Dongkeun Yoon
@dongkeun_yoon
Followers
464
Following
493
Media
17
Statuses
170
PhD student @kaist_ai. Researching multilinguality in LLMs.
Joined March 2022
🎉🎉 Super pround to share that our work is accepted to #NeurIPS2025 !! Huge thanks to all the co-authors. 👏👏
🙁 LLMs are overconfident even when they are dead wrong. 🧐 What about reasoning models? Can they actually tell us “My answer is only 60% likely to be correct”? ❗Our paper suggests that they can! Through extensive analysis, we investigate what enables this emergent ability.
2
3
31
We are gathering problems to build a challenging math benchmark (collaboration between @AiEleuther and @withmsit). The compensation per problem is up to ~$3,623 and the due date is Nov 10th! https://t.co/TdUG5xvTr2
We are announcing an opportunity for paid question writers to contribute to a new PhD-level math benchmark. Accepted contributors will be paid per question and will be invited to be authors on the resulting dataset paper. Check out the link below for more information!
0
2
22
Every LM needs a way of encoding data, and any choice of encoding is a design choice. When using bytes, you borrow choices from the makers of UTF8, and there’s generally no reason to believe that the most common encoding on the internet is also the best one for language modeling.
I have a new blog post about the so-called “tokenizer-free” approach to language modeling and why it’s not tokenizer-free at all. I also talk about why people hate tokenizers so much!
2
8
91
If you're at ACL, join us for our tutorial on Synthetic Data in the Era of LLMs with @vijaytarian @xiangyue96 @yizhongwyz @gneubig!! 🕑 2pm - 5:30pm 📍 Hall B
4
14
121
🥳Excited to share that I’ll be joining @unccs as postdoc this fall. Looking forward to work with @mohitban47 & amazing students at @unc_ai_group. I'll continue working on retrieval, aligning knowledge modules with LLM's parametric knowledge, and expanding to various modalities.
20
31
161
🚀 Tower+: our latest model in the Tower family — sets a new standard for open-weight multilingual models! We show how to go beyond sentence-level translation, striking a balance between translation quality and general multilingual capabilities. 1/5 https://t.co/WKQapk31c0
1
9
25
Check out the latest iteration of Tower models, Tower+. Ideal for translation tasks and beyond, and available at three different scales: 2B, 9B, 72B. All available on huggingface: https://t.co/XWJqTeht7R Kudos to everyone involved!
huggingface.co
🚀 Tower+: our latest model in the Tower family — sets a new standard for open-weight multilingual models! We show how to go beyond sentence-level translation, striking a balance between translation quality and general multilingual capabilities. 1/5 https://t.co/WKQapk31c0
0
1
10
🚨 Want models to better utilize and ground on the provided knowledge? We introduce Context-INformed Grounding Supervision (CINGS)! Training LLM with CINGS significantly boosts grounding abilities in both text and vision-language models compared to standard instruction tuning.
2
47
125
🚨 New Paper 🧵 How effectively do reasoning models reevaluate their thought? We find that: - Models excel at identifying unhelpful thoughts but struggle to recover from them - Smaller models can be more robust - Self-reevaluation ability is far from true meta-cognitive awareness
3
27
131
🚨 Excited to share that our paper was accepted to #ACL2025 Findings 🎉 "When Should Dense Retrievers Be Updated in Evolving Corpora? Detecting Out-of-Distribution Corpora Using GradNormIR" Huge thanks to my amazing collaborators! 🙌 @jinyoung__kim @ohmyksh We propose
arxiv.org
Dense retrievers encode texts into embeddings to efficiently retrieve relevant documents from large databases in response to user queries. However, real-world corpora continually evolve, leading...
0
7
38
[1/6] Ever wondered why Direct Preference Optimization is so effective for aligning LLMs? 🤔 Our new paper dives deep into the theory behind DPO's success, through the lens of information gain. Paper: "Differential Information: An Information-Theoretic Perspective on Preference
5
22
67
Check out our latest work on self-improving LLMs, where we try to see if LLMs can utilize their internal self consistency as a reward signal to bootstrap itself using RL. TL;DR: it can, to some extent, but then ends up reward hacking the self-consistency objective. We try to see
4
27
142
🚨 New Paper co-led with @bkjeon1211 🚨 Q. Can we adapt Language Models, trained to predict next token, to reason in sentence-level? I think LMs operating in higher-level abstraction would be a promising path towards advancing its reasoning, and I am excited to share our
4
43
168
New preprint 📄 (with @jinho___park ) Can neural nets really reason compositionally, or just match patterns? We present the Coverage Principle: a data-centric framework that predicts when pattern-matching models will generalize (validated on Transformers). 🧵👇
2
31
126
Imagine you’re collaborating with an AI co-scientist: you ask it to proofread your manuscript and flag any errors. Which LLM would you choose? 🤔 We evaluated the new Claude 4 models on SPOT. It looks like o3 is still the best model for this.
2
5
8
❓What if your RAG didn’t need a separate retrieval model at all? We present 🧊FREESON, a new framework for retriever-FREE retrieval-augmented reasoning. With FREESON, a single LRM acts as both generator and retriever, shifting the focus from seq2seq matching to locating
1
5
29
@dongkeun_yoon Congrats to the team for this fantastic work! Had a chance to try the code on my reasoning VLM and found consistent results. https://t.co/UcwzXoxuZM
Tried out the code using SpaceThinker-Qwen2.5-VL-3B Plots indicate steady increase in accuracy and confidence while reducing calibration error as CoT increases Fitting TriviaQA with linear regression, slope: 0.042, -0.034, -0.02, 0.011 all statsig
1
1
1
🙁 LLMs are overconfident even when they are dead wrong. 🧐 What about reasoning models? Can they actually tell us “My answer is only 60% likely to be correct”? ❗Our paper suggests that they can! Through extensive analysis, we investigate what enables this emergent ability.
9
49
302
[CL] Reasoning Models Better Express Their Confidence D Yoon, S Kim, S Yang, S Kim... [KAIST & CMU & UCL] (2025) https://t.co/w7bwk1voj8
0
5
19
