Alireza Mousavi

CS PhD student @UofT and @VectorInst. Ex research intern @Apple ML Research. Interested in deep learning theory and generative modeling.

[Alireza Mousavi @alirezamh_]

Can RL with outcome rewards alone just efficiently explore outside the support of the base model and learn completely new capabilities?

We study it in a theoretically tractable setup, and prove that outcome rewards are not enough, but process rewards can get you there! [1/3]

[Alireza Mousavi @alirezamh_]

I’m at ICML 🇰🇷 presenting a 🔦 spotlight 🔦 poster today. Check it out if you want to know how process rewards help RLVR go beyond the base model support when outcome rewards alone can’t.

📆 Thu. 5-6:45 pm. Hall A. Poster # 1307.

[Lily Goli @lily_goli]

🚀 🚀 🚀 Excited to share our new paper:

Remember to be Curious: Episodic Context and Persistent Worlds for 3D Exploration

What does it take for an agent to stay curious in a 3D world?
The answer is memory.

🌐 Project: https://t.co/G4SjLoFJht
📄 Paper: https://t.co/iUFwp5NvRu
💻 Code: https://t.co/KZRaQLyzyh

[Alireza Mousavi @alirezamh_]

Under the spotlight at ICML 🔦 Take a look at what theory tells us about RLVR and base models.

[Alireza Mousavi @alirezamh_]

Can RL with outcome rewards alone just efficiently explore outside the support of the base model and learn completely new capabilities?

We study it in a theoretically tractable setup, and prove that outcome rewards are not enough, but process rewards can get you there! [1/3]

[Alireza Mousavi @alirezamh_]

How do you choose which mixer to use at each layer? Why not use all of them? Check out this work.

[ServiceNow AI Research @ServiceNowRSRCH]

(1/8)
🚀 Introducing Super Apriel: One Checkpoint, Many Speeds

Train once → serve at any speed-quality tradeoff

We release:
✓ 15B supernets with 4 mixers/layer
✓ Training code (Fast-LLM)
✓ vLLM serving extension

🧵 How it works ↓

[Alberto Bietti @albertobietti]

Attention and MLP matrices can store as much knowledge as # params, but in practice, finite samples/compute and long context noise can further limit this capacity. We characterize these trade-offs in a needle-in-a-haystack model.

#ICLR2026 poster Thurs, led by @nurimertvural45

[Lily Goli @lily_goli]

Excited to share @YaldaForoutan and @pekzta4's work on robust 3D reconstruction from 360 captures!
Casual 360 video capture gives you the coverage you need for calibration and reconstruction; FullCircle makes it hassle-free!
No more wrestling with stitching artifacts or floaters from the camera operator. Data and code are released, enjoy!

[Alireza Mousavi @alirezamh_]

@thegautamkamath Wouldn’t the authors ultimately take a reputational hit once their paper is public (e.g. arxiv)?

[Alireza Mousavi @alirezamh_]

This is no longer an issue if you use process rewards. Number of reward queries will depend on a “token-level likelihood quantile” that is never exponentially small in sequence length. [3/3]

📄 arXiv: https://t.co/PwdlP3oRfP

[Alireza Mousavi @alirezamh_]

Can RL with outcome rewards alone just efficiently explore outside the support of the base model and learn completely new capabilities?

We study it in a theoretically tractable setup, and prove that outcome rewards are not enough, but process rewards can get you there! [1/3]

[Alireza Mousavi @alirezamh_]

We prove the number of reward queries in RLVR depends on the “likelihood quantile” function of the base model.

Unfortunately, LQ becomes exponentially small if we want to go significantly below the base model error, then RL needs exponentially many reward queries. [2/3]

[Michael Kirchhof @mkirchhof_]

New paper 🥳 RL relies a lot on an agent’s capability to explore. Our strategy-guided exploration makes the agent find new solutions more efficiently. It learns faster, and in some environments its Pass@1 surpasses the base model’s Pass@128. 🧵1/6

📄 https://t.co/IZEyeMQG9g

[VAIBHAV SINGH @VAIBHAV22155287]

Masked Diffusion LMs (MDLMs) are the most exciting paradigm shift in AR generation because they can decode in parallel, infill, and self-correct.

But they are bottlenecked by the transformer's quadratic attention, making throughput fall apart for long contexts.

We offer a simple solution. Introducing DiffuMamba: first diffusion LM with a bidirectional Mamba backbone. Better quality. Up to 8.2x faster.

🧵1/N

[Andrea Montanari @Andrea__M]

Fantastic workshop on High-Dimensional Learning Dynamics at BIRS. All talks here:
https://t.co/7LCE3pa0J2

[Alireza Mousavi @alirezamh_]

@konglk1203 Congrats to you as well, nice work! See you hopefully at ICLR :D

[Alireza Mousavi @alirezamh_]

Now accepted at ICLR 2026. Check it out if you’re a diffusion/flow/OT enjoyer.

[Alireza Mousavi @alirezamh_]

Super excited to share what @stephenz_y and I’ve been up to during our internship at🍎:

Using optimal transport makes flows straighter and generation faster in flow matching, but small batch OT is biased and large batch OT is slow.

What to do? Use semidiscrete OT!
🧵

[Alireza Mousavi @alirezamh_]

@mufan_li @roydanroy @miniapeur @jasondeanlee @andrewgwils @OmarRivasplata @deepcohen @alex_damian_ I also liked this one https://t.co/i9qP7zCoZU

[stephen zhang @stephenz_y]

- Sat Dec 6 @ SPIGM workshop, On Fitting Flow Models with Large Sinkhorn Couplings, work with @alirezamh_ , Michal Klein, Marco Cuturi 4/4

[Michael Kirchhof @mkirchhof_]

Our research team is hiring PhD interns 🍏 Spend your next summer in Paris and explore the next frontiers of LLMs for uncertainty quantification, calibration, RL and post-training, and Bayesian experimental design.

Details & Application ➡️ https://t.co/YHXWPhJ5K7

[Alireza Mousavi @alirezamh_]

For my usual ML theory math questions Gemini 3 doesn't feel different from GPT 5.1, both can answer small very well-defined questions and fail otherwise. Maybe the 2x improvement on ARC-AGI-2 isn't enough to replace mathematicians yet 🤔

[Statistics (Machine Learning) Papers @StatsPapers]

From Information to Generative Exponent: Learning Rate Induces Phase Transitions in SGD. https://t.co/klMuw7Yd5H