Risheek Garrepalli

Risheek Garrepalli

Staff Research Scientist at Qualcomm AI Research. Generative modeling, RL, representation learning. Graduate research advised by Tom Dietterich, Alan Fern.

Selected Publications

Below is a curated selection of recent preprints and foundational papers that best reflect my current research trajectory. For my full list of peer-reviewed publications (including NeurIPS, CVPR, etc.), please see my Google Scholar.

Garrepalli = me; * = equal contribution.

Reinforcement Learning & Generative Modeling

How iterative models learn from their own rollouts without losing coverage. (Diffusion distillation reframed as sequential decision-making)

Covariate shift → DAgger-style correction → On-policy learning
Train on student-visited denoising states; retain teacher/reference states to preserve diversity and coverage. This reframes diffusion distillation as rollout learning rather than one-step matching.

  1. Garrepalli, Mahajan, Hayat, Porikli. “DDIL: Improved Diffusion Distillation With Imitation Learning.” arXiv 2024. [arXiv]
  2. Yasarla, Hegde, Han, Cheng, Shi, Sadeghigooghari, Mahajan, Bhattacharyya, Liu, Garrepalli, et al. “Generative Scenario Rollouts for End-to-End Autonomous Driving.” arXiv 2025. [arXiv]

Representation, Competence & Uncertainty

What a representation retains sets the ceiling on what downstream monitors can detect.

Key takeaway: Open-set/novelty detection cannot recover information discarded by the representation; pairing discriminative and generative objectives preserves more of the data manifold and raises that ceiling.

  1. Liu*, Garrepalli*, Dietterich, Fern, Hendrycks. “Open Category Detection with PAC Guarantees.” ICML 2018.
  2. Garrepalli. “Oracle Analysis of Representations for Deep Open Set Detection.” arXiv 2022. [arXiv]

Representation Analysis & Training Objectives

Training objectives shape how a model organizes intermediate information.

Key takeaway: Score matching promotes global structure, while masking promotes local composition. MAgD combines those priors; AR vs dLLM representations measures the distinct layerwise geometry and redundancy of diffusion and autoregressive LLMs.

  1. Garrepalli*, Kadambi*, Borse, Hayat, Porikli. “MADI: Masking-Augmented Diffusion with Inference-Time Scaling for Visual Editing.” arXiv 2025. [arXiv]
  2. Garrepalli*, Goel*, Agrawal, Lott, Lee, Porikli. “Skip to the Good Part: Representation Structure & Inference-Time Layer Skipping in Diffusion vs. Autoregressive LLMs.” arXiv 2026. [arXiv]

Generative Priors in Perception & Embodied AI

Generative priors encode task-relevant structure for perception and control.

Key takeaway: Hierarchical refinement structures correspondence, future prediction improves depth, and uncertainty-aware expressivity addresses long-tail driving.

  1. Garrepalli, Jeong, Ravindran, Lin, Porikli. “DIFT: Dynamic Iterative Field Transforms for Memory Efficient Optical Flow.” CVPR Workshop 2023.
  2. Yasarla, Han, Cheng, Liu, Mahajan, Bhattacharyya, Shi, Garrepalli, Cai, Porikli. “RoCA: Robust Cross-Domain End-to-End Autonomous Driving.” arXiv 2025. [arXiv] (originated framing: representational expressivity as the bottleneck for long-tail and rare cross-domain scenarios)
  3. Yasarla, Singh, Cai, Shi, Jeong, Zhu, Han, Garrepalli, Porikli. “FutureDepth: Learning to Predict the Future Improves Video Depth Estimation.” ECCV 2024. (motivated generative prediction prior grounded in oracle analysis; advocated for on-policy/non-teacher-forcing training formulation)