Yanjun Fu

CS Ph.D. Student @UofMaryland | Prev. Research Intern @AIatMeta

[Sanghamitra Dutta @Sangha26Dutta]

✈️Travelling to #NeurIPS2025 in San Diego with my PhD students @FaisalHamman @yanjun_fu @BarprodaH. Looking forward to catching up with old and new friends! Reach out if you want to chat about #TrustworthyAI #Interpretability #Efficiency #DataCentricAI #InformationTheoreticMeasures #GroupSimilarityRewards #TabularLLM

[Yanjun Fu @yanjun_fu]

Joint work with @FaisalHamman and @Sangha26Dutta.

[Yanjun Fu @yanjun_fu]

📢Excited to present our work at #NeurIPS2025 in San Diego!
👕T-SHIRT: Token-Selective Hierarchical Data Selection for Instruction Tuning
#SFT #DataSelection #EfficientPostTraining
⏰ Wednesday, Dec 3, 4:30–7:30 PM PST
➡️ Exhibit Hall C/D/E, #200

[Yanjun Fu @yanjun_fu]

Previous work selects high-quality SFT data by scoring entire samples and splitting them into high- and low-quality sets using a hard threshold.
In contrast, our approach uses fine-grained token-level quality and neighborhood quality to guide sample selection.