Research Intern @AdobeResearch, CS Ph.D. student @UWMadison | prev. intern: @AdobeResearch; @Krafton_AI
Madison, WI
Joined October 2023
- Tweets65
- Following679
- Followers135
- Likes403
Pinned Tweet
We all knew LLM agents struggle to explore, but we had to eyeball it ๐. We couldn't measure exploration errors. Until now. ๐บ๏ธ๐ค
We built a policy-agnostic metric to quantify exploration and exploitation errors in LLM agents.
Spoiler: Exploration error is what kills๐ agent performance in our setting ๐๐งต(1/8)
Our lab has a robot now! And a robotics paper. Amazing work @trungdt880 ๐ฆพ๐ฆพ
thaw-vla.trung-dt.com
My first robotics paper is out ๐ค
World models are basically robotics' smartest friend who thinks 7 times before answering any question. great with text, terrible at parties ๐ฅบ
So we asked: can we keep the brain and lose the lag? THAW-VLA ๐งต
(1/n)
I am attending #SIGGRAPH2026 to present โMAOAM: Unified Object & Material Selection with Vision-Language Modelsโ next week (July 23rd Thursday, 11:35-11:45 am at room 408)!
MAOAM is a unified segmentation model that can select both objects and materials under click- and text-based interactions. ๐งต(1/9)
Also, feel free to try our model out and check out the data released! ๐งต(8/9)
- HF demo: huggingface.co/spaces/jpark6โฆ
- HF dataset: huggingface.co/datasets/jparโฆ
Huge thanks to my co-authors @vdeschaintre, Jason Kuen, Kangning Liu, @iliyang, @Krishnakusin, @yong_jae_lee, @michi_fischer! ๐งต(9/9)
+ If you are attending SIGGRAPH, Iโd love to connect, especially if you are working on physical AI, robotics or world (action) models! Feel free to reach out :)
Jaden Park retweeted
The reversal curse. Edits that don't suppress negations. Multi-hop updates that don't propagate. These look like separate bugs.
Our ICML 2026 spotlight argues they may share a common geometric origin, visible only when you study how representations move under updates ๐งต
(1/11)
Jaden Park retweeted
Almost all "flagship" models are now MoEs.
But smaller models still prefer to be dense as they target memory-constrained scenarios where total params matter.
So we ask: Can we leverage an MoE to produce dense models without having to train them from scratch?
๐งต๐
Lots of good news this week! ๐
1. My internship project from @AdobeResearch has been accepted to #SIGGRAPH2026!
("MAOAM: Unified Object & Material Selection with Vision-Language Models")
Special thanks to my wonderful mentor @michi_fischer who has made this project possible!
2. Paper accepted to #ICML2026!
("DocHop: Benchmarking Out-of-domain Multi-hop Reasoning in Information-Dense Documents")
3. Paper accepted (with minor revisions) at #DMLR!
("Decomposing Complex Visual Comprehension into Atomic Visual Skills for Vision Language Models")
In both papers, we generate carefully designed benchmarks to tackle compositional/multi-hop reasoning in VLMs. Proud to have contributed in these projects.
More detailed posts soon :) Stay tuned!
Jaden Park retweeted
We all knew LLM agents struggle to explore, but we had to eyeball it ๐. We couldn't measure exploration errors. Until now. ๐บ๏ธ๐ค
We built a policy-agnostic metric to quantify exploration and exploitation errors in LLM agents.
Spoiler: Exploration error is what kills๐ agent performance in our setting ๐๐งต(1/8)
I will be at #ICLR2026 to present my work on data contamination in VLMs! (Fri, Apr 24, 2026 โข 8:30 AM โ 11:00 AM, Pavilion 3 P3-917)
I am currently interested in VLA/physical AI, agents and robustness/generalization.
Would love to chat and connect with anyone with similar interests :)
Me: memorize past exams ๐๐ฏ
Also me: fail on a slight tweak ๐คฆโโ๏ธ๐คฆโโ๏ธ
Turns out, we can use the same method to ๐ฑ๐ฒ๐๐ฒ๐ฐ๐ ๐ฐ๐ผ๐ป๐๐ฎ๐บ๐ถ๐ป๐ฎ๐๐ฒ๐ฑ ๐ฉ๐๐ ๐! ๐งต(1/10)
- Project Page: mm-semantic-perturbation.gitโฆ
Can we improve exploration failures in LM agents? ๐ ๏ธ
๐บ๏ธ Exploration Prompts: Explicitly injecting exploration strategies increases success rate by 17%.
๐ Explicit Harness: Providing the agent with structured summaries of its past observations; success rate boost by 29.4%! ๐งต(7/8)
This was a joint co-first author work with @jungtaek_kim and @jongwonjeong123, with the guidance from @rdnowak, @Kangwook_Lee and @yong_jae_lee.
If this sounds interesting, please check out our paper ๐ arxiv.org/abs/2604.13151!
If you have any questions, feedback, or new ideas, I'd be more than happy to discuss!๐งต(8/8)
