@ShovalMessica

Audio & Speech MSc student at HUJI @cseHUJI

Joined May 2024
  • Tweets18
  • Following87
  • Followers41
1Media
🚨I’m excited to share our #INTERSPEECH2024 paper "NAST: Noise Aware Speech Tokenization for Speech Language Models” 🥳 W/ @adiyossLC Paper-arxiv.org/abs/2406.11037 Code-github.com/ShovalMessica/NAS…
1
14
1
32
4,227
Shoval Messica retweeted
David Bau@davidbauOct 12
Do VLMs reason spatially or verbally? Kelly Cui @nikhil07prakash @ShovalMessica trace binding within VLMs. Is "X to the left of Y" about geometry or words?Interestingly, they see *dual* spatial+verbal mechanisms. Remove one, and the other kicks in. They think *both* ways. ➡️
Nikhil Prakash@nikhil07prakashOct 12
VLMs use pointer-like signals for spatial variable binding. But where does the binding actually happen? We find a dual mechanism: the vision encoder is the primary source of binding, with the LM backbone playing a secondary role.
1
3
57
3,043
Shoval Messica retweeted
Nikhil Prakash@nikhil07prakashOct 12
VLMs use pointer-like signals for spatial variable binding. But where does the binding actually happen? We find a dual mechanism: the vision encoder is the primary source of binding, with the LM backbone playing a secondary role.
3
18
3
73
7,628
Shoval Messica retweeted
Michael Hassid@MichaelHassidApr 29
Have LLMs become supervised learners (once again)?! In our new paper, we argue that current LLMs’ post-training methods have effectively reverted to the "pre-train then fine-tune" era, explicitly tailoring models to desired behaviors. 1/n
🤖 Made with AI
1
20
1
50
2,431
Shoval Messica retweeted
Omri Fahn@OmriFahnFeb 5
MFA is a beautiful visualization, but not only that. It’s a practical tool: competitive localization and better steering, while revealing that concepts live in regions, not just single directions. Grateful I got to contribute to this with an awesome team!
Or Shafran@OrShafranFeb 5
It's time to look past dictionary learning for decomposing LM activations. What happens when we instead leverage local geometry? We find a natural region-based decomposition that yields better steering and localization 🧵 1/
1
1
2
77
Shoval Messica retweeted
Omri Fahn@OmriFahn4 Jun 2025
8/ 📄 Read the full paper: arxiv.org/abs/2505.21218 Thanks for reading! #AI #LLM #Uncertainty #Interpretability #Research

Pretrained LLMs Learn Multiple Types of Uncertainty

Large Language Models are known to capture real-world knowledge, allowing them to excel in many downstream tasks. Despite recent advances, these models are still prone to what are commonly known...

arxiv.org
1
1
84
Shoval Messica retweeted
Omri Fahn@OmriFahn4 Jun 2025
🤔Can an LLM "unconscious" feel a lie bubbling up before it speaks? 🤔And what mysteries hide in its linear space of uncertainty? Thrilled to share our new paper: “Pre-trained LLMs Learn Multiple Types of Uncertainty” (@roicohen9•@OmriFahn•@GerardDeMelo) arxiv.org/abs/2505.21218

Pretrained LLMs Learn Multiple Types of Uncertainty

Large Language Models are known to capture real-world knowledge, allowing them to excel in many downstream tasks. Despite recent advances, these models are still prone to what are commonly known...

arxiv.org
2
3
21
231
Shoval Messica retweeted
Gallil Maimon@GallilMaimon12 Sep 2024
🚨New paper on SLM evaluation🚨 We present SALMon🍣 which is a suite of benchmarks for evaluating how much Speech Language Models model acoustic elements like sentiment or background noise. Project: pages.cs.huji.ac.il/adiyoss-… 🧵👇🏻
2
21
2
54
4,443
Which is better, running a 70B model once, or a 7B model 10 times? The answer might be surprising! Presenting our new @COLM_conf paper: "The Larger the Better? Improved LLM Code-Generation via Budget Reallocation" arxiv.org/abs/2404.00725 1/n
5
43
9
206
31,529
I am excited to share my first work: "Dataset Size Recovery from LoRA Weights". Ever wondered if you could find out how many samples was a model trained on using just its weights? Well now you can! Project: vision.huji.ac.il/dsire/ 👇
3
27
3
51
2,963
Shoval Messica retweeted
Yossi Adi@adiyossLC21 Jun 2024
Speech tokenizers are fundamental in building speech LMs. In a recent study we show the common tokenization method (a.k.a “semantic tokens”) is not robust to different signal variations. We then present NAST! a noise aware speech tokenizer w. @ShovalMessica Code and models 👇
Shoval Messica@ShovalMessica19 Jun 2024
🚨I’m excited to share our #INTERSPEECH2024 paper "NAST: Noise Aware Speech Tokenization for Speech Language Models” 🥳 W/ @adiyossLC Paper-arxiv.org/abs/2406.11037 Code-github.com/ShovalMessica/NAS…
7
37
2,643
Shoval Messica retweeted
Guy Yariv@guy_yariv21 Jun 2024
1/ Commonsense reasoning needs multimodal knowledge, yet current LLMs focus mostly on text, limiting their integration of crucial visual information. We introduce vLMIG, a method that enhances LLMs' visual commonsense by integrating images into the decision-making process
1
24
1
75
10,288
NAST: Noise Aware Speech Tokenization for Speech Language Models. arxiv.org/abs/2406.11037

NAST: Noise Aware Speech Tokenization for Speech Language Models

Speech tokenization is the task of representing speech signals as a sequence of discrete units. Such representations can be later used for various downstream tasks including automatic speech...

arxiv.org
1
3
148
Shoval Messica retweeted
arXiv Sound@ArxivSound18 Jun 2024
``Joint Audio and Symbolic Conditioning for Temporally Controlled Text-to-Music Generation,'' Or Tal, Alon Ziv, Itai Gat, Felix Kreuk, Yossi Adi, ift.tt/tP4Hua8
1
5
17
1,282
Shoval Messica retweeted
Or Tal@Or__Tal18 Jun 2024
Today we released JASCO 🥁🎵🪇! JASCO uses time-aligned conditions (e.g. drums, melody) as guiding controls for text-to-music generation. 📜Full Paper: arxiv.org/abs/2406.10970 🔊Samples: pages.cs.huji.ac.il/adiyoss-… @lonziks @itai_gat @FelixKreuk @adiyossLC Stay tuned for updates.
2
18
6
51
3,447
Shoval Messica retweeted
arXiv Sound@ArxivSound18 Jun 2024
``NAST: Noise Aware Speech Tokenization for Speech Language Models,'' Shoval Messica, Yossi Adi, ift.tt/TwrHMcg
5
23
1,460
Load more