Giosue Migliorini

LLMs are autoregressive and slow? No! Parallel Token Prediction decodes multiple consistent tokens in one model call. PTP allows arbitrary dependencies in one call, unlike discrete diffusion. Practical: 2.4x speedup github.com/mandt-lab/ptp ICLR: Apr 23, morning poster P3-#608

@JIRIGESI Hi Jiri, I am a fourth year PhD candidate at UCI interested in probabilistic modeling, RL, and multi modal generative models. I’d love to grab a coffee if you are available!

I’ll be at NeurIPS, if you’re interested in a 2026 PhD research internship with Amazon Store Foundation AI and want to work on agents, RL, and multi-modal, I’d love to connect at the conference.

Efficient Inference for Coupled Hidden Markov Models in Continuous Time and Discrete Space. arxiv.org/abs/2510.12916

What a great event at @AIHealthMIT , happy to be a part of #MoML2025 and thanks to the organizers!

@cloneofsimo Ideally we should sample with replacement from the dataset (potentially repeated datapoints in a single batch!). Epochs & not shuffling introduce periodic behaviors

Guys do we REALLY need to shuffle at the end of epoch? like REALLY REALLY ?

@jxmnop Amazing. It is probably a combination of all of these things.

posted the other day about model distillation. pretty much everyone responded with their theories professors, leading lab researchers, students, pseudoanonymous anime-profile posters seems there's no clear consensus why it works, but here are the theories 🧵

@historyinmemes This has been debunked. youtu.be/ikg3-GQLg3g?si…

There were about 180 towers in Bologna in the 12th century. The tallest, 97 meters high, still stands.

We often think of an "equilibrium" as something standing still, like a scale in perfect balance. But many equilibria are dynamic, like a flowing river which is never changing—yet never standing still. These dynamic equilibria are nicely described by so-called "detailed balance"

Bregman divergences are convex distance-like functionals that are locally Euclidean. Most algorithms handling Euclidean distances generalize to Bregman divergences. en.wikipedia.org/wiki/Bregman_d…

Future be like tab tab tab

@BalintMucsanyi @mkirchhof_ @coallaoh Mine might be a naive question. Why try to measure uncertainty through the predictive entropy, when the predictive variance admits such a nice and interpretable decomposition? I have never seen a confidence interval based on entropy.

If you've been using formulas like this to split aleatoric from epistemic uncertainty: They don't work. To find out why, come visit our poster #68 at the #ICML2024 #SPIGM workshop at 15:10! @mkirchhof_ @coallaoh

@PreetumNakkiran Flow matching should recover the optimal vector field in the Benamou-Brenier perspective of optimal transport if data is sampled from the optimal coupling (in the static ot problem). In that case it would recover the identity

easiest way to see that Flow Matching does not always produce an optimal transport: observe that the marginal flow from a distribution *to itself* is not the Identity (eg for linear flows & independent coupling)

Dynamic Conditional Optimal Transport through Simulation-Free Flows. arxiv.org/abs/2404.04240

Have you ever done a dense grid search over neural network hyperparameters? Like a *really dense* grid search? It looks like this (!!). Blueish colors correspond to hyperparameters for which training converges, redish colors to hyperparameters for which training diverges.

Introducing Voicebox, a new breakthrough generative speech system based on Flow Matching, a new method proposed by Meta AI. It can synthesize speech across six languages, perform noise removal, edit content, transfer audio style & more. More details on this work & examples ⬇️

Functional Flow Matching. (arXiv:2305.17209v1 [cs.LG]) ift.tt/BmXJcbn