Shaswati Saha

Graduate Research Assistant, UMBC.

[Shaswati Saha @_Shaswati]

🎉 Excited to share that our paper, PARSE, has been accepted to #ECCV2026!

Can a text-to-image model forget one unwanted concept without forgetting everything around it?

PARSE is our training-free approach to doing exactly that. More below 🧵

[Shaswati Saha @_Shaswati]

Our biggest takeaway: good concept erasure isn’t about deleting as much as possible. It’s about drawing the boundary carefully.

Huge thanks to Rajasekhar Anguluri and @manasgaur90 for all the discussions and guidance. More details and code soon! 🙌

[Shaswati Saha @_Shaswati]

🎉 Excited to share that our paper, PARSE, has been accepted to #ECCV2026!

Can a text-to-image model forget one unwanted concept without forgetting everything around it?

PARSE is our training-free approach to doing exactly that. More below 🧵

[Shaswati Saha @_Shaswati]

We compared PARSE against 14 recent concept-erasure approaches.

Overall, PARSE achieved a strong balance: robustly erasing unwanted concepts without unnecessarily affecting the rest of the model.

[Shaswati Saha @_Shaswati]

@Eleni30fillou Hi, my research focuses on localizing and editing knowledge to ensure safe generation. My EMNLP25 paper reveals vulnerability of the edited models. [https://t.co/uL8ER81YLo]

I’m at NeurIPS and would love to chat more!

[Shaswati Saha @_Shaswati]

I’ll be at NeurIPS Dec 2–7. Excited to meet folks working on diffusion models, model editing, and safety! I'm interested in localizing and editing knowledge in diffusion models (concept erasure, robustness, and side effects).

Come see our poster on Dec 6 (Upper Level Room 23ABC)

[Shaswati Saha @_Shaswati]

@sourajitCS @manasgaur90 @trgokhale In short, erasing is easy to say, hard to do safely. Concept Erasure Techniques introduce side effects—raising important questions about the reliability of current unlearning methods in diffusion models. (7/n)

[Shaswati Saha @_Shaswati]

🚨 Our paper “Side Effects of Erasing Concepts from Diffusion Models” has been accepted to EMNLP 2025 (Findings)! #EMNLP2025
We investigate the vulnerabilities of Concept Erasure Techniques (CETs)
Big shoutout to my amazing collaborators @sourajitCS @manasgaur90 @trgokhale
1/n

[Shaswati Saha @_Shaswati]

@sourajitCS @manasgaur90 @trgokhale (2) CETs can be easily fooled with hierarchical or compositional prompts. (3) CETs suffer from post-erasure attribute leakage and counterintuitive attention shifts. (6/n)

[Shaswati Saha @_Shaswati]

@jbhuang0604 Get well soon.

[Shaswati Saha @_Shaswati]

I am happy to share that our paper "Towards Robust Evaluation of Unlearning in LLMs via Data Transformations" got accepted to @emnlpmeeting 2024. It was an exciting collaboration between the KAI2 lab (@manasgaur90) at @umbccsee, @IITKanpur, @MSFTResearch

[Shaswati Saha @_Shaswati]

@m2saxon @sharakelyan Haha, still no luck with the director tho!

[Shaswati Saha @_Shaswati]

@m2saxon @sharakelyan The closer I could get till now is this, lol

[Shaswati Saha @_Shaswati]

@m2saxon @sharakelyan I got this