aryopg.bsky.social
PhD student @BioMedAI_CDT @EdinburghNLP @EdiClinicalNLP | ex: AI Safety Fellow @Anthropic | Opinions are my own. Personal page: https://aryopg.com
👤 Joined on 17 November 2024
Followers
Following
Total Posts
@aryopg.bsky.socialInterested? Please check out the paper and code below for more! Work with @neuralnoise.com @edoardo-ponti.bsky.social and @nolovedeeplearning.bsky.social Arxiv: arxiv.org/abs/2505.13138 Code: github.com/HEmile/neuro...
Show PostOn benchmarks for visual path planning and problems affected by reasoning shortcuts, we achieve state-of-the-art performance while showing scalability and strong calibration!
Show PostWe derive a principled loss function using a new result that may be of broader interest to the masked diffusion community. Our loss function reconstructs the output dimensions separately, decomposing the problem!
Show PostHow to go beyond independence? This is where powerful discrete diffusion models come in! Every unmasking step is an independent distribution… Neurosymbolic Diffusion Models (NeSyDMs) exploit this for efficient probabilistic reasoning while guaranteeing global dependencies 🚀
Show PostClassical NeSy methods assume conditional independence over concepts! But in arxiv.org/abs/2404.08458 (ICML’24) we show many issues with this assumption, from challenging optimisation to poor calibration 👇 x.com/EmilevanKrie...
Show Post