Speakers
Description
Reproducing modern AI research often sounds straightforward: read the paper, clone the repository, download the checkpoint, and run the experiments.
In practice, that is often where the problems begin.
What happens when the model weights are missing? When the training configuration in the paper does not quite match the repository? When the dataset format has changed, dependencies no longer work, or your reproduced results differ from what was reported?
In this talk, I’ll walk through the practical process of turning an AI research paper into a working, reproducible system, using my experience reproducing vision-language-action models as a running example. In one project, official trained weights were not available, so I had to train ACoT-VLA and π0.5 from scratch using the AgiBot challenge dataset and reconstruct the surrounding training and evaluation workflow.
I’ll cover the parts of reproduction that papers often leave implicit: understanding an unfamiliar research codebase, reconstructing configurations, preparing datasets, debugging training instability and GPU memory problems, validating intermediate results, and distinguishing an implementation bug from an undocumented research assumption.
I’ll also discuss reproducibility from an open-source engineering perspective: environment pinning, checkpoints, experiment metadata, deterministic evaluation, useful failure logs, and how researchers can structure repositories so another person can realistically reproduce their work months later.
The goal is not to present another model architecture. It is to share a practical workflow for going from “this paper looks interesting” to “I can actually run, test, modify, and trust this system.”
Attendees should leave with a repeatable approach for reproducing unfamiliar AI projects and a better sense of where ML reproducibility usually breaks in the real world.
Session author's bio
Atul Saxena is a 20-year-old ML researcher and Founder of BlackdromeAI Labs, where he is changing the landscape of future Agents and how RL environments work. His work spans machine learning systems, reinforcement-learning environments, multimodal AI, mechanistic interpretability, and open-source developer tooling. He has published research studying computational pathways inside GPT-2 and has built GS-DroneGym, a vision-language-action drone simulation environment, and LEMMA, a neuro-symbolic mathematical reasoning system. He has also worked on RL and agent evaluation at Bespoke Labs and has experience reproducing and training research models when official checkpoints or complete implementation details are unavailable. He enjoys turning research ideas into systems that can actually be run, tested, reproduced, and extended by other developers. Also the Co-Founder of PkBoostAI Labs building next gen Neural Network based Systems and deterministic models. Also contributed to various open source communities such as Eigent-AI, Apache, Open-telemetry and more, had also worked with Snorkel AI.
Divyanshi Sharma is an AI and data science researcher pursuing a B.S. in Data Science at IIT Madras. She is currently a Researcher at IIT Mandi’s IKSHMA Lab, working on vision-and-language navigation, multimodal reasoning, and 3D scene understanding. Her work includes reproducing and extending Open-Nav, an ICRA 2025 vision-language navigation system, and contributing to AXON, a real-time mechanistic interpretability platform for visualising internal feature activations in language models. Her broader interests span multimodal AI, interpretable machine learning, autonomous systems, and open-source AI research.
Any other info we should know?
This will be a technical but practical session aimed at developers, students, ML engineers, and researchers who have tried to run or reproduce open-source AI projects.
No prior knowledge of VLA models or robotics is required. Basic familiarity with Python, machine learning concepts, and using Git/GitHub will make some examples easier to follow, but the talk is designed to be understandable without specialist knowledge.
I plan to keep the session implementation-focused rather than mathematical, with examples of real reproduction failures, debugging decisions, training/evaluation workflows, and lessons for making open-source research easier for others to reproduce.
Questions and discussion from the audience will be welcome throughout or at the end depending on the session format.
| In Person Attendance | In-person |
|---|---|
| Level of Difficulty | Intermediate |
| Social Media | https://x.com/Blackdromedang1 |
| Agree to Privacy Policy and Notice | I agree |
| Please confirm that there are included headshots of all speakers in their profiles | Yes |