Sahil Verma
@Sahil1V
Followers
586
Following
6K
Media
17
Statuses
583
PhD student @uwcse. Robustness and Interpretability. Currently at @MSFTResearch. Former intern at @amazon, @itsArthurAI. Undergrad @IITKanpur
Seattle, WA
Joined September 2013
Glad to share that our paper was accepted the main EMNLP 2025 Conference!.
๐จ New Paper! ๐จ.Guard models slow, language-specific, and modality-limited?. Meet OmniGuard that detects harmful prompts across multiple languages & modalities all using one approach with SOTA performance in all 3 modalities!! while being 120X faster ๐.
4
7
69
RT @soumyesinghal: Llama Nemotron model just got Super-Charged โก๏ธWe released Llama-Nemotron-Super-v1.5 today! The best open model that canโฆ.
huggingface.co
0
7
0
RT @_shruti_joshi_: I will be at the Actionable Interpretability Workshop (@ActInterp, #ICML) presenting *SSAEs* in the East Ballroom A froโฆ.
0
6
0
RT @fengyao1909: ๐ตโ๐ซ Struggling with ๐๐ข๐ง๐-๐ญ๐ฎ๐ง๐ข๐ง๐ ๐๐จ๐?. Meet ๐๐๐ง๐ฌ๐๐๐ข๐ฑ๐๐ซ โ an MoE post-training method that offers more ๐ฉ๐ซ๐๐๐ข๐ฌ๐ ๐ซ๐จ๐ฎ๐ญ๐๐ซ ๐ ๐ซ๐๐๐ข๐โฆ.
0
60
0
RT @avibose22: ๐จ Code is live! Check out LoRe โ a modular, lightweight codebase for personalized reward modeling from user preferences. ๐ฆ Fโฆ.
github.com
Code to reproduce results of our experiments using LoRe - facebookresearch/LoRe
0
6
0
Using retrieval? --> check out this work by my awesome collaborator on how to increase diversity when retrieving!.
1/8 ๐ How can retrieval augmentation be made both relevant and non-redundant for few-shot adaptation? I'm excited to introduce COBRA. Catch our poster at #CVPR25 (ExHall D, Poster #450) on Sat 14 Jun, 5โ7 p.m. CDT:
0
2
7
RT @fengyao1909: ๐ฅ "Vibe coding" is everywhereโbut is it really care-free?. We introduce ๐๐๐๐, an RL framework that trains LLMs with automaโฆ.
0
41
0
Joint work with awesome collaborators:.@keeghin, @jbilmes, @_siskac, @LukeZettlemoyer,.@hila_gonen, and @csinva. Also huge thanks to @arnaved, @butcher_jasper, @BhattGantavya, @zvez11, @faeze_brh, @MakeshNarsimhan, @soumyesinghal, @_shruti_joshi_ for helpful discussions.
0
0
7
OmniGuard is also the fastest and most accurate at rapid adaptation to new settings using few-shot examples (one of the purported benefits of separate Guard models built using LLMs).
1
0
3
Using the internal representations for safety classification bypasses the need for a separate Guard model while making OmniGuard 120X faster than the fastest baseline Guard model.
1
0
3
How does OmniGuard detect harmful prompts across languagesโeven ciphersโand modalities?. 1๏ธโฃ Identifies universal internal representations from models (LLMs/MLLMs). 2๏ธโฃ Builds powerful classifiers using these shared representations. A single unified guard ! ๐๐
1
0
3
๐จ New Paper! ๐จ.Guard models slow, language-specific, and modality-limited?. Meet OmniGuard that detects harmful prompts across multiple languages & modalities all using one approach with SOTA performance in all 3 modalities!! while being 120X faster ๐.
1
39
82
RT @jinaycodes: Introducing soarXiv โ๏ธ, the most beautiful way to explore human knowledge. Take any paper's URL and replace arxiv with soarโฆ.
0
1K
0
RT @soumyesinghal: โกโก Llama-Nemotron-Ultra-253B just dropped: our most advanced open reasoning model.๐งต๐
0
13
0
RT @soumyesinghal: โก๏ธ Llama-Nemotron-Ultra is fully open โ weights and post-training data. Achieves 76.0% on GPQA via FP8 RL training withโฆ.
0
4
0
RT @soumyesinghal: ๐ Meet Llama-Nemotron-Super-49B, our teamโs new reasoning model released at #GTC25! Proud to have contributed ๐ง . Optimizโฆ.
huggingface.co
0
4
0
RT @snehaark: Come for the ridiculous 30 column spreadsheet created at @sweetgreen, stay for a critical discussion of how you *actually* scโฆ.
0
3
0
RT @_shruti_joshi_: 1\ Hi, can I get an unsupervised sparse autoencoder for steering, please? I only have unlabeled data varying across mulโฆ.
0
10
0