Georgios Batzolis

I’m a Research Scientist at Devotion AI Labs. Before that, I spent several years at Cambridge, completing my PhD with Prof. Carola-Bibiane Schönlieb in the Department of Applied Mathematics and Theoretical Physics (DAMTP), then working with Prof. Mark Girolami as a Postdoctoral Research Associate in the Department of Engineering.

My current research focuses on diffusion language models (DLMs) and closing the capability gap with autoregressive language models, commonly known as LLMs. In CoBit (Continuous Bitstream Diffusion), we represent text as bitstreams and model it with continuous diffusion. At the time of our arXiv submission, our GenPPL–entropy and GSM8K results suggested an advance in the state of the art among continuous DLMs on those evaluations. The field has moved on rapidly since, so this should be read as a historical result.

The collective progress of the DLM research community makes me optimistic that the capability gap could close sooner than we expect. Matching strong autoregressive models across a broad range of tasks remains an open challenge, but the momentum is exciting. I look forward to sharing new work soon and contributing to that progress.

As the capability gap narrows, I expect speed and inference cost to become increasingly decisive. I’m excited by the prospect of DLMs competing as standalone models in production, and by their potential as parallel draft models within speculative decoding, proposing tokens for an autoregressive model to verify. Both directions offer opportunities for faster generation, provided the gains hold up in practical workloads without compromising quality.

My interest in DLMs also connects to my earlier work on the geometry of continuous data. There, I have studied how trained diffusion models and normalising flows can define Riemannian metrics that reveal intrinsic structure and give meaningful notions of distance and interpolation. I’m particularly interested in how this perspective might extend to discrete distributions, including text and DNA sequences. What notions of geometry are useful for these data, and how can they guide the choice of representations and the design of more efficient diffusion models?

At NeurIPS 2026: I will be presenting CoBit in Sydney and would be happy to meet others working on DLMs. If you will be there, feel free to get in touch.

Research interests

• Language and multimodal generation: designing unified generative models that capture relationships across modalities and support both joint and conditional generation. Our multimodal CoBit work explores one approach.
• Fast, capable diffusion models: accelerating inference through caching, parallel generation, and distillation, while understanding the trade-offs between speed and quality.
• Information theory and optimal sampling: understanding what makes denoising difficult and designing sampling algorithms with guarantees on accuracy and computational cost.
• Learning the geometry of data: constructing Riemannian metrics from trained generative models and extending geometric methods from continuous data to text and biological sequences.

If any of these topics interest you, feel free to reach out for a chat at g.batz97@gmail.com or Georgios@devotionlabs.ai.

Publications

arXiv Preprints

* Equal contribution.

Contact

Email: g.batz97@gmail.com or Georgios@devotionlabs.ai. You can also find me on Google Scholar or LinkedIn.