Shreyas Kowshik
skowshik@andrew.cmu.edu

I am a Master of Science in Machine Learning student at Carnegie Mellon University, where I work on robot learning, reinforcement learning, vision-language-action models, and LLM agents. I am particularly interested in post-training foundation models, with the broad goal of fully autonomous improvement: Given a new task and an interaction protocol, agents must be able to adapt autonomously and safely to the new environment.

At CMU, I work with Prof. Aviral Kumar and Prof. Max Simchowitz on data-efficient adaptation of generalist robot policies and continual learning for LLMs. Before graduate school, I was an Applied Scientist in Microsoft's Speech and Language Group, where I built and shipped large-scale speech-to-text systems.

Outside of work, I enjoy playing the keyboard and guitar, table tennis, badminton, rock music, and horror films.

CV  /  Email  /  Github  /  LinkedIn  /  Google Scholar

• [October 2026]  First-author work on minimal-data adaptation of generalist robot policies accepted at R2RL@CoRL'26/RL4VLA@RSS'26 and was submitted to ICLR'26.
• [Fall 2026]  Teaching Assistant for Deep Reinforcement Learning at CMU.
• [June 2026]  Started leading a research project on continual learning of LLMs with Prof. Aviral Kumar and Prof. Max Simchowitz.
• [March 2026]  COACH selected for an oral presentation at the HCAIR Workshop at ICLR'26.
• [2024]  Completed my time at Microsoft after shipping four transformer-based speech-to-text models.

My research focuses on post-training foundation models with the long-standing goal of fully autonomous improvement. Given a new task where a foundation model fails or is not reliable and an environment-interaction protocol, we would want the model to be able to collect data, learn from it and improve itself with minimial (perhaps no) human intervention. To this end, I am particularly interested in vision-language-action models, imitation and reinforcement learning and continual learning for LLM agents. My recent work has focussed on online-RL adpatation of extremely brittle initial policies (with success as low as 4%) in data-constrained settings. I have also been working on continual-learning of LLM coding agents, where I am working on a recipe to compress unsupervised interaction experience to induce issue solving capabilities outside of the pretrained model.

Leading research on continual learning for LLMs. I am working on a recipe to compress unsupervised interaction experience to induce issue solving capabilities outside of the pretrained model.

Developed MiDAS, an offline-to-online reinforcement learning method that adapts pretrained VLA policies to new tasks from as little as one demonstration. Combined behavior cloning with value-based residual learning over frozen VLA representations, improving task success in LIBERO and RoboCasa and on a bimanual YAM setup. Overall, MiDAS helps isolate the role that pretraining, demonstrations and online-learning play in successful adaptation, and prescribes a recipe to get RL working with extremely brittle initializations.

Proposed COACH, a framework in which one LLM generates worked examples that condition a second LLM for code solutions. Jointly trained both models with GRPO and execution-based rewards, improving pass@1 by 32% over vanilla GRPO-based RLVR on MBPP with 40% better sample efficiency.

Shipped four transformer speech-to-text models, scaling distributed training with DeepSpeed to more than 100 million parameters. Built pipelines over 100,000 hours of speech and a neural language-model rescoring system, reducing word error rate by 10% and accelerating training by 5×.

Built an inference-only intervention that remaps the VLM prefix KV-cache of a frozen π_0.5 flow-matching policy. Found that success collapses below 12 of 18 VLM layers, while a single denoising step nearly matches ten on LIBERO-10.

Proposed test-time adaptation for VLAs by refining actions online with GradQ ascent on a warm-started Q-function. Improved task success by 18% over behavior-cloning baselines on the hardest long-horizon LIBERO-Long tasks.

Implemented core deep and offline RL algorithms, model-based learning, and imitation learning from scratch; derived foundational computer-vision methods; and built a Llama-3 model with RoPE, RMSNorm, grouped-query attention, SwiGLU, training, and sampling.

Built a MobileNet-SSD traffic-sign detector running at more than 50 FPS with an integrated tracking pipeline. Designed perception for the Eklavya 6.0 and 7.0 robots, which finished as runners-up at IGVC 2018 and 2019.

Languages and Libraries: Python, JAX, C++, PyTorch, TensorFlow, Hugging Face, NumPy, ROS, CUDA, FAISS

Tools and Infrastructure: Git, Docker, Ray, HPC, Linux, Weights & Biases, Distributed and Mixed-Precision Training (DeepSpeed), MuJoCo, Isaac

Robot Post-Training and ML: Embodied Foundation Models, VLAs, Fine-Tuning and Domain Adaptation, Imitation Learning, RL/RLHF/RLVR, Distillation, Synthetic Data, Curriculum Learning, Flow-Matching Policies, Real-Robot Validation, Sim-to-Real, Robustness and Controllability, Evaluation and Failure Analysis

Template adopted from Jon Barron.