Pinned Tweet
🎉 Our paper has been accepted to #ICLR2026! 😆💖
This work was done during my internship at LG AI Research – Superintelligence Lab. As summarized in the project:
Deep research requires broad evidence coverage and reliable synthesis.
HybridDeepSearcher achieves both by parallel retrieval for breadth with sequential reasoning for depth, supporting scalable search.
🔗 Project page: hybriddeepsearcher.github.io
📄 OpenReview: openreview.net/forum?id=rXpTZ…
Huge thanks to my mentors and co-workers for their guidance and support throughout this project. We also plan to release related work soon. Stay tuned! 😊
This work was done during my internship at LG AI Research – Superintelligence Lab. As summarized in the project:
Deep research requires broad evidence coverage and reliable synthesis.
HybridDeepSearcher achieves both by parallel retrieval for breadth with sequential reasoning for depth, supporting scalable search.
🔗 Project page: hybriddeepsearcher.github.io
📄 OpenReview: openreview.net/forum?id=rXpTZ…
Huge thanks to my mentors and co-workers for their guidance and support throughout this project. We also plan to release related work soon. Stay tuned! 😊
4
5
58
5,739
Dayoon Ko retweeted
Check out this work co-led by @SophieLWang. We find that base models can match much of RL’s reasoning gains with the right first tokens, which we call 𝘵𝘰𝘬𝘦𝘯 𝘤𝘶𝘦𝘴. The effective cue differs across models and can be as simple as “.\n\nOkay,” or “To determine,” with no explicit instruction to reason. We dissect this phenomenon through a set of interventions on RL and the base-model training data itself. We also extend our analysis to a safety case study of refusal and compliance.
Token cues can arise from simple associations in reasoning-related training data, and we can even turn the word “chicken” into a reasoning cue through counterfactual mid-training runs.
Super excited to find a rich underlying phenomenology! Check out Sophie’s thread for the main findings. I’ll share a few other observations below that I found especially cool. 👇
Token cues can arise from simple associations in reasoning-related training data, and we can even turn the word “chicken” into a reasoning cue through counterfactual mid-training runs.
Super excited to find a rich underlying phenomenology! Check out Sophie’s thread for the main findings. I’ll share a few other observations below that I found especially cool. 👇
Can “chicken” make a base model reason? 🐔
Yes!
Why? With the right first tokens, base models can match RL in reasoning. We trace this to learned training data associations and use a simple data edit to make “chicken” elicit reasoning too.
blog: sophielwang.com/cues
4
19
74
22,219
Dayoon Ko retweeted
AI detectors read your sentences. We read your scientific reasoning and score its slop.
Every ICLR year, 2017–2025: papers with more our "scientific slop" got lower reviewer ratings and were rejected more often.
Calculate yours: yerimoh.github.io/scientific-slo…
Every ICLR year, 2017–2025: papers with more our "scientific slop" got lower reviewer ratings and were rejected more often.
Calculate yours: yerimoh.github.io/scientific-slo…
8
13
62
31,812
Dayoon Ko retweeted
AI can increasingly make progress on open problems like Navier-Stokes. Yet defining a new problem still relies on human research taste.
Introducing ScholarCatalyst, a benchmark built from AI researchers’ firsthand accounts of what inspired their work.
Introducing ScholarCatalyst, a benchmark built from AI researchers’ firsthand accounts of what inspired their work.
9
63
294
60,902
Dayoon Ko retweeted
Excited to introduce 🎶 EvoDuet: co-evolving web search and task solutions to push the state of the art across 8 scientific optimization tasks.
1
9
38
5,693
Dayoon Ko retweeted
Your emails, chats, docs, and tickets keep piling up, and the right ones get harder to find. 😡
🗺️ Introducing CorpusMap, which organizes a growing corpus around recurring entities, so agents follow connections instead of searching from scratch.
📈 Better answers, fewer tokens.
🗺️ Introducing CorpusMap, which organizes a growing corpus around recurring entities, so agents follow connections instead of searching from scratch.
📈 Better answers, fewer tokens.
1
17
48
3,531
Super excited to join @sewon__min's group at UC Berkeley as a Visiting Student Researcher from October to March!🐻 I've been a big fan of her work for years!
I'll be exploring on-device multimodal retrieval. If you're in the Bay area and into RAG or multimodal learning, let's thanks to my advisor @gunheekim at SNU for supporting this visit ✨
I'll be exploring on-device multimodal retrieval. If you're in the Bay area and into RAG or multimodal learning, let's thanks to my advisor @gunheekim at SNU for supporting this visit ✨
2
2
59
3,385
Dayoon Ko retweeted
Long-context LMs seem great at tasks like RULER, but what about tasks like "find every contradiction in the given corpus"?
- Performance suffers, especially as the input corpus scales (even with in-domain fine-tuning, not just zero-shot)
- It changes a lot of core modeling decisions, e.g., whether to use hybrid models, whether to use block sparse attention
TL;DR: if you are a model developer, include our CTC-Bench in your evals! ✨
- Performance suffers, especially as the input corpus scales (even with in-domain fine-tuning, not just zero-shot)
- It changes a lot of core modeling decisions, e.g., whether to use hybrid models, whether to use block sparse attention
TL;DR: if you are a model developer, include our CTC-Bench in your evals! ✨
Could long-context architectures “find all contradictions” in a science literature? Not yet! 🧵
We study a new class of "high-complexity” tasks whose difficulty scales quadratically with corpus size (as opposed to linearly), reversing common LCLM decisions! (block-sparse s…)
10
18
159
14,454
Dayoon Ko retweeted
Excited to share that our paper “MM-SCALE: Evaluating Evidence-Grounded Moral Judgment in Vision-Language Models” has been accepted to the @NeurIPSConf NeurIPS 2026 Evaluations & Datasets Track! 🎉
We study whether VLMs’ moral judgments are actually grounded in the visuans find to be important, not just whether they reach the same final verdict. An updated Preprint + code coming very soon. Stay tuned! 👀
This was the project that started the collaboration between @SNUVLLab and @cmuhcii, I am greatful that we finally are able to publish it! 🚀
✨Kudos to all my collaborators and mentors; Wesley Deng (@wes_deng ), Cheyon Jin, Matheus Maldaner (@matheusmaldaner), Jordan Wheeler, Jason Hong (@jas0nh0ng), Hong Shen (@hongshenus), Adam Perer (@adamperer ), Ken Holstein (@d19fe8), Motahhare Eslami (@meslami2) and my advisor, Gunhee Kim (@gunheekim).
We study whether VLMs’ moral judgments are actually grounded in the visuans find to be important, not just whether they reach the same final verdict. An updated Preprint + code coming very soon. Stay tuned! 👀
This was the project that started the collaboration between @SNUVLLab and @cmuhcii, I am greatful that we finally are able to publish it! 🚀
✨Kudos to all my collaborators and mentors; Wesley Deng (@wes_deng ), Cheyon Jin, Matheus Maldaner (@matheusmaldaner), Jordan Wheeler, Jason Hong (@jas0nh0ng), Hong Shen (@hongshenus), Adam Perer (@adamperer ), Ken Holstein (@d19fe8), Motahhare Eslami (@meslami2) and my advisor, Gunhee Kim (@gunheekim).
0
4
13
832
Dayoon Ko retweeted
🚀 Excited to share Meta^n: Recursive Self-Improvement through Emergent Depth, a project I’m grateful to have worked on.
What excites me most about Meta^n is that it moves beyond improving an agent’s answers to improving the process that produces them. By recursively applying a fixed meta-operation, the system builds increasingly deep layers of executable strategies—and those layers naturally develop distinct roles without being explicitly designed to do so.
A huge thank you to @zaemyung 🙏
What excites me most about Meta^n is that it moves beyond improving an agent’s answers to improving the process that produces them. By recursively applying a fixed meta-operation, the system builds increasingly deep layers of executable strategies—and those layers naturally develop distinct roles without being explicitly designed to do so.
A huge thank you to @zaemyung 🙏
1/ Self-improving agents refine answers, not the process that produces them. Meta^n keeps one meta-operation Ω and recurses on its input: each layer writes code that rewires the layer below. Depth emerges, and so do roles.
📄 arxiv.org/abs/2608.24735
💻 github.com/minnesotanlp/m…
1
4
28
1,749
Dayoon Ko retweeted
We’re asking the research community to help us build a benchmark for research taste.
Scientific discovery starts with a fundamental step: which prior work is worth building on? We want to capture this undocumented layer through our collective knowledge.
Please sign up: tinyurl.com/45b9ykae ↓
Scientific discovery starts with a fundamental step: which prior work is worth building on? We want to capture this undocumented layer through our collective knowledge.
Please sign up: tinyurl.com/45b9ykae ↓
6
36
215
77,598
Dayoon Ko retweeted
I'll be speaking at the EMM-QA Workshop tomorrow (Saturday)! #icml2026
Come say hi and hear about PixelRAG -- doing RAG but without text (?) 👀
Don't miss @andylizf's poster as well🙌
pixelrag.ai
arxiv.org/abs/2606.28344
Come say hi and hear about PixelRAG -- doing RAG but without text (?) 👀
Don't miss @andylizf's poster as well🙌
pixelrag.ai
arxiv.org/abs/2606.28344
🚀The ICML 2026 EMM-QA Workshop is happening tomorrow!🚀
Join us at 8:00 AM in ASEM Ballroom 201 at COEX.
The program features four excellent invited talks, paper presentations, shared challenge highlights, and the live AI QA competition!
Program:
qanta-org.github.io/competition/20…
2
5
69
6,982
Dayoon Ko retweeted
Excited to see our FlexOlmo (arxiv.org/abs/2507.07024), released a year ago, is now being incorporated into the Danish Foundation Models (DFM) project!
The Danish Foundation Models (DFM) project is adapting our modular FlexOlmo architecture into a lighter-weight system that runs on commodity hardware—putting collaborative model building within reach of smaller research groups & organizations. 🧵
1
8
50
8,153
Dayoon Ko retweeted
If you’re coming to #ICML2026 🇰🇷, this Seoul Guide might be All You Need!
I’ve lived in Seoul for 20+ years and curated this all-in-one guide: Top 5 must-visits, 100+ recs with Google Maps, practical local tips, and more.
Check it out: ohmyksh.github.io/icml2026/index…
DMs are open and happy to give more details!
I’ve lived in Seoul for 20+ years and curated this all-in-one guide: Top 5 must-visits, 100+ recs with Google Maps, practical local tips, and more.
Check it out: ohmyksh.github.io/icml2026/index…
DMs are open and happy to give more details!
22
60
492
123,534
Dayoon Ko retweeted
🤔Can experience optimizing in one domain (e.g., faster kernel design) transfer to a very different domain (e.g., open mathematical conjectures)?
💡We found that this kind of cross-domain optimization is possible through Evolution Fine-Tuning (EFT).
🤗huggingface.co/papers/2606.29…
Fine-Tuning (EFT), a mid-training paradigm that teaches LLMs how to discover from evolutionary search trajectories. We train 🐦 Finch 2B–9B on the Finch Collection — 156K trajectories spanning 371 tasks — and it surpasses its base models by +10.2% on 22 held-out tasks, with emergent cross-domain transfer. ⬇️⬇️⬇️
💡We found that this kind of cross-domain optimization is possible through Evolution Fine-Tuning (EFT).
🤗huggingface.co/papers/2606.29…
Fine-Tuning (EFT), a mid-training paradigm that teaches LLMs how to discover from evolutionary search trajectories. We train 🐦 Finch 2B–9B on the Finch Collection — 156K trajectories spanning 371 tasks — and it surpasses its base models by +10.2% on 22 held-out tasks, with emergent cross-domain transfer. ⬇️⬇️⬇️
2
23
100
19,508
Dayoon Ko retweeted
🚨Most AI agents solve only the problems users explicitly ask about.
But what about the problems users haven’t noticed yet?
🌊TIDE enables proactive multi-problem discovery, helping agents uncover hidden issues 🔍 and recommend actionable next steps ✅.
huggingface.co/papers/2606.04…
But what about the problems users haven’t noticed yet?
🌊TIDE enables proactive multi-problem discovery, helping agents uncover hidden issues 🔍 and recommend actionable next steps ✅.
huggingface.co/papers/2606.04…
2
16
27
4,112
Dayoon Ko retweeted
🇰🇷Despite rapid progress in AI agent research, Korean agentic benchmarks remain largely absent!
To narrow this gap, we release K-BrowseComp, a benchmark that requires searching across Korean websites and Korean-language content.
arxiv.org/abs/2606.02404
To narrow this gap, we release K-BrowseComp, a benchmark that requires searching across Korean websites and Korean-language content.
arxiv.org/abs/2606.02404
5
26
112
21,118
Dayoon Ko retweeted
Existing work on AI Reviewers primarily rely on superficial metrics such as correlation against human scores. However, even if the scores are matching, the actual content of the reviews drastically differ between human and AI.
Here, 45 scientists actually review the contents.
Here, 45 scientists actually review the contents.
Recently, there's been complaints on low-quality AI reviews at conferences and journals. What if we put the frontier LMs into an agent harness?
With the right setup, on 82 Nature-family papers, 45 expert scientists judged that AI reviewers outperform the best human reviewer!
🤗 huggingface.co/papers/2605.20…
0
3
8
1,421
Dayoon Ko retweeted
Today's a special day for me! We released Nemotron-Personas-Korea, the 1st Korean persona dataset🇰🇷💚 Built the largest persona PGM ever from 62 census data, capturing up to 10^46 states to closely simulate Korea. Already trending Top5 on 🤗 plz hit like❤️huggingface.co/datasets/nvidi…
11
145
446
155,242
Dayoon Ko retweeted
🌎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 g models’ online adaptation to streaming, continually updating knowledge.
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 g models’ online adaptation to streaming, continually updating knowledge.
3
21
65
14,630
Dayoon Ko retweeted
Hybrid Deep Searcher: Integrating Parallel and Sequential Search Reasoning
HybridDeepSearcher from LG is a Qwen3-8B model fine-tuned on HDS-QA (1,987 hybrid-hop questions, 2,111 correct trajectories) to distinguish parallel from sequential queries. It integrates both modes in ery–retrieval loops, cutting latency, broadening evidence retrieval, and scaling accuracy where sequential or naive multi-query baselines plateau.
- HDS-QA: synthetic hybrid-hop dataset from Natural Questions, mixing independent parallel queries with dependent sequential steps
- Training / Mechanism: Qwen3-8B fine-tuned for 1 epoch (lr 3e-5, batch 4, grad accum 32) on supervised trajectories with <think> reasoning and multi-query blocks; structured reasoning alternates with query–retrieval cycles, issuing parallel queries when possible to reduce turns and scale with budget
Results:
- FanOutQA: 15.9 F1 improvement
- BrowseComp-50: +11.5 F1 with fewer turns
- Evidence retrieval: 61% (FanOutQA), 55.8% (FRAMES), 40.7% (MuSiQue) vs 53/49/38 baselines
- Efficiency: highest AUC, answers in fewer turns, keeps improving with more turns/calls
HybridDeepSearcher from LG is a Qwen3-8B model fine-tuned on HDS-QA (1,987 hybrid-hop questions, 2,111 correct trajectories) to distinguish parallel from sequential queries. It integrates both modes in ery–retrieval loops, cutting latency, broadening evidence retrieval, and scaling accuracy where sequential or naive multi-query baselines plateau.
- HDS-QA: synthetic hybrid-hop dataset from Natural Questions, mixing independent parallel queries with dependent sequential steps
- Training / Mechanism: Qwen3-8B fine-tuned for 1 epoch (lr 3e-5, batch 4, grad accum 32) on supervised trajectories with <think> reasoning and multi-query blocks; structured reasoning alternates with query–retrieval cycles, issuing parallel queries when possible to reduce turns and scale with budget
Results:
- FanOutQA: 15.9 F1 improvement
- BrowseComp-50: +11.5 F1 with fewer turns
- Evidence retrieval: 61% (FanOutQA), 55.8% (FRAMES), 40.7% (MuSiQue) vs 53/49/38 baselines
- Efficiency: highest AUC, answers in fewer turns, keeps improving with more turns/calls
1
2
8
975