iiinova — where your digital self keeps evolving

Oscar Mañas

@oscmansan.bsky.social

180 followers 172 following 9 posts

Research scientist at Meta, PhD from Mila and Université de Montréal. Working on multimodal vision+language generation. Català a Zúric.

Posts Replies Media

Oscar Mañas @oscmansan.bsky.social · 10/06/2025

Headed to @cvprconference.bsky.social in Nashville! I'll be presenting our work on Multimodal Reward-guided Decoding. Let's connect if you're around!

020

Oscar Mañas @oscmansan.bsky.social · 15/05/2025

TFW you find a memory leak in your code two days before the rebuttal's deadline

010

Oscar Mañas @oscmansan.bsky.social · 22/04/2025

Heading to Singapore for the next 1.5 weeks for @iclr-conf.bsky.social. If you're around and want to meet up, hit me up!

010

Reposted by Oscar Mañas

Simons Institute for the Theory of Computing @simonsinstitute.bsky.social · 01/04/2025

"Tokenize Everything!" Luke Zettlemoyer of @uofwa.bsky.social on using GPT-like autoregressive techniques for training multimodal models (text, images, audio etc.) at the Simons Institute workshop on The Future of Language Models and Transformers simons.berkeley.edu/workshops/fu...

031

Oscar Mañas @oscmansan.bsky.social · 22/12/2024

I quite like this analogy by Oriol Vinyals: * LLM ~= core electric brain * Agent ~= LLM with a digital body youtu.be/78mEYaztGaw

youtu.be

Gemini 2.0 and the evolution of agentic AI with Oriol Vinyals

YouTube video by Google DeepMind

030

Oscar Mañas @oscmansan.bsky.social · 14/12/2024

Curious about how to effectively steer the behavior of multimodal LLMs during inference to improve their visual grounding? Join me today at 4:30pm at the AFM workshop at @NeurIPSConf, where I'll be presenting a poster on my work. Come by to learn more! openreview.net/forum?id=VWJ...

031

Oscar Mañas @oscmansan.bsky.social · 10/12/2024

Tomorrow at 3:15pm I'll be presenting my work at @mila-quebec.bsky.social's booth (#104) at @neuripsconf.bsky.social. Come to learn more about controlling multimodal LLMs via reward-guided decoding! 🔗 openreview.net/forum?id=VWJ...

openreview.net

Controlling Multimodal LLMs via Reward-guided Decoding

As Multimodal Large Language Models (MLLMs) gain widespread applicability, it is becoming increasingly desirable to adapt them for diverse user needs. In this paper, we study the adaptation of...

0113

Reposted by Oscar Mañas

François Fleuret @francois.fleuret.org · 05/12/2024

All this being said, Meta/FAIR remains the only place where you can do open AI research with a group of stellar colleagues ten times larger than any university + big-tech computational capabilities level.

2391