Preprint 2026Diffusion
Representation-Augmented Diffusions
F. M. Sofian*, K. Pandey*, S. Mandt
TL;DR
A new class of diffusion models that are faster to train, faster to sample, and parameter-efficient, with a smooth latent space.
arXiv coming soon
Farrin Marouf Sofian
I'm an early fourth-year Ph.D. student at UC Irvine, advised by Prof. Stephan Mandt and supported by the Chan Zuckerberg Initiative.
I work on generative modeling with diffusion models: designing new foundation models, conditional sampling via test-time steering, and reward-guided fine-tuning, with applications from images and language to science. Previously, I was an AI Resident at the Chan Zuckerberg Initiative and at Adobe.
01
News & Honors
- Sep 2026Paper
Two papers accepted to NeurIPS 2026: Hierarchical Variational Policies for Reward-Guided Diffusion and Control-Augmented Autoregressive Diffusion for Data Assimilation.
- Jan 2026Paper
Parallel Token Prediction for Language Models accepted to ICLR 2026.
- May 2025Paper
Variational Control for Guidance in Diffusion Models accepted to ICML 2025.
- 2025HonorSelected as an AI Resident at the Chan Zuckerberg Initiative.
- Summer 2024Career
Joined Adobe for the summer to work on LLMs.
02
Publications
9 papers
Topic
Year
01
Foundation Model Training
New model classes and training schemes for large diffusion and language models.
Preprint 2026LLMs
Parallel Language Models via Variational Decoding
F. Draxler*, R. C. Geyer*, D. Geyfman*, F. M. Sofian*, J. Will*, et al.
TL;DR
Pre-training and post-training a large language model that decodes many tokens in parallel.
arXiv coming soon
ICLR 2026LLMs
Parallel Token Prediction for Language Models
F. Draxler, J. Will, F. M. Sofian, T. Karaletsos, S. Mandt
TL;DR
Predicts several tokens in one forward pass by moving randomness into the model's inputs, giving a 2.4× speedup in speculative decoding.
* Equal contribution
02
Test-time Scaling & Reward Fine-tuning
Steering pretrained diffusion models toward rewards and constraints — at inference time or through lightweight learned control.
NeurIPS 2026DiffusionInverse Problems
Hierarchical Variational Policies for Reward-Guided Diffusion
K. Pandey*, F. M. Sofian*, J. Groeneveld, F. Draxler, S. Mandt
TL;DR
Amortizes test-time control into a lightweight stochastic policy, matching test-time scaling baselines at a fraction of the compute.
NeurIPS 2026DiffusionAI for Science
Control-Augmented Autoregressive Diffusion for Data Assimilation
P. Srivastava, F. M. Sofian, F. Immorlano, K. Pandey, S. Mandt
TL;DR
Adds an offline-trained controller to autoregressive diffusion models for fast, stable data assimilation in chaotic PDEs and weather.
Preprint 2026Diffusion
Variational Test-time Optimization for Diffusion Synchronization
H. Lee*, F. M. Sofian*, K. Pandey, S. Mandt
TL;DR
Derives diffusion synchronization from optimal control to coordinate multiple diffusion trajectories at test time, with no extra training.
ICML 2025DiffusionInverse Problems
Variational Control for Guidance in Diffusion Models
K. Pandey*, F. M. Sofian*, F. Draxler, T. Karaletsos, S. Mandt
TL;DR
Unifies diffusion guidance as trajectory matching under variational control, reaching state of the art on inverse problems without retraining.
* Equal contribution
03
Earlier Work
Multimodal generation and NLP from my undergraduate research at Koç University.
Preprint 2024DiffusionMultimodal
SonicDiffusion: Audio-Driven Image Generation and Editing with Pretrained Diffusion Models
B. C. Biner, F. M. Sofian, U. B. Karakaş, D. Ceylan, E. Erdem, et al.
TL;DR
Conditions pretrained image diffusion models on audio, enabling sound-driven image generation and editing.
IJCNLP-AACL 2023NLP
GECTurk: Grammatical Error Correction and Detection Dataset for Turkish
A. Kara, F. M. Sofian, A. Bond, G. G. Sahin
TL;DR
The first large-scale dataset and baselines for Turkish grammatical error correction and detection.
03
Get in touch
I'm always happy to talk about research collaborations or new opportunities.
If you're a woman in STEM and would like guidance, a sounding board, or just a chat, please reach out. Supporting other women in the field means a lot to me, wherever you are in your journey.
Email is the best way to reach me.