Yogesh Kulkarni

Research Intern @AdobeResearch | CS PhD @SCAI_ASU | Prev: @CSatUSC, @BellLabs | MLLMs & RL

[Yogesh Kulkarni @__yogesh007]

EgoVITA is accepted at ECCV 2026! 🎉🇸🇪
MLLMs give plausible-but-ungrounded answers on egocentric video. We fix it by — planning in first person, verify from a third-person view.
#ECCV2026 @eccvconf

w/ @PooyanFazli @gameschoolasu @SCAI_ASU @ASUEngineering

[Yogesh Kulkarni @__yogesh007]

@AryanEverywhere DM

[Yogesh Kulkarni @__yogesh007]

@OmeadMoses It is two step generate plan then use generated plan along with video to verify.

[Yogesh Kulkarni @__yogesh007]

EgoVITA is accepted at ECCV 2026! 🎉🇸🇪
MLLMs give plausible-but-ungrounded answers on egocentric video. We fix it by — planning in first person, verify from a third-person view.
#ECCV2026 @eccvconf

w/ @PooyanFazli @gameschoolasu @SCAI_ASU @ASUEngineering

[Yogesh Kulkarni @__yogesh007]

TL;DR: EgoVITA makes egocentric reasoning work:

1️⃣ Plan first-person, verify third-person
2️⃣ Ground every step in the frames where it happens

Result: attention sharpens onto task-relevant objects and hands, not the background.

📄: https://t.co/FgH8k3ubSG

[Yogesh Kulkarni @__yogesh007]

🚀 +7.7 EgoBlind, +4.4 EgoOrient over Qwen2.5-VL
🚀 Holds across 3 backbones (Qwen2.5-VL, InternVL3.5, Qwen3-VL)
🚀 No forgetting (+2.7 Video-MME)
🚀 Just 47k samples

[Chaoyu Li @chaoyu72033]

🚀🚀🚀 Excited to share our new paper: “Video Generation Models: A Survey of Post-Training and Alignment.”

As video generation advances, the key challenge is no longer just scaling, but reliability, controllability, and alignment with human intent. Post-training alignment hence has emerged as the central stage for shaping model behavior in video systems.

Our work provides the first comprehensive review of post-training alignment in video generation, introducing a unified perspective on implicit vs. explicit alignment and organizing the landscape across:
• Supervised Fine-tuning Methods
• Self-training & Distillation Methods
• Preference- & Reward-Based Methods
• Inference-Time Methods

Beyond methods, we systematically review datasets, benchmarks, evaluation protocols, and broader research challenges in post-training alignment.

Paper: https://t.co/tz2hIy9pTU
GitHub (continuously updated): https://t.co/0SkyXht4Tw https://t.co/dPZQLOpc4T

#VideoGeneration #GenerativeAI #PostTraining #Alignment #AI #ComputerVision

[Yogesh Kulkarni @__yogesh007]

@CVPR @PooyanFazli @gameschoolasu @SCAI_ASU TL;DR: AVATAR makes RL for video work:
1️⃣ Replay hard rollouts instead of wasting them
2️⃣ Credit to tokens that matter — planning & synthesis

📄: https://t.co/Ch9sYAgq7P

[Yogesh Kulkarni @__yogesh007]

Happy to share AVATAR is accepted at CVPR 2026! On-policy RL wastes video rollouts, stalls when rewards collapse, and spreads credit uniformly across tokens. We fix all three — replay what's hard, reward what's critical.
@CVPR #CVPR2026
w/ @PooyanFazli @gameschoolasu @SCAI_ASU

[Yogesh Kulkarni @__yogesh007]

@CVPR @PooyanFazli @gameschoolasu @SCAI_ASU Qualitative:
GRPO: "He looks worried which indicates stress"
AVATAR: links "tense expression, eyes darting" (visual) w/ "hurried tone" (audio), tracks "tone shifting from calm to anxious"