Sign in

Shivani Kumar

@shivanikumar.bsky.social
28 followers 30 following 15 posts

Postdoc @ University of Michigan | PhD from LCS2, IIITDelhi | Working in Computational Social Science #NLProc More info: kumarshivani.com

PostsRepliesMedia
Shivani Kumar @shivanikumar.bsky.social · 01/03/2025
Are models better at psychological vs. real-world dilemmas? 👍 Yes, models perform better on psychological scenarios than Reddit dilemmas. The gap is larger in predicting ethics & decision factors. Why? Structured scenarios align with values, while Reddit dilemmas add noise and ambiguity. (7/n)
120
Shivani Kumar @shivanikumar.bsky.social · 01/03/2025
Do the responder's values improve predictions? 👍 Yes, context matters! Values aid action prediction, but models rely on surface patterns. Surprisingly, a short self-authored persona works as well as values in personalizing predictions. Examples also help in identifying decision factors. (6/n)
110
Shivani Kumar @shivanikumar.bsky.social · 01/03/2025
Can models reason equally well in different languages? 👎 No! Moral reasoning varies. English, Spanish & Russian outperform. Arabic & Hindi show lower confidence due to limited data & complex morphology. ➕ Identifying decision factors lags behind action prediction. (5/n)
110
Shivani Kumar @shivanikumar.bsky.social · 01/03/2025
Can AI reason morally? We tested LLMs with UniMoral to: ⚖️ Make action choices 🏛️ Identify ethical preferences ✅ Recognize influences 🔮 Predict consequences Insights: LLMs excel at action & consequence but lag in ethics & factors. But, how well do they generalize across languages and contexts? (4/n)
120
Shivani Kumar @shivanikumar.bsky.social · 01/03/2025
What’s inside? 💭 Multilingual Hypothetical + Reddit based dilemmas 🌐 Action choices of people across 46 countries! 🔎 Ethical principles preferences 📊 Cultural & moral profiles of annotators 🔁 Consequence modeling Think of it as a "CT scan" of human moral judgment! (3/n)
121
Shivani Kumar @shivanikumar.bsky.social · 01/03/2025
Why care?🤔 AI thrives on decision-making, yet most NLP research in moral reasoning relies on fragmented, western-centric data. What’s missing? A dataset capturing the full cycle: actions ⚖️, ethics 🏛️, consequences 🔄, and cultural nuance 🌏. That’s where UniMoral comes in. (2/n)
110
Shivani Kumar @shivanikumar.bsky.social · 01/03/2025
Are models better at psychological vs. real-world dilemmas? 👍 Yes, models perform better on psychological scenarios than Reddit dilemmas. The gap is larger in predicting ethics & decision factors. Why? Structured scenarios align with values, while Reddit dilemmas add noise and ambiguity. (7/n)
010
Shivani Kumar @shivanikumar.bsky.social · 01/03/2025
Do the responder's values improve predictions? 👍 Yes, context matters! Values aid action prediction, but models rely on surface patterns. Surprisingly, a short self-authored persona works as well as values in personalizing predictions. Examples also help in identifying decision factors. (6/n)
100
Shivani Kumar @shivanikumar.bsky.social · 01/03/2025
Can models reason equally well in different languages? 👎 No! Moral reasoning varies. English, Spanish & Russian outperform. Arabic & Hindi show lower confidence due to limited data & complex morphology. ➕ Identifying decision factors lags behind action prediction. (5/n)
100
Shivani Kumar @shivanikumar.bsky.social · 01/03/2025
Can AI reason morally? We tested LLMs with UniMoral to: ⚖️ Make action choices 🏛️ Identify ethical preferences ✅ Recognize influences 🔮 Predict consequences Insights: LLMs excel at action & consequence but lag in ethics & factors. But, how well do they generalize across languages and contexts? (4/n)
120
Shivani Kumar @shivanikumar.bsky.social · 01/03/2025
What’s inside? 💭 Multilingual Hypothetical + Reddit based dilemmas 🌐 Action choices of people across 46 countries! 🔎 Ethical principles preferences 📊 Cultural & moral profiles of annotators 🔁 Consequence modeling Think of it as a "CT scan" of human moral judgment! (3/n)
100
Shivani Kumar @shivanikumar.bsky.social · 01/03/2025
Why care?🤔 AI thrives on decision-making, yet most NLP research in moral reasoning relies on fragmented, western-centric data. What’s missing? A dataset capturing the full cycle: actions ⚖️, ethics 🏛️, consequences 🔄, and cultural nuance 🌏. That’s where UniMoral comes in. (2/n)
110