Portrait de Parisa Kordjamshidi n'est pas disponible

Parisa Kordjamshidi

Alumni
Site web
Google Scholar

Publications

One Flow-Transformer for Imagination and Control
Rabiul Awal
Jinseong Jeong
Ankur Sikarwar
Parisa Kordjamshidi
Andrii Zadaianchuk
Sai Rajeswar
Paul Hongsuck Seo
Aishwarya Agrawal
Diffusion and flow models are effective world models for visual reinforcement learning, but existing agents treat them as black-box simulato… (voir plus)rs, leaving the backbone’s representations unused for control. We introduce DRIFT, an online agent in which a single Flow-Transformer serves as both world model and policy backbone, trained from scratch. We find that denoising features alone are suboptimal for control; DRIFT bridges this gap with a next-latent prediction objective that gives the backbone an explicit dynamics signal. Shortcut flow matching reduces imagination to a single denoising step per frame. Across Atari 100k, Craftium, and Crafter, DRIFT is competitive with both latent-dynamics and diffusion world-model baselines.
2026-07-30
WMW @ Reinforcement Learning Conference (publié)
openreview.net
openreview.net
Discovering Failure Modes in Vision-Language Models using RL
Kanishk Jain
Qian Yang
Shravan Nayak
Parisa Kordjamshidi
Nishanth Anand
Aishwarya Agrawal
Vision-language Models (VLMs), despite achieving strong performance on multimodal benchmarks, often misinterpret straightforward visual conc… (voir plus)epts that humans identify effortlessly, such as counting, spatial reasoning, and viewpoint understanding. Previous studies manually identified these weaknesses and found that they often stem from deficits in specific skills. However, such manual efforts are costly, unscalable, and subject to human bias, which often overlooks subtle details in favor of salient objects, resulting in an incomplete understanding of a model's vulnerabilities. To address these limitations, we propose a Reinforcement Learning (RL)-based framework to automatically discover the failure modes or blind spots of any candidate VLM on a given data distribution without human intervention. Our framework trains a questioner agent that adaptively generates queries based on the candidate VLM's responses to elicit incorrect answers. Our approach increases question complexity by focusing on fine-grained visual details and distinct skill compositions as training progresses, consequently identifying 36 novel failure modes in which VLMs struggle. We demonstrate the broad applicability of our framework by showcasing its generalizability across various model combinations.
2026-04-05
arXiv (prépublication)
doi.org
arxiv.org