Ayhan Suleymanzade

On my way to #ICLR2026 🇧🇷✈️ Hmu if you want to chat about latent/continuous reasoning and flow/diffusion language models. I’ll be presenting #MUX:
→ compress reasoning into continuous latent space
→ multiplex multiple reasoning paths
→ fewer tokens, better reasoning

🧠We introduce "Generative Recursive Reasoning"! Recursive Reasoning Models like HRM, TRM, and Looped Transformers are deterministic — same input, same reasoning, every time. They collapse the entire space of plausible reasoning paths into a single attractor. Our model GRAM (Generative Recursive reAsoning Models) turns recursion itself into a stochastic latent trajectory. Multiple hypotheses, alternative solution strategies, and inference-time scaling not just by depth, but by width — parallel trajectory sampling. And here's the kicker: the same formulation that gives us conditional reasoning p(y|x) also makes GRAM a general generative model p(x). With only 10M params: • Sudoku-Extreme: 97.0% (TRM 87.4%) • ARC-AGI-1: 52.0% • ARC-AGI-2: 11.1% • N-Queens coverage: 90%+ 📄 Paper: arxiv.org/abs/2605.19376 🌐 Project page: ahn-ml.github.io/gram-website w/ Junyeob Baek @JunyeobB (KAIST), Mingyu Jo @pyross0000 (KAIST), Minsu Kim @minsuuukim (KAIST & Mila), Mengye Ren @mengyer (NYU), Yoshua Bengio @Yoshua_Bengio (Mila), Sungjin Ahn @SungjinAhn_ (KAIST)

This is extremely important for anyone interested in language diffusion. We need better metric to move forward!

Pretraining diffusion/flow-matching is simple and scales: noise a clean sample, regress against a closed-form target. RL post-training can be made just as simple. Reinforce Adjoint Matching (RAM) matches Flow-GRPO on SD3.5M in 50× fewer steps. Highest reward, no reward hacking.

🎉 Our paper "Infinite Mask Diffusion for Few-Step Distillation" is accepted to #ICML2026 🇰🇷! We prove MDMs have a structural lower bound on factorization error from using a single deterministic mask, then break it with stochastic infinite-state masks (IMDM). With proper distillation, IMDM beats MDM-based few-step baselines on LM1B and OpenWebText 🚀

Can LLMs help relational learning while avoiding high inference cost? Introducing RelAgent: an LLM agent that searches over SQL feature programs and classical ML models, then deploys the predictor without further LLM calls. With @louistichelman @jw9730 @kolejnyyyy @ismaililkanc

Now that the NeurIPS deadline is over, consider submitting your work to the ICML 2026 Workshop on Graph Foundation Models (GFM)! We would love to see more exciting work from the community! 🗓️ Submission Deadline: May 8, 2026 AoE

@LucaAmb that's why continuous flows will beat autoregression and discrete diffusion ;)

I'm at AISTATS in Tangier this week to present PJAX, our framework for training neural networks using projections instead of gradients. Happy to see you at the spotlight presentation tomorrow, or chat at the poster. Arxiv: arxiv.org/pdf/2506.05878

i'll be talking about our new work on flow map language models (arxiv.org/abs/2602.16813) at the GRaM workshop at ICLR (gram-workshop.github.io) from 15:15 – 15:45 in room 101 A on sunday come check it out if you're interested in the future of language modeling

Super excited to be at @iclr_conf in Rio! I'll be presenting "Topological Flow Matching" in collaboration with the amazing @ismaililkanc and @AlexanderTong7. We improve flow matching performance for modelling signals on graphs and simplicial complexes by aligning sample paths with heat diffusion. Find out more at the poster! 🗓️ Friday, April 24, 2026 ​🕒 3:15 PM - 5:45 PM BRT ​📍 Pavilion 3 · Poster P3-#820

Join us at #AITHYRA. It's a fantastic research environment. If you are considering applying for the AI/ML PI position and will be attending #ICLR2026, send me a message. I’d love to meet up and answer any questions you have!

This work provides theoretical grounding for some of the design decisions (cross-entropy loss, learnable embeddings, self-conditioning, entropy-based schedule) in CDCD (arxiv.org/abs/2211.15089), and brings it into the modern era. Continuous text diffusion is still competitive!

We are seeking a highly motivated postdoctoral researcher to work on fundamental challenges toward AGI, particularly in reasoning, abstraction, and world modeling. The position also offers potential opportunities for co-advising with Yoshua Bengio (Mila) and/or Mengye Ren (NYU). Research areas include: • World Model Learning & Planning • Compositional Generalization & Neuro-Symbolic World Learning • Causal Discovery, Reasoning, and Abstraction This position is supported by the InnoCORE Fellowship Program 2026, with: • Competitive salary of KRW 90M+ (~USD 60K+) • Renewable yearly contract For more information and recent publications: mlml.kaist.ac.kr If you are interested, please send me your CV by email.

Great presentation by @nmboffi and @FEijkelboom about discrete diffusion models that operate in continuous space. This seems most promising for discrete data in the long term 1. can represent uncertainty over tokens 2. generalized genration order youtu.be/OGIFBwh0KGE

Reading group tomorrow: One-step Language Modeling via Continuous Denoising arxiv.org/abs/2602.16813 With @nmboffi 🤗 Join us on zoom Monday at 9am PT / 12pm ET / 6pm CEST: hannes-stark.com/starkly-speaki…

AITHYRA is recruiting an AI/ML Starting PI! Come join us in Vienna! 🇦🇹 Couldn't ask for a better place to do AI/ML research. Don't hesitate to reach out if you're thinking of applying and have questions! Application Deadline: 30 April 2026 More info: aithyra.at/fileadmin/down…

A vibrant Novruz celebration organized by Azerbaijanis in Austria together both locals, internationals

@kpetrovvic @mmbronstein @ismaililkanc @AlexanderTong7 Welcome 🙏

Excited to share that I joined AITHYRA in Vienna 🚀 to work with @mmbronstein @ismaililkanc @AlexanderTong7

Mercury 2 is live 🚀🚀 The world’s first reasoning diffusion LLM, delivering 5x faster performance than leading speed-optimized LLMs. Watching the team turn years of research into a real product never gets old, and I’m incredibly proud of what we’ve built. We’re just getting started on what diffusion can do for language.