Akshita Jha

Research Scientist @apple | PhD @VT_CS | Prev: @GoogleAI @iiit_hyderabad @debian @linuxfoundation

[sanchit kabra @sanchitkabra4]

Can reasoning help mitigate stereotypical bias in language models?
Excited to share our latest work "Reasoning Towards Fairness: Mitigating Bias in Language Models through Reasoning-Guided Fine-Tuning" that explores this direction.
Work w/ @akshitajha and @chandankreddy 1/5

[Akshita Jha @akshitajha]

Excited to share our latest work, "Reasoning Towards Fairness", led by @sanchitkabra4, where we explore how reasoning can help mitigate stereotypical bias in language models.
Paper: https://t.co/WhzQVK5hj9

[Chandan Reddy @chandankreddy]

Can reasoning reduce bias in language models?
Turns out — yes.
We introduce ReGiFT, a simple yet powerful way to transfer fairness by leveraging reasoning traces.
Paper: https://t.co/DtCjDIKtlc
#AI #LLM #NLP #Fairness #BiasMitigation

[Akshita Jha @akshitajha]

If you are curious about how targeting flaws can mitigate stereotypical outputs, check out our paper here: https://t.co/KBtMmGhQkN n/n

[Akshita Jha @akshitajha]

Are generative language models biased or just "flawed"?

Our new paper "Biased or Flawed? Mitigating Stereotypes in Generative Language Models by Addressing Task-Specific Flaws" unpacks this critical distinction.💡

Work w/ @sanchitkabra4 and @chandankreddy 🧵 1/n

[Akshita Jha @akshitajha]

This approach works across multiple dimensions like nationality, age, gender, disability, and appearance. 🌍 5/n

[Simran Khanuja @simi_97k]

Stellar cast and an extremely relevant+timely topic of discussion ✨ looking forward to attending this💕

[Akshita Jha @akshitajha]

@keviv9 @SCAI_ASU @ASU @CityofPhoenixAZ Congratulations!!

[Akshita Jha @akshitajha]

@josephradhik This is a more widespread problem, and we recently documented such "stereotypical pull" in other text-to-image generation models as well.

https://t.co/eKv2mWj3Wa

[Akshita Jha @akshitajha]

@sunipa17 @vinodkpg @cephaloponderer @shachi_dave @qadrida @chandakreddy We also observe a stronger phenomenon, which we formalize and quantify using the term "stereotypical pull", which measures how the 'default' representations of generated images of identity groups have a pull toward stereotypical depictions...... 5/n

[Vinodkumar Prabhakaran @vinodkpg]

And this one too! Another upcoming #acl2024 paper on how various geo-cultural stereotypes are reflected in text-to-image generations, and how persistent they are, for which @akshitajha devised this new eval/analysis called "stereotypical pull". Work led by @akshitajha & @sunipa17

[Akshita Jha @akshitajha]

@rnav_arora Thanks, Arnav!

[Akshita Jha @akshitajha]

New #ACL2024 paper alert📢: Introducing ViSAGe: "Visual Stereotypes Around the Globe" - a dataset enabling evaluation of known nationality-based stereotypes in Text-to-Image models. Work w/ @sunipa17, @vinodkpg, @cephaloponderer, Sarah, @shachi_dave, @qadrida, @chandakreddy 1/n

[Akshita Jha @akshitajha]

Find the full paper here: https://t.co/rDlrNikwkB and the dataset of visual stereotypes at https://t.co/NHTCGZVBEO. n/n

[Akshita Jha @akshitajha]

…..and we find that the stereotypical pull is disproportionately more prominent for identity groups from the Global South even when prompted with ‘neutral’ or ‘non-stereotypical’ prompts. 6/n