Masoud Jalili Sabet

I am a Research Scientist in the AI Research Agents (AIRA) Lab at FAIR, Meta. My current work focuses on ML world models, recursive self-improvement, and rubric-based agentic judges. I am also a core contributor to Muse Glimmer 30B, an open-weight agentic model. I developed a synthetic agentic data generation pipeline used for its mid-training and post-training.

Previously, I was an NLP Research Lead in the AI Innovation team at Volkswagen, where I built production RAG and agentic systems, multimodal pipelines, and LLM evaluation tools. I completed my PhD in CIS at LMU Munich under the supervision of Hinrich Schütze. During my doctoral studies, I interned at Meta AI and worked on end-to-end training of RAG models for open-domain question answering.

My broader research interests include AI research agents and ML world models, explainability for LLMs/VLMs, multimodal models, and multilingual NLP.

Prior to LMU, I received my master’s degree in Computer Engineering at University of Tehran, and my bachelor’s degree at IUST. During my master’s, I interned with the MT group at FBK and had the privilege of working with Marco Turchi and Matteo Negri.

I’m happy to talk about research and my work. If you have questions about one of my papers or my broader research, feel free to reach out!

news

Aug 14, 2026	Our paper AI Research Preference Models is now available on arXiv.
Jun 05, 2024	Spotlight paper at CVPR 2024 (T4V) in collaboration with the Torr Vision Group.
Jul 25, 2023	Our paper Glot500 got accepted at ACL 2023 and was selected for Area Chair Award.
Dec 01, 2022	Our paper got accepted at EMNLP 2022.
Jul 18, 2022	I just defended my PhD. #PhDone

selected publications

In prep.

Muse Glimmer 30B: An On-Device Agentic Model

Muse Glimmer Team

2026

Core contributor: Masoud Jalili Sabet. Manuscript in preparation.

Bib

@unpublished{muse-glimmer-2026,
  title = {Muse Glimmer 30B: An On-Device Agentic Model},
  author = {{Muse Glimmer Team}},
  year = {2026},
  note = {Core contributor: Masoud Jalili Sabet. Manuscript in preparation.},
}

arXiv

AI Research Preference Models

Thomas Simon Foster , Bassel Al Omari , Tingchen Fu , Thomas Mann , Carl Domond , and 28 more authors

arXiv, 2026

Abs arXiv Bib PDF

AI research agents can carry machine learning experiments from proposal through implementation and evaluation, but frontier evaluations can require days of GPU time. AI Research Preference Models predict which candidate solutions are most promising under a fixed execution budget. Integrated into the AIRA-dojo research agent and evaluated on AIRS-Bench, the proposed inference-only and agentic models improve normalized performance while reducing the execution budget required to reach the unguided agent’s 24-hour performance.

@article{foster-etal-2026-ai-research-preference-models,
  title = {AI Research Preference Models},
  author = {Foster, Thomas Simon and Al Omari, Bassel and Fu, Tingchen and Mann, Thomas and Domond, Carl and Cipolina-Kun, Lucia and Gauri, Bhavul and Aghamelu, Muna and Goldie, Alexander D. and Helenowski, Eryk and Gagnon-Audet, Jean-Christophe and Pepe, Alberto and Nazir, Saba and Izcovich, Daniel and Levi, Noam and Hazra, Rishi and Hambardzumyan, Karen and Baldwin, Nicolas and Li, Xian and Josifoski, Martin and Giampouras, Paris and Jalili Sabet, Masoud and Sims, Anya and Momand, Hela and Shavrina, Tatiana and Magka, Despoina and Weston, Jason and Wang, Yulin and Goyal, Anirudh and Henriques, Jo{\~a}o and Bachrach, Yoram and McMilin, Emily and Foerster, Jakob Nicolaus},
  journal = {arXiv},
  year = {2026},
  url = {https://arxiv.org/abs/2608.13940},
}

T4V@CVPR

Learning Visual Prompts for Guiding the Attention of Vision Transformers

Razieh Rezaei* , Masoud Jalili Sabet*, Jindong Gu , Daniel Rueckert , Philip Torr , and 1 more author

In Transformers for Vision (T4V) Workshop at CVPR 24 , 2024

Bib PDF

@inproceedings{rezaei2024learningvisualpromptsguiding,
  title = {Learning Visual Prompts for Guiding the Attention of Vision Transformers},
  author = {Rezaei*, Razieh and Jalili Sabet*, Masoud and Gu, Jindong and Rueckert, Daniel and Torr, Philip and Khakzar, Ashkan},
  year = {2024},
  archiveprefix = {arXiv},
  booktitle = {Transformers for Vision (T4V) Workshop at CVPR 24},
  url = {https://arxiv.org/abs/2406.03303},
}

ACL

Glot500: Scaling Multilingual Corpora and Language Models to 500 Languages

Ayyoob ImaniGooghari , Peiqin Lin , Amir Hossein Kargaran , Silvia Severini , Masoud Jalili Sabet, and 6 more authors

In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , Jul 2023

Bib PDF

@inproceedings{imanigooghari-etal-2023-glot500,
  title = {Glot500: Scaling Multilingual Corpora and Language Models to 500 Languages},
  author = {ImaniGooghari, Ayyoob and Lin, Peiqin and Kargaran, Amir Hossein and Severini, Silvia and Jalili Sabet, Masoud and Kassner, Nora and Ma, Chunlan and Schmid, Helmut and Martins, Andr{\'e} and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
  booktitle = {Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  month = jul,
  year = {2023},
  publisher = {Association for Computational Linguistics},
  url = {https://aclanthology.org/2023.acl-long.61},
  bib = {https://aclanthology.org/2023.acl-long.61.bib},
  doi = {10.18653/v1/2023.acl-long.61}
}

ACL
Graph Neural Networks for Multiparallel Word Alignment

Ayyoob Imani , Lütfi Kerem Senel , Masoud Jalili Sabet, François Yvon , and Hinrich Schuetze

In Findings of the Association for Computational Linguistics: ACL 2022 , May 2022

Abs Bib PDF

After a period of decrease, interest in word alignments is increasing again for their usefulness in domains such as typological research, cross-lingual annotation projection and machine translation. Generally, alignment algorithms only use bitext and do not make use of the fact that many parallel corpora are multiparallel. Here, we compute high-quality word alignments between multiple language pairs by considering all language pairs together. First, we create a multiparallel word alignment graph, joining all bilingual word alignment pairs in one graph. Next, we use graph neural networks (GNNs) to exploit the graph structure. Our GNN approach (i) utilizes information about the meaning, position and language of the input words, (ii) incorporates information from multiple parallel sentences, (iii) adds and removes edges from the initial alignments, and (iv) yields a prediction model that can generalize beyond the training sentences. We show that community detection algorithms can provide valuable information for multiparallel word alignment. Our method outperforms previous work on three word alignment datasets and on a downstream task.
@inproceedings{imani-etal-2022-graph, title = {Graph Neural Networks for Multiparallel Word Alignment}, author = {Imani, Ayyoob and Senel, L{\"u}tfi Kerem and Jalili Sabet, Masoud and Yvon, Fran{\c{c}}ois and Schuetze, Hinrich}, booktitle = {Findings of the Association for Computational Linguistics: ACL 2022}, month = may, year = {2022}, address = {Dublin, Ireland}, publisher = {Association for Computational Linguistics}, url = {https://aclanthology.org/2022.findings-acl.108}, bib = {https://aclanthology.org/2022.findings-acl.108.bib}, doi = {10.18653/v1/2022.findings-acl.108}, pages = {1384--1396} }
EMNLP
Graph Algorithms for Multiparallel Word Alignment

Ayyoob Imani* , Masoud Jalili Sabet*, Lutfi Kerem Senel , Philipp Dufter , François Yvon , and 1 more author

In EMNLP , May 2021

Abs Bib PDF

With the advent of end-to-end deep learning approaches in machine translation, interest in word alignments initially decreased; however, they have again become a focus of research more recently. Alignments are useful for typological research, transferring formatting like markup to translated texts, and can be used in the decoding of machine translation systems. At the same time, massively multilingual processing is becoming an important NLP scenario, and pretrained language and machine translation models that are truly multilingual are proposed. However, most alignment algorithms rely on bitexts only and do not leverage the fact that many parallel corpora are multiparallel. In this work, we exploit the multiparallelity of corpora by representing an initial set of bilingual alignments as a graph and then predicting additional edges in the graph. We present two graph algorithms for edge prediction: one inspired by recommender systems and one based on network link prediction. Our experimental results show absolute improvements in F1 of up to 28% over the baseline bilingual word aligner in different datasets.
@inproceedings{imanigooghari-etal-2021-graph, title = {Graph Algorithms for Multiparallel Word Alignment}, author = {Imani*, Ayyoob and Jalili Sabet*, Masoud and Senel, Lutfi Kerem and Dufter, Philipp and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich}, booktitle = {EMNLP}, year = {2021}, publisher = {Association for Computational Linguistics}, url = {https://aclanthology.org/2021.emnlp-main.665}, bib = {https://aclanthology.org/2021.emnlp-main.665.bib}, doi = {10.18653/v1/2021.emnlp-main.665}, pages = {8457--8469}, }
EMNLP
SimAlign: High Quality Word Alignments Without Parallel Training Data Using Static and Contextualized Embeddings

Masoud Jalili Sabet*, Philipp Dufter* , François Yvon , and Hinrich Schütze

In EMNLP Findings , May 2020

Abs Bib PDF

Word alignments are useful for tasks like statistical and neural machine translation (NMT) and cross-lingual annotation projection. Statistical word aligners perform well, as do methods that extract alignments jointly with translations in NMT. However, most approaches require parallel training data and quality decreases as less training data is available. We propose word alignment methods that require no parallel data. The key idea is to leverage multilingual word embeddings – both static and contextualized – for word alignment. Our multilingual embeddings are created from monolingual data only without relying on any parallel data or dictionaries. We find that alignments created from embeddings are superior for four and comparable for two language pairs compared to those produced by traditional statistical aligners – even with abundant parallel data; e.g., contextualized embeddings achieve a word alignment F1 for English-German that is 5 percentage points higher than eflomal, a high-quality statistical aligner, trained on 100k parallel sentences.
@inproceedings{jalili-sabet-etal-2020-simalign, title = {{S}im{A}lign: High Quality Word Alignments Without Parallel Training Data Using Static and Contextualized Embeddings}, author = {Jalili Sabet*, Masoud and Dufter*, Philipp and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich}, booktitle = {EMNLP Findings}, year = {2020}, publisher = {Association for Computational Linguistics}, url = {https://aclanthology.org/2020.findings-emnlp.147}, bib = {https://aclanthology.org/2020.findings-emnlp.147.bib}, doi = {10.18653/v1/2020.findings-emnlp.147}, pages = {1627--1643}, }