Taeyun Roh

AI Researcher @ Korea Univ

[Taeyun Roh @nrbsld237]

I am attending #ACL2026 and #ICML workshop in person!
Especially excited to present my work CLAG at ACL which is an memory system optimized for SLMs.
Always glad to chat with fellow researchers interested in multimodal agents,agentic training and anything else!
See y'all soon!

[Taeyun Roh @nrbsld237]

Excited to share that i will be joining @Samsung_RA as a Research Intern.

I will be working on memory system for AI Agents.

Excited to meet new friends in the Bay area.

Feel free to reach out

#Samsung #LifeatSRA #SRA

[Yein Park @parkyein522]

I’ll present our team’s recent work at #ICML2026 #SD4H!

📄 Distilling Expert-level Planning for Real-world Diabetes Prescribing

We train medical LLMs to follow expert prescribing workflows using EMR context, insurance-aware search, and RL.

🗓️ Jul 11, 9:50–10:50
📍Hall D2, COEX

[Taeyun Roh @nrbsld237]

@Jonathan_Ivey8 Nice post! I have also finished my first year of graduate school, and it feels nice to know that the worries I had are not mine alone. Let's try hard to enjoy our PhD journey on our own way!

[Taeyun Roh @nrbsld237]

@CSVisionPapers Nice work

[Taeyun Roh @nrbsld237]

@harveyiyun Nice work!
I have been doing a similar research in Vqa especially in mcqa domains. Models chooses choices without seeing images. In this situation substracting text only distribution helps. Reading my work could potentially add insight to you amazing work. https://t.co/P7kETMHqPL

[Taeyun Roh @nrbsld237]

@ShuozheL @COLM_conf Couldnt agree more

[Taeyun Roh @nrbsld237]

@ahatamiz1 I can review DM me if needed

[Taeyun Roh @nrbsld237]

@penzant I can review

[Taeyun Roh @nrbsld237]

Paper accepted at #acl2026 (findings)

Yeaaah!!!

[Taeyun Roh @nrbsld237]

@windx0303 Why does this it fundamentally happen??

[Taeyun Roh @nrbsld237]

@SciFi Nice work by the authors!

[Taeyun Roh @nrbsld237]

Intoducing SCICON, contrastive decoding method designed to tackle choice-induced prior bias in scientific MCQA. By discounting text-only preferences,we ensure that models stay grounded in visual evidence rather than relying on text shortcuts.
​📄 Paper: https://t.co/L77yhqlfwj

[Taeyun Roh @nrbsld237]

@HEI Nice paper

[Taeyun Roh @nrbsld237]

CLAG: ​A clustering-based agentic memory framework that organizes data into "semantic neighborhoods."
​🔹 SLM Routing: Auto-profiles & clusters memories.
🔹 Local Evolution: Reduces interference
🔹 2-Stage Retrieval: Faster, cleaner search space. https://t.co/IS253qkIZs