Woongyeong Yeo

38 posts

Woongyeong Yeo

Woongyeong Yeo

@wgcyeo

M.S. Student @KAIST_AI

Katılım Eylül 2024
79 Takip Edilen50 Takipçiler
Sabitlenmiş Tweet
Woongyeong Yeo
Woongyeong Yeo@wgcyeo·
🔍 Is a single embedding space really enough for multimodal RAG? Excited to share that UniversalRAG has been accepted to the #ACL2026 main conference! 🥳 We introduce the first any-to-any multimodal RAG framework, enabling retrieval across diverse modalities and granularities.
Woongyeong Yeo tweet media
English
1
11
27
2.9K
Woongyeong Yeo retweetledi
Jinheon Baek
Jinheon Baek@jinheonbaek·
Privacy for LLM agents is not simply about hiding information. A useful assistant should know what to use, what to ignore, and what is appropriate to disclose. And, we propose a complementary self-distillation framework for it. Excited to have co-advised this work!
Woongyeong Yeo@wgcyeo

📢 New preprint out on contextual integrity (CI) and a new Product-of-Experts (PoE) view of self-distillation! Introducing SelfCI, a novel self-distillation framework that operationalizes CI by optimizing for the intersection of task utility and minimal disclosure. 🧵👇

English
1
1
12
918
Woongyeong Yeo
Woongyeong Yeo@wgcyeo·
(3/n) 🔥 Consistently outperforms online RL algorithms like GRPO without costly external supervision. ⏱️ Cuts GPU wall-clock training time per step by nearly half compared to online RL. 📈 Scales robustly across model families and excels in out-of-domain agentic workflows.
Woongyeong Yeo tweet media
English
1
0
2
137
Woongyeong Yeo
Woongyeong Yeo@wgcyeo·
📢 New preprint out on contextual integrity (CI) and a new Product-of-Experts (PoE) view of self-distillation! Introducing SelfCI, a novel self-distillation framework that operationalizes CI by optimizing for the intersection of task utility and minimal disclosure. 🧵👇
Woongyeong Yeo tweet media
English
1
10
27
2.4K
Woongyeong Yeo retweetledi
Yumin Choi
Yumin Choi@yumin_choi_·
Can LLM agents build memory before seeing any user task? Memory is usually built from human tasks or deployment interactions. New tool environments often have neither, creating cold-start gap. Introducing PREPING: building agent memory without tasks. dozi01.github.io/preping-projec…
Yumin Choi tweet media
English
5
9
21
2.4K
Woongyeong Yeo retweetledi
Kangsan Kim
Kangsan Kim@kangsan_kim_·
💻 🧠 Does SWE memory help ML programming tasks in coding agents? Super excited to introduce 𝗠𝗲𝗺𝗼𝗿𝘆 𝗧𝗿𝗮𝗻𝘀𝗳𝗲𝗿 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴, a framework that leverages cross-domain coding memory, enabling agents to reuse experiences beyond task boundaries and improve memory utilization. MTL improves coding agent by 𝟯.𝟳% 𝗼𝗻 𝗮𝘃𝗲𝗿𝗮𝗴𝗲 over a zero-shot baseline across six benchmarks. 💡Key Insights 1. 𝐌𝐞𝐦𝐨𝐫𝐲 𝐓𝐫𝐚𝐧𝐬𝐟𝐞𝐫 𝐖𝐨𝐫𝐤𝐬! Memory Transfer Learning significantly improves coding agent performance and outperforms self-evolving methods in effectiveness and efficiency. 2. 𝐓𝐫𝐚𝐧𝐬𝐟𝐞𝐫𝐚𝐛𝐥𝐞 𝐤𝐧𝐨𝐰𝐥𝐞𝐝𝐠𝐞 𝐢𝐬 𝐦𝐨𝐬𝐭𝐥𝐲 𝐦𝐞𝐭𝐚-𝐦𝐞𝐦𝐨𝐫𝐲 Transferable knowledge exists across distinct task types, and its primary form is meta-memory encoding procedural and behavioral guidance, not domain-specific knowledge 3. 𝐀𝐛𝐬𝐭𝐫𝐚𝐜𝐭𝐢𝐨𝐧 𝐢𝐬 𝐚 𝐤𝐞𝐲 𝐝𝐫𝐢𝐯𝐞𝐫 𝐨𝐟 𝐞𝐟𝐟𝐞𝐜𝐭𝐢𝐯𝐞 𝐭𝐫𝐚𝐧𝐬𝐟𝐞𝐫 More abstract and generalized memory representations yield higher transfer effectiveness by avoiding brittle implementation anchoring. Project Page: lnkd.in/gHp8VPrb @KAIST_AI @nyuniversity
Kangsan Kim tweet media
English
1
28
88
5.5K
Woongyeong Yeo retweetledi
Soyeong Jeong
Soyeong Jeong@SoyeongJeong97·
Super excited to share that one of my favorite papers, “When Thoughts Meet Facts: Reusable Reasoning for Long-Context LMs,” has been accepted to #ACL2026 Findings! 🎉
Soyeong Jeong@SoyeongJeong97

🧠📚 When thoughts meet facts. How can LLMs reuse their thoughts to reason better over long contexts even without direct retrieval? Reusable reasoning templates + iterative refinement → better factual multi-hop reasoning 🧩 📄 arxiv.org/abs/2510.07499

English
1
12
42
2.8K
Woongyeong Yeo
Woongyeong Yeo@wgcyeo·
Happy to share that our latest work on a video memory agent has been selected as a Highlight at #CVPR2026 ✨ Looking forward to discussing memory systems and video reasoning in Denver. Huge thanks again to my co-authors for making this possible! 🌐 worldmm.github.io
Woongyeong Yeo@wgcyeo

🚨 Excited to introduce 🌏 WorldMM, a novel agent that constructs & leverages multimodal and multi-scale memories for long video reasoning. Day- or week-long videos often exceed context limits, and textual abstractions fail to capture crucial visual details. WorldMM tackles these with: ➡️ Multimodal Memory combining multi-scale episodic graphs, evolving semantic knowledge, and fine-grained visual details ➡️ Adaptive Retrieval that decides the required memory, query, and timescale by reasoning over retrieval history ➡️ Comprehensive Evaluation with an average +8.4% over prior SOTA on hour- to week-long benchmarks 🧵👇

English
1
2
23
2.2K
Woongyeong Yeo retweetledi
Jinheon Baek
Jinheon Baek@jinheonbaek·
Excited to see UniversalRAG accepted to #ACL2026! 🎉 Proud to be part of this work on any-to-any multimodal RAG with modality-aware routing to bridge the modality gap.
Woongyeong Yeo@wgcyeo

🔍 Is a single embedding space really enough for multimodal RAG? Excited to share that UniversalRAG has been accepted to the #ACL2026 main conference! 🥳 We introduce the first any-to-any multimodal RAG framework, enabling retrieval across diverse modalities and granularities.

English
0
5
40
2.4K
Woongyeong Yeo
Woongyeong Yeo@wgcyeo·
While recent embedding models perform well on homogeneous data, they often struggle with heterogeneous corpora due to the modality gap. To address this, we introduce modality-aware routing, allowing UniversalRAG to achieve both effective and efficient multimodal retrieval.
Woongyeong Yeo tweet media
English
1
0
1
184
Woongyeong Yeo
Woongyeong Yeo@wgcyeo·
🔍 Is a single embedding space really enough for multimodal RAG? Excited to share that UniversalRAG has been accepted to the #ACL2026 main conference! 🥳 We introduce the first any-to-any multimodal RAG framework, enabling retrieval across diverse modalities and granularities.
Woongyeong Yeo tweet media
English
1
11
27
2.9K
Woongyeong Yeo retweetledi
Kangsan Kim
Kangsan Kim@kangsan_kim_·
🎉 Happy to share that UniversalRAG has been accepted to the #ACL2026 main conference! We introduce the first any-to-any multimodal RAG framework that integrates diverse modalities and granularities into a unified workflow via modality-aware routing. 🔗 Link: universalrag.github.io
Kangsan Kim tweet media
English
0
10
32
1.3K
Woongyeong Yeo retweetledi
AK
AK@_akhaliq·
MA-EgoQA Question Answering over Egocentric Videos from Multiple Embodied Agents paper: huggingface.co/papers/2603.09…
AK tweet media
English
4
6
17
4.8K
Woongyeong Yeo retweetledi
Kangsan Kim
Kangsan Kim@kangsan_kim_·
Excited to share that one of our papers was accepted to CVPR 2026 and another to CVPR 2026 Findings! 🌏WorldMM: Dynamic Multimodal Memory Agent for Long Video Reasoning (CVPR 2026) worldmm.github.io 🛡️HoliSafe: Holistic Safety Benchmarking and Modeling for Vision-Language Model (CVPR 2026 Findings) youngwanlee.github.io/holisafe/
Kangsan Kim tweet media
English
0
8
64
4.4K
Woongyeong Yeo
Woongyeong Yeo@wgcyeo·
Marking my first first-authored conference paper, and hope to continue making meaningful steps forward in this field. Thanks again to my best research mate @kangsan_kim_, and sincere appreciation to @jaeh0ng_yoon and @SungJuHwang1 for the insightful advice. See you in Denver 🇺🇸
English
0
1
4
597
Woongyeong Yeo
Woongyeong Yeo@wgcyeo·
🥳 Excited to share that WorldMM is accepted to #CVPR2026! We introduce a multimodal memory agent with adaptive retrieval over multi-scale multimodal memories, integrating both textual and visual information, achieving SOTA performance on long-video reasoning.
Woongyeong Yeo@wgcyeo

🚨 Excited to introduce 🌏 WorldMM, a novel agent that constructs & leverages multimodal and multi-scale memories for long video reasoning. Day- or week-long videos often exceed context limits, and textual abstractions fail to capture crucial visual details. WorldMM tackles these with: ➡️ Multimodal Memory combining multi-scale episodic graphs, evolving semantic knowledge, and fine-grained visual details ➡️ Adaptive Retrieval that decides the required memory, query, and timescale by reasoning over retrieval history ➡️ Comprehensive Evaluation with an average +8.4% over prior SOTA on hour- to week-long benchmarks 🧵👇

English
1
9
54
6.3K

Keşfet

@bahahdlax @seanie_12 @yumin_choi_ @pwnhyo @kangsan_kim_ @jinheonbaek @coallaoh @SungJuHwang1