Doyoung Kim (김도영)

My name is Doyoung Kim, and I am a PhD student at NYU CS, advised by Prof. Sherry Yang. Prior to this, I completed my MS at KAIST AI under Prof. Minjoon Seo. Before studying AI, I completed my BS in Mathematics & Computer Science (double major) at KAIST.

I am currently interested in understanding and improving agentic capabilities of (large) neural networks in language and robotics fields, including reasoning, planning, and test-time adaptation.

Email  /  Google Scholar  /  X  /  Github  /  LinkedIn  /  CV

I believe mentorship and collaboration are essential parts of a research journey. They are mutually beneficial: mentees can learn and grow, while mentors gain new perspectives and become better researchers and collaborators.

I currently have several potential research projects that I would be excited to mentor students on:

• Training LLMs on challenging tasks, such as MLE-bench [ref] and Terminal-Bench [ref], primarily using verl [ref].
• Test-time training [ref] for large models, including world models and LLMs.
• Code as Policy [ref].
• Code as Reward [ref].

If you have the background to pursue ML research, please email me. I would be happy to discuss potential projects and mentorship opportunities.

I also welcome students who are new to ML research but eager to learn.

Please see my Semantic Scholar or Google Scholar profiles for the full list.

RL for agents compresses each trial into a scalar reward, while harnesses such as ACE let the model write notes (a playbook) for its next attempt, but the model is never trained to write them. We propose ACE-RL, which trains one model to act, reflect, and curate the playbook inside a fixed harness, rewarding each playbook edit by its effect on a future rollout. ACE-RL reaches the highest held-out reward on four agentic benchmarks at equal training compute.

We propose Tree Structured Language Modeling (TSLM), a new paradigm that enables language models to generate complete search trees using special tokens rather than sequential reasoning. TSLM achieves efficient inference-time scaling by producing divergent reasoning paths in a single forward pass.

While humans can learn complex reasoning from few examples, AI struggles to generalize beyond its training. We enable language models to generate "cognitive maps" - tree-structured expansions of future states - before planning. In maze-solving tasks, this cognitive mapping approach proves to be the only effective method for helping language models extrapolate their planning abilities to larger, unseen mazes.

We propose a self-training method that helps LLMs identify their first incorrect reasoning step ("pit") and use it as a reward signal. Through preference optimization, this method enables LLMs to improve their reasoning process, leading to enhanced mathematical performance.

We introduce a semiparametric model superposing two embedding spaces: parametric token embeddings and nonparametric sequence embeddings. The model is co-trained using weighted cross-entropy loss for language modeling and InfoNCE loss for sequence retrieval to enable generation with citations.

We propose a fine-grained evaluation framework for generative language models based on 12 alignment skill sets, which show a strong correlation between model-based and human-based evaluations.

We introduce a new instruction-tuning dataset called the COT COLLECTION dataset, containing 1.84 million rationales across 1,060 tasks. These rationales were extracted from the FLAN Collection using OpenAI Codex with in-context learning (ICL). We fine-tune Flan-T5 (3B & 11B) with the COT COLLECTION to show both zero-shot and few-shot improvements.

Website design from Jon Barron