Anjishnu

PhD student @GeorgeMasonU, 3x intern @Amazon, multimodal LLMs, understanding culture and fairness for NLP. Kpop, video games, dogs, cars. He/him

[Anjishnu @iamshnoo]

🎉 Our paper "Global Voices, Local Biases: Socio-Cultural Prejudices across Languages" has been accepted at #emnlp2023 ! Thread 🧵⬇️

📜 arXiv : https://t.co/1hK8sL1rLW
🤗 dataset : https://t.co/EbX4jBdGw5 (@huggingface)
🕸️ website : https://t.co/aKGw7VqUfF

[Anjishnu @iamshnoo]

@chien_eli Lol, we also have a paper with the name MAPLE which uses metadata augmented samples for pre-training (different task, but we also find that it makes the pre-training more efficient). Here is a link to an older version of the work that’s up on arxiv: https://t.co/3yD3A4apDZ

[Chahat Raj @chahatsaidit]

#ACL2026 has been eventful ✨

Presented Vignette as an Oral and had amazing conversations over Talent or Luck in the poster session.

Also received an SAC Highlight award for Vignette 🫶🏼

@ZzwWilliam @anas_ant @aylin_cim @iamshnoo
#NLProc @aclmeeting #ACL2026NLP

[Anjishnu @iamshnoo]

I am in San Diego till Tuesday, attending @aclmeeting . Let me know if you are around and would like to chat!

[Anjishnu @iamshnoo]

@wonderingishika @nbbozdag This is an interesting set of findings! We looked at a similar question some time back in this paper https://t.co/hGY9HBbsUa

[Anjishnu @iamshnoo]

@PranavVenkit Congrats Pranav!

[Anjishnu @iamshnoo]

Presenting this at #WACV2025 today + oral session tomorrow. Come say hi if you are around!

[Anjishnu @iamshnoo]

🧵1/ New Paper! 🚀 Our latest work with @anas_ant and @ZzwWilliam is now on arXiv. https://t.co/eaJH1FFrJq! We look at cultural understanding in LLMs and develop an approach for automatically updating the culture represented in images. Thread 🧵⬇️

[Antonis Anastasopoulos @anas_ant]

Stop by Jasmine #20 right now to see how @iamshnoo and @chahatsaidit explore multimodal LLMs to uncover the implicit associations they make in their BiasDora work

[Chahat Raj @chahatsaidit]

I'll be at #EMNLP2024 with @iamshnoo presenting BiasDora. Catch us on Thursday, Nov 14, Session 11. @aylin_cim @anas_ant @ZzwWilliam @emnlpmeeting

[Chahat Raj @chahatsaidit]

I'll be at @AIESConf with @aylin_cim and @iamshnoo to show how LLM responses imitate human psychology and how we use it to mitigate social bias using the Contact Hypothesis.
Listen to our talk on Wednesday, Oct 23rd, when we present https://t.co/SlRkIkTvGq
@anas_ant @ZzwWilliam

[Alex Choi @alex_choi_cs]

🧑🤖The LLM Effect: Are Humans Truly Using LLMs, or Are They Being Influenced By Them Instead❓

This is obviously a massive question that is both important and timely.

Accepted to #EMNLP2024 Main, our paper examines the integration of LLMs into specialized expert workflows, and explores both the potential benefits and worrisome pitfalls of this emerging partnership.

📜Preprint available: https://t.co/bq7qYCIUZ6

w/ @SyedaSabrina11, @Prof_JPSingh, and @anas_ant

1/6

[Chahat Raj @chahatsaidit]

We have seen the usual man:doctor🧑‍⚕️ and woman:nurse👩‍⚕️ stereotypes.
But guess what? VLMs are also throwing out wild associations like blonde:dumb🤦‍♀️, old person:dinosaur🦕, or college student:broke. 🧑‍🎓💸

BiasDora, our #EMNLP2024 Findings Paper: https://t.co/3qYn6YWNHP, explores them all! 🧵

[Anjishnu @iamshnoo]

@aylin_cim Congrats Aylin!

[Anjishnu @iamshnoo]

@simi_97k Congrats Simran! See you in Miami :)

[Anjishnu @iamshnoo]

@faridlazuarda 😂

[Anjishnu @iamshnoo]

🧵6/ What next: Given the feasibility of this cultural adaptation pipeline, we can improve its individual components and then use it to augment training data for more culturally balanced models. We could also use it to make near real-time edits to the outputs from generative systems to make them more relevant to a target user! 🔥🌏🤖

[Anjishnu @iamshnoo]

🧵1/ New Paper! 🚀 Our latest work with @anas_ant and @ZzwWilliam is now on arXiv. https://t.co/eaJH1FFrJq! We look at cultural understanding in LLMs and develop an approach for automatically updating the culture represented in images. Thread 🧵⬇️

[Anjishnu @iamshnoo]

🧵5/CultureAdapt: We use the identified cultural artifacts from the previous step as inputs to an object detector, get bounding boxes, and then replace them with a randomly selected object from the same concept class (eg. front door, wardrobe) but from a different country, using stable diffusion inpainting. We measure changes in CLIPScore from source and target country to measure the effectiveness of this adaptation process!