Publications

Apertus: Democratizing Open and Compliant LLMs for Global Language Environments
Apertus Team , 2025.

ArXiv

RL for Reasoning by Adaptively Revealing Rationales
Mohammad Hossein Amani , Aryo Lotfi , Nicolas Mario Baldwin , Samy Bengio , Mehrdad Farajtabar , Emmanuel Abbe , Robert West , 2025.

ArXiv | Code

Show description

We proposed an adaptive curriculum-based RL algorithm that reveals partial rationales based on per-sample learning estimation of the value of a state. We demonstrate — both empirically and theoretically — a separation result: on complex reasoning tasks, neither SFT nor vanilla RL generalizes with increasing complexity, while AdaBack enables gradual abstraction learning — without ever fine-tuning on full demonstrations.

Symbolic autoencoding for self-supervised sequence learning
Mohammad Hossein Amani , Nicolas Mario Baldwin , Amin Mansouri , Martin Josifoski , Maxime Peyrard , Robert West , 2024.

ArXiv | Code | Poster | ICML'24 Workshop

Show description

In this project, we used straight-through estimators to optimize over a discrete latent space of sentences—effectively treating language itself as an optimizable hidden variable.

Sharp asymptotics on the compression of two-layer neural networks
Mohammad Hossein Amani , Simone Bombari , Rattana Pukdee , Stefano Rini , Marco Mondelli , 2022.

ArXiv | ITW'22

Memorization and optimization in deep neural networks with minimum over-parameterization
Simone Bombari , Mohammad Hossein Amani , Marco Mondelli , 2022.

ArXiv | Neurips'22