Victor S. Bursztyn

Research Scientist at Adobe Research. PhD (Northwestern University) specialized in NLP. Opinions my own.

[Victor S. Bursztyn @victorsb]

Our work on Compositional Fine-Tuning has been accepted to findings of EMNLP!
We study if incrementally fine-tuning GPT-3 on a curriculum of sub-tasks helps it learn a complex end task, such as making recommendations. Preprint: https://t.co/le6UMs5INN
#NLProc #emnlp2022

[Jiale Liu @Jiale_Leo]

My favorite takeaway: AI should not replace teamwork. It should help teams understand each other faster, preserve diverse voices, and make the path to consensus clearer. 🌱

📄 Paper: https://t.co/cBZcRvOuRI
📰 Adobe Research blog (check out the demo here): https://t.co/E0VsJD3BiC

Huge thanks to my amazing collaborator Lin Ai, my mentors @victorsb, Haoliang Wang, Sunav Choudhary, Saayan Mitra, and my advisor @qingyun_wu for making this work happen! 🙏

[Jiale Liu @Jiale_Leo]

We put real professional designers in the loop on 50 actual ad design scenarios. 👩‍🎨👨‍🎨 The results:

✅ They agreed with their agent's comments (< 3% strong disagreement)
✅ They preferred TeamFusion's remixed designs over the originals
✅ Team-wide consensus rose by about 17% on average

(It also beat direct-summary baselines across a 500-team public-opinion benchmark.)

[Jiale Liu @Jiale_Leo]

TeamFusion takes a different approach, in 4 steps:
1️⃣ Represent: build a proxy agent for each teammate from their written feedback
2️⃣ Discuss: agents debate in turns, so no single voice dominates
3️⃣ Remix: synthesize the debate into an editable, shared deliverable
4️⃣ Refine: critique and repeat 🔁

[Jiale Liu @Jiale_Leo]

The Motivation: Many team decisions are not about finding one correct answer. Fields like design, product planning, and creative work all involve different tastes, priorities, and trade-offs. 🎨🧠

But today's AI often compresses everyone's input into one "average" answer, which quietly hides the disagreements that matter most.

[Jiale Liu @Jiale_Leo]

Excited to share that TeamFusion, my 25 summer intern work at @AdobeResearch, is featured on Adobe Research blog, and accepted to #ACL2026 main conference! 🎉🎉🎉

TL;DR: TeamFusion helps teams make open-ended decisions by simulating teammate discussions and remixing ideas into stronger team-wide outcomes, all with minimal communication overhead.🤝✨

#Agents #MultiAgentSystems #HumanAI #LLMAgents

[Victor S. Bursztyn @victorsb]

@lindolagodoamor Vim aqui pra fazer a mesma sugestão: a bola tá quicando pra chamar de "Horácia". #ahorapodcast

P.S. De um ouvinte que trabalha com IA no Vale do Silício e acompanha vocês desde o doutorado (justamente) em language models na Northwestern University.

[Victor S. Bursztyn @victorsb]

@thais_bilenky @MidiaTrovao @revistapiaui Só pra sublinhar a coincidência de o Alexandre ter tomado uma decisão sobre o inquérito dos empresários golpistas horas depois desse episódio (!). Será que ele escuta na segunda mesmo?

[Victor S. Bursztyn @victorsb]

@Xfinity @Xfinity had the brilliant idea of renting a residential space (where I live) to throw a corporate party and DISRUPTED ALL indoor common areas. My wife has Covid (I don't), and this brand left me with zero place to work. Accounts cancelled w/ this ANTI-MARKETING @SAC_VillageApts

[Victor S. Bursztyn @victorsb]

@tallinzen I found this to be a fun paper about human data-efficiency: https://t.co/mwFI8VqA4X

[Victor S. Bursztyn @victorsb]

@ndiakopoulos @percyliang's group started a knowledge graph to keep track of this space: https://t.co/7TpdqScnPX

[Percy Liang @percyliang]

The ecosystem graph is a community resource for tracking all the foundation models, datasets, and products in this ridiculously fast moving space:
https://t.co/ye1ZMAysHT
Help us maintain it by adding new models, datasets, and products: https://t.co/xxZnWR8Kg0

[Victor S. Bursztyn @victorsb]

At the same time, ChatGPT is overconfident (outputs confidently even when wrong) and totally lacks self-confidence (allows users to correct it even when clearly right). Funny how both can co-exist. #NLProc

[Victor S. Bursztyn @victorsb]

@WilliamWangNLP Agreed. But even more than incremental example difficulty (curriculum learning), ordering examples by compositional task structure (what inspired CL) can be specially useful for LLMs: https://t.co/le6UMs5INN

[Victor S. Bursztyn @victorsb]

@anmarasovic Maybe something like log analysis e.g. clustering logs by occurrences of certain exceptions.

[Victor S. Bursztyn @victorsb]

@madiator So it's just a bunch of if-then rules?

[Victor S. Bursztyn @victorsb]

@ericzelikman @Abmld2 @Scobleizer This combination is fun! We tried something along this line with GPT-3 and CLIP: https://t.co/j63TX5blUR
(You can preserve the style if you condition CLIP on the concatenation of the previous image &amp; text prompt.)

[Victor S. Bursztyn @victorsb]

@yoavgo There seems to be bigger pragmatic gaps, even with the model knowing a lot about the underlying facts (softball from https://t.co/hhkbWDpRuK tested by @_DougDowney).

[Victor S. Bursztyn @victorsb]

This pragmatic gap seems to disappear after a factual comparison is explicitly added to a conversation. Full details about the task are available at: https://t.co/hhkbWDpRuK 5/5

[Victor S. Bursztyn @victorsb]

TLDR: Testing our EMNLP task on #ChatGPT, and it ignores the pragmatics of the task even though it clearly knows all the base facts. #nlproc

🧵1/5

[Victor S. Bursztyn @victorsb]

Boston vs Miami (or vs Oaxaca) are softballs. I wonder if ChatGPT is more inclined to surface relevant facts than it’s to solve the pragmatics of the question… Humans probably do the opposite order. 4/5