expert modules and expert merging strategy to facilitate training on multiple skills to T2I models.
Our findings:
▶️SELMA enhances SD models' T2I generation performance w/o human-annotated datasets.
▶️SELMA learns multiple skill-specific experts effectively while mitigating task interference.
▶️SELMA shows promising weak-to-strong generalization in T2I.
Can we teach multiple skills to a text-to-image (T2I) model (w/o expensive annotations), while minimizing knowledge conflicts between skills? 🤔
👉SELMA improves T2I models by fine-tuning on automatically generated multi-skill image-text datasets, with skill-specific LoRA expert learning & merging.
-- On SD models, SELMA boosts 5 metrics (TIFA/DSG/PickScore/ImageReward/HPS) + human eval.
-- Finetuning on auto-generated data matches the performance of finetuning on GT data.
-- Weak-to-strong generalization (SDv2 images can help SDXL finetuning).
arxiv.org/abs/2403.06952
🧵
