
Intern @MSFTResearch #Data#AI PhD@virginia_tech w Prof.@ruoxijia | Ex. intern @Meta FAIR @Meta SuperintelligenceLabs/MSL @NVIDIAAI | Alum @BU_Tweets @ZJU_China
Virginia, VA
Joined October 2023
- Tweets63
- Following96
- Followers60
- Likes98
Pinned Tweet
#OpenToWork I’m completing AI/ML PhD soon and actively looking for full-time industrial positions starting Winter 2026 - Spring 2027! 👨🏻💻
👨🏻💻LinkedIn: linkedin.com/in/fykang/
📩mailto: [email protected]
Always happy to chat :) Please feel free to reach out via LinkedIn or email.
I’ll be attending #COLM2026 (SF) and #NeurIPS2026 (Atlanta), and would love to connect in person! If you’re working in related fields, share research interests, or are hiring for full-time positions, I’d be more than happy to meet/chat :)
1
2
67
1+ years of full-time industrial experience in frontier research labs (#Microsoft Research (#MSR), #Meta Superintelligence Labs (MSL), Meta #FAIR, #NVIDIA AI).
26
#OpenToWork 1+ years of full-time industrial experience in frontier research labs (#Microsoft Research (#MSR), #Meta Superintelligence Labs (MSL), Meta #FAIR, #NVIDIA AI).
I’ll be attending #COLM2026 (SF) and #NeurIPS2026 (Atlanta), and would love to connect in person!
#OpenToWork I’m completing AI/ML PhD soon and actively looking for full-time industrial positions starting Winter 2026 - Spring 2027! 👨🏻💻
👨🏻💻LinkedIn: linkedin.com/in/fykang/
📩mailto: [email protected]
Always happy to chat :) Please feel free to reach out via LinkedIn or email.
2
3
182
- Prior to PhD, I was a full-time research assistant in advanced optimization and operational research for 4 years, and obtained a Master’s degree in #IEOR with journal publications.
19
I’ve published 15+ papers in flagship AI/ML conferences, leading to multiple invited talks, production impacts, and Spotlight presentation at ICLR. I’m a main contributor to the establishment of the workshop series on ”Data Problems for Foundation Models," held at ICLR’{24,25,26}
75
Broader projects involve diagnostic/interpretability tools for reliable ML (e.g., <Data Influence>, <Data Valuation>, <Scaling Laws>, and <Trustworthy Machine Learning> applications).
21
PhD research focuses on **data problems** and **foundation models**, from #DataCuration and #SyntheticData for #Pretraining, #SFT-#RL #Postraining, to #skills for #reasoning and AI agents. Recent work on <Agentic AI > and <Automated Data Curation> attracted broad invitations
1
1
36
I’m a final-year #PhD student in #AI/#ML at @virginia_tech (graduating in Winter 2026-Spring 2027), working with Prof. @ruoxijia on <#DataCentricAI> for <#FoundationModels>, incl. <#AIAgents>, #LargeLanguageModels (#LLMs), #ComputerVision (#CV), #Multimodal, #TeamCollaborativeAI
1
45
📍Poster: Imperial Ballroom #36 at 11 AM PDT
⏱️ Poster Session 1 (Tuesday 10/06)
An amazing research venture and fantastic experience working with the wonderful @Mahavir_Dabas18 @KoMyeongseob, and @ruoxijia! --Come chat with us!!
Also at COLM workshops (Friday, 10/09)
#COLM2026 in #SF🌉! Presenting: Characterizing Model-Native Skills arxiv.org/pdf/2604.17614 **From human-defined skill taxonomies to skills as patterns in model activations.** Human skills: Algebra, Probabilities; model skills: step-by-step derivation vs. symbolic deduction #COLM
2
6
265
Referrals, recommendations, and shares would be greatly appreciated! Thank you :)
#OpenToWork #ResearchScientist #AppliedScientist #LLM #Data #AIAgents
1
36
#COLM2026 in #SF🌉! Presenting: Characterizing Model-Native Skills arxiv.org/pdf/2604.17614 **From human-defined skill taxonomies to skills as patterns in model activations.** Human skills: Algebra, Probabilities; model skills: step-by-step derivation vs. symbolic deduction #COLM
3/n) **Method:** This work proposes AutoSkill: PCA directions on activations as Skill Basis.
Interpreting the Skill Basis
- Skills as a direction in the activation (cosine similarity)
- Interpret based on representative examples consistent with the activation direction
22
2/n) Existing skill-based approaches manually define skills based on task knowledge. If we could identify the skills a model already expresses—we could
• selecting training data by skill,
• steering behavior along skill directions, or
• balancing coverage across skills.
1
28
Generalist agents can already run your data-curation loop, but not necessarily data research---if used out of the box, agents match published baselines in 10 iterations, then grind source ratios instead of trying new methods. The fix wasn't a better model, but better scaffolds.
Thanks for sharing our work on building agents for data work in AI.
TL;DR:
1. Generalist agents can already match strong human-designed techniques.
2. Most gains come from iterative local tuning of familiar strategies, rather than exploring and inventing new data strategies.
3. The biggest surprise to us: when we gave agents paper references and required evidence-grounded method adaptation, the agent composed a new data-selection policy that significantly outperformed strong published baselines.
4. As high-quality data becomes harder to find, agent compute spent on improving data may become a new scaling lever.
So far, we have focused on data selection. But this is a much broader space, with many open questions: how to extend this to other types of AI data work, how to scale up, etc.
Happy to connect with people interested in this direction.
Joint work with the amazing team: Hanze Li, Adam Nguyen, @Mahavir_Dabas18 (@virginia_tech), Prof. @Jiaqi_Ma_ (@UofIllinois), Prof. @fredsala (@UWMadison), Prof. @dawnsongtweets (@UCBerkeley ), and Prof. @ruoxijia (@virginia_tech)
37




