Twitxy Saved

Pratinav Seth ( ✈️ EMNLP 2026)

Verified account This account is verified. Learn more

@pratinavseth

AI explorer | (Just my thoughts)
Kolkata, India linktr.ee/pratinavseth Joined July 2020
6.7K Following 1.1K Followers 328 Posts

Latest posts 5

Pratinav Seth ( ✈️ EMNLP 2026) Verified account @pratinavseth
Find me switching across Demo at D8 and Poster 67 today 🙌.
#EMNLP2026
Pratinav Seth ( ✈️ EMNLP 2026) Verified account @pratinavseth Oct 25
At #EMNLP2026 in Budapest! 🇭🇺

We're presenting 5 works from Lexsi_labs across the main conference, system demonstrations, and workshops — covering post-training data curation, inference-time alignment, LLM safety, and legal AI reliability.

2 main conference papers, 1 system demo, and 2 workshop papers.

Here in Budapest with Saisab Sadhu! If you're attending EMNLP, we'd love to connect, discuss our research, and exchange ideas.

Presentation details below 👇
View quoted post on X
0 replies 0 reposts 24 likes 674 views View on X
Pratinav Seth ( ✈️ EMNLP 2026) Verified account @pratinavseth
At #EMNLP2026 in Budapest! 🇭🇺

We're presenting 5 works from Lexsi_labs across the main conference, system demonstrations, and workshops — covering post-training data curation, inference-time alignment, LLM safety, and legal AI reliability.

2 main conference papers, 1 system demo, and 2 workshop papers.

Here in Budapest with Saisab Sadhu! If you're attending EMNLP, we'd love to connect, discuss our research, and exchange ideas.

Presentation details below 👇
5 replies 2 reposts 29 likes 1.6K views View on X
Pratinav Seth ( ✈️ EMNLP 2026) Verified account @pratinavseth
Can we correct unsafe model behaviour without compromising utility?

Over the past few months, I’ve been working with our team at @Lexsi_labs on mechanistic interpretability, safety evaluation, and targeted model repair.

Our latest work explores multilingual and cybersecurity safety through targeted weight-level interventions, improving safety and multi-turn jailbreak resistance while largely preserving model capabilities.

Some improvements generalized beyond the repair settings.

#AISafety #MechanisticInterpretability

Papers, models & code below ↓