I am a master's student in AI at École Polytechnique, interested in robot learning, representation learning, diffusion/flow matching, and multimodal LLMs. Most recently I was a research intern at Inria Paris in the Willow team, with Cordelia Schmid and Shizhe Chen, on pretraining 3D representations for vision-language-action models and downstream manipulation. Before that I worked with Georgia Tech on multimodal AI for finance and co-authored two papers in the area, at KDD 2025 (oral) and an ICCV 2025 workshop. I came to research after three and a half years as a software engineer at Urban Company.
Looking for research internships from April 2027 in reinforcement learning and robot learning, multimodal learning, diffusion/flow matching, and LLM and VLM pre- and post-training.
Yash Bhardwaj*, Siddhant Sukhani*, Riya Bhadani, Veer Kejriwal, Michael Galarnyk, Sudheer Chava
The first benchmark for topic-aligned captioning of financial short videos: 624 YouTube clips, 5 topics and all 7 combinations of transcript, audio and video across 5 multimodal LLMs. Video alone is the best single input on 4 of 5 topics, while a well-chosen pair such as audio + video often beats all three.
Michael Galarnyk*, Veer Kejriwal*, Agam Shah*, Yash Bhardwaj, Nicholas Watney Meyer, Anand Krishnan, Sudheer Chava
A benchmark of 288 finfluencer videos (43 hours, 6,000+ expert annotations) that asks which stock is recommended, what action is advised and how strongly the speaker believes it. Across 16 LLMs and 6 video MLLMs, ticker extraction reaches 86% F1, but adding the action caps the best model at 54%, and adding conviction at 28%.
Repairing a latent world model on the planner's own data cut its top-choice regret by 14% against random data (pre-registered, 12 seeds, p = 0.0015), yet random data did more for closed-loop control, and held-out loss predicted neither.
Continuing a trajectory already seen is 43–62% cheaper per success, but 54–67% of those groups return identical rewards, which gives RLOO zero gradient. RLOO itself lifts Qwen3-1.7B on ALFWorld from 58.2% to 64.6% (p = 0.019).
Fine-tuned Qwen3-VL-2B with LoRA on a 4-bit base (17M trainable parameters, 0.8%) to tell a wheel loader where to dig: a target bounding box, its spatial position and a discrete action. Trained on 848 frames labelled automatically by a Florence-2 data engine, in about 10 GB of GPU memory.
Reproducing differentiable MPC on the pendulum: every seed imitates the expert and recovers the identifiable ratios g/l and 1/(ml²) within 2%, while g itself lands anywhere from 7 to 17.
Distilled Stable Diffusion 1.5 into a LoRA student that generates images in 2 to 8 steps, up to 18× faster than the 50-step model (0.12 s against 2.17 s per image). Built a CUDA-timed benchmark harness with CLIP text-image alignment and qualitative comparison grids.
Fine-tuned Stable Diffusion 1.5 to generate images that evoke one of 8 emotions, on 100K images with LoRA (0.4% of parameters trainable): 15× faster training and 80% less memory than full fine-tuning. Compares learned emotion tokens, latent classifier guidance and multimodal conditioning.
A CLIP-style dual encoder (ResNet-18 for images, DistilBERT for text) trained with a symmetric InfoNCE loss on ROCO radiology image-caption pairs, reaching 28% Recall@1 and 55% Recall@10 on image-text retrieval, served through a FAISS index.
