Hello,
I’m Łukasz
Borchmann
ABOUT
Today, I am primarily focused on large language models, LVLMs, and document understanding, which includes developing novel architectures and inference methods.
Most of the recent research was conducted at Applica.ai, where I have worked since 2018, most recently as a Senior Research Scientist. This journey continues under the new banner after Applica was acquired by Snowflake in 2022.
In the same year, I defended my PhD thesis focusing on neural network architectures shifting paradigm toward what is now called GenAI with honors.
At the same time, I am involved in developing models to prevent fake news, hoaxes, and disinformation at Adam Mickiewicz University. My previous work in this area resulted in winning the SemEval 2020 propaganda detection shared task and receiving the Best Paper Award at this venue.
Every once in a while, I serve as a reviewer for NeurIPS, ICML, CVPR, ICLR, ACL, EMNLP, and a couple of smaller venues.
SELECTED PUBLICATIONS
Some of my recently published papers.
STable: Table Generation Framework for Encoder-Decoder Models
2024.03Document Understanding Dataset and Evaluation (DUDE)
2023.10Sparsifying Transformer Models with Trainable Representation Pooling
2022.05DUE: End-to-End Document Understanding Benchmark
2021.12Going Full-TILT Boogie on Document Understanding with Text-Image-Layout Transformer
2021.09
WHAT'S NEW
Since you’ve made it this far, you might be interested in my recent activity.
It has been a while since I started my career in Computer Science, but in the middle of the 2000s, it was mainly software development that I found no longer attractive. In the early 2010s, I moved into applied ML and Natural Language Processing, whereas a few years ago, I started conducting serious research in the field. Here are some updates regarding the most recent period of my life.
-
2024
I worked on Snowflake's Arctic dense-MoE hybrid LLM and published two pieces on large language model evaluation: one introduced the FeatEng benchmark, designed to test a model’s capabilities in feature engineering tasks, and the other discussed the challenges of assessing LLMs through multiple-choice problems.
Concerning the document understanding field, I was invited to give a keynote at the ICDAR's Document Analysis Systems event and have written technical reports on the Arctic-TILT model and the applicability of GPT-4 to this domain. -
2023
I created the DUDE data set for multipage Document VQA evaluation. It was presented at ICCV and used for the ICDAR competition I co-organised. I presented a pitch on Scaling up Document Image Understanding at ICDAR's ScalDoc and gave educational talks on LLMs at the Snowflake BUILD event.
-
2022
My paper on sparsifying transformer models was presented at ACL. I published a preprint on text-to-table inference and defended my Computer Science Ph.D. thesis with honors.
-
2021
I published a paper introducing the TILT model, a state-of-the-art Document Understanding solution. It won ICDAR's InfographicsVQA shared task. Subsequently, I proposed the DUE benchmark and was invited by Huawei Research for a talk on End-to-End Document Understanding. Other pieces from this year tackle the problem of semantic retrieval and trainable top-k mechanism.
-
...2020
Since 2014, I have been involved in commercial and scientific ML projects. This resulted in several mildly interesting papers and a 2016 master's thesis focused on automating the segmentation of words into morphemes.
Starting in 2018, I did a lot of applied research related to semantic retrieval and document understanding at Applica.ai. With my Named Entity Recognition model, I won the nationwide PolEval 2018 competition. -
2005...
Between 2005 and 2013, I did boring stuff related to software development I taught myself. In the meantime, I became interested in general and quantitative linguistics, ultimately leading me to Natural Language Processing.