Tim Tomov

Are LLM outputs just language? We often don’t know what the answer to a question will be - but we do know its structure. Numbers, lists, sets, ... Our framework models outputs in their underlying structure for better answers + uncertainty. w/ @dfuchsgruber @guennemann [1/4]🧵

A reminder that you cannot learn from data without making assumptions. You should never apologize for making assumptions. Embrace them, be transparent about what they are, and enabled by those assumptions, update based on new information.

I want to take this post to thank the amazing people of X for reviewing and reading our paper: arxiv.org/pdf/2602.18292… - where we claim that decoding is an optimisation layer on top of LLMs. Thank you for letting us know which references we missed. We are adding those and will update the arXiv soon. So far, the list is: @amritsinghbedi3 Controlled decoding from language models arxiv.org/pdf/2310.17022 @amritsinghbedi3 Transfer Q⋆: Principled Decoding for LM Alignment arxiv.org/pdf/2405.20495 @elon_lit Scaled-Dot-Product Attention as One-Sided Entropic Optimal Transport arxiv.org/abs/2508.08369 @timtomov Task-Awareness Improves LLM Generations and Uncertainty arxiv.org/pdf/2601.21500 Let me know if I missed any! Please keep them coming! Would be happy to cite the missing refs! Also make sure to read those papers above, I think they are super awesome :D #AI #MachineLearning

@hbouammar Looks really interesting! 2 things I will check out are how MBR (minimum bayes risk) decoding fits into this, and thinking about an application of your ideas in a space different from the token space. (We recently proposed MBR decoding in such a space arxiv.org/pdf/2601.21500)

I am happy to announce our paper, where we formalise LLM decoding as a special case of optimisation over a probability simplex. 📜arxiv.org/pdf/2602.18292 This is work in progress! How can you help? 1⃣ We probably missed citations! If you think your work is relevant to this topic, make sure to leave a comment with the paper that we should cite. We will review it and update the arXiv. 2⃣ Try to derive new special cases. We will review that if that works, we are happy to add you to our team! Why not work together? We can do more! #AI #machinelearning

Interesting research from Anthropic: When you have increasingly large models and increasingly complex tasks it's more likely that the models will give you different answers if you run the same query multiple times. On easy tasks, larger models actually become more coherent. Think of a "cone" of possible trajectories and the branching factor gets bigger with more possibilities (due the larger models "knowing more options to explore" and more complex problems having more "possible aspects"). The amount of time reasoning (trajectory length) then makes it multiplicatively more incoherent at the end state. Having a large model with an easy task means the correct answer is definitely "in there" and it's less likely to become distracted. They are arguing this is relevant for AI safety because some might have assumed that larger models would have convergent "instrumental goals" and would give a consistently wrong rather than randomly wrong answer. Apparently the "the hot mess theory of intelligence" (Sohl-Dickstein, 2023) argues that "as entities become more intelligent, their behaviour tends to become more incoherent, and less well described through a single goal."

Better uncertainty. We similarly outperform existing UQ methods across a wide range of tasks. The key: variation ≠ uncertainty. Our Bayes risk is task-aware and distance-sensitive, capturing uncertainty with higher granularity. 📝 arxiv.org/abs/2601.21500 [4/4]🧵

Better answers. The Bayes-optimal answer outperforms other decoding methods across a wide range of tasks. The key: We leverage information from the full predictive distribution and even enable the synthesis of new responses beyond those produced by the model [3/4]🧵

Are LLM outputs just language? We often don’t know what the answer to a question will be - but we do know its structure. Numbers, lists, sets, ... Our framework models outputs in their underlying structure for better answers + uncertainty. w/ @dfuchsgruber @guennemann [1/4]🧵

Wondering how to attend an ML conference the right way? ahead of NeurIPS 2025 (30k attendees!) here are ten pro tips: 1. Your main goals: (i) meet people (ii) regain excitement about work (iii) learn things – in that order. 2. Make a list of papers you like and seek them out at poster sessions. Try to talk to the authors– you can learn much more from them than from a PDF. 3. Pick one workshop and one tutorial that sounds most interesting. Skip the rest. 4. Cold email people you want to meet but haven't. Check Twitter and the accepted papers list. PhD students are especially responsive. 5. Practice a concise pitch of unpublished research you're working on for "what are you interested in rn?". Focus on big unanswered questions and exciting new directions, *not* papers. 6. Skip the orals. Posters are a higher-bandwidth, more engaging, more invigorating. Orals are a good time to go for a walk or talk in the hallway. 7. for the love of god, do NOT work on other research in your hotel room. Save mental bandwidth for the conference. (This may seem obvious; you'd be surprised.) 8. Talk to people outside your area. There are many smart people working on niches <10 people understand. Learn about one or two that won't help your own work. 9. Attend one social each night. Don't overthink it or get caught up in status games. They're all fun. 10. Take breaks. You can't go to everything, and conferences consume more energy than a normal workweek. hope this helps, and sad i'm not attending neurips, have fun :)

Excited to share that my first PhD paper, which introduces a new diffusion model for relational databases, has been accepted to #NeurIPS2025! We will be presenting it this week in San Diego. ☀️🌴 Joint work with @ludke_david, @SchwinnLeo, and @guennemann. 🧵 1/

@guerrantif @MarcelKollovieh @nifleisch @Bertrand_Charp @guennemann Very cool work to extend them to hierarchies !

1️⃣ #TreeGen: A Bayesian Generative Model for Hierarchies We learn full distributions over hierarchies with an application to jet clustering. 📍 Main Conf (Hall C/D/E) | Poster #3715 🕒 Thu 4 Dec, 11:00–14:00 PST (w/ @MarcelKollovieh @nifleisch @Bertrand_Charp @guennemann)

Heading to San Diego for #NeurIPS2025! 🌴☀️ I’ll be presenting 3 recent papers covering generative models for hierarchies, spatiotemporal tissue dynamics, and long-range graph learning. If you're around, drop by and say hi! 👋 Here’s the schedule 🧵👇

Found this great post by @alemi: blog.alexalemi.com/kl-is-all-you-…. It shows how many ML methods boil down to minimising the KL between the joint of the real causal world and the world we want (e.g., real images → latents, but we want latents → images (VAE))

No Free Lunch Theorem: en.wikipedia.org/wiki/No_free_l…

Our paper “Modeling Microenvironment Trajectories on Spatial Transcriptomics with NicheFlow” has been accepted at #NeurIPS2025! @ale__palmaa @guerrantif @fabian_theis @guennemann (@TU_Muenchen × @HelmholtzMunich × @MunichCenterML × @MdsiTum ) 🧵

These results reveal a deeper issue: current uncertainty estimators only work when the world is unambiguous. But language never is. To make uncertainty reliable, we need to rethink current paradigms. arxiv.org/abs/2511.04418 [4/4]🧵

Why do these estimators seem to work? For single answers, the true distribution is one-hot. We theoretically show that entropy and mutual information track model uncertainty (EU). Once ambiguity enters, the truth can lie anywhere on the simplex — and the link breaks. [3/4]🧵

Can we actually tell when LLMs know or don’t know? For questions with a single answer, that works. But once ambiguity enters - several answers are correct - current methods collapse, confusing model with data uncertainty. w/ @dfuchsgruber @TomWollschlager @guennemann [1/4]🧵

Open-source Diffusion LLMs easily break GPT-5! In “Diffusion LLMs are Natural Adversaries for any LLM,” we show that Inpainting on Diffusion LLMs yields efficient, transferable jailbreaks without any model access. @TomWollschlager, Paul Ungermann, @guennemann, @leoschwinn 🧵