Anvesh Rao (@nvshrao) - Twitter Profili | Zamantika Mersobahis Locabet

Sabitlenmiş Tweet

Presenting our latest work at San Diego! Persona-assigned LLM agents show realistic but sometimes unsafe socio-cognitive effects. Happy to chat about personas & personalization. Also, on the 2026 RS/AS job market, reach out! #ACL2026

Anvesh Rao@nvshrao

(1/4) Excited to share our new work w/ @snigdhac25 Do LLM Agents Mirror Socio-Cognitive Effects in Power-Asymmetric Conversations? To be presented at ACL 2026 (main) We study how LLMs behave in roles like manager–employee, doctor–nurse 📄 Paper: arxiv.org/pdf/2605.17694 🧵👇

English

1

7

1.3K

Anvesh Rao@nvshrao·6d

@snigdhac25 (4/4) ⚠️ Implication: LLMs are realistic, but that includes risky social biases. We need models that replicate realistic power dynamics without amplifying biases #AI #LLM #Safety #Bias

English

0

28

Anvesh Rao@nvshrao·6d

@snigdhac25 (3/4) 📊 General findings: • Low-status agent → adapts style (realistic) but more compliant (unsafe) ⚠️ • High-status agent → more persuasive 💬 • Effects are strongest early in conversations ⏱️ • GPT models show the weakest power effects!

English

1

0

31

Anvesh Rao@nvshrao·6d

(1/4) Excited to share our new work w/ @snigdhac25 Do LLM Agents Mirror Socio-Cognitive Effects in Power-Asymmetric Conversations? To be presented at ACL 2026 (main) We study how LLMs behave in roles like manager–employee, doctor–nurse 📄 Paper: arxiv.org/pdf/2605.17694 🧵👇

English

2

1

12

1.3K

Anvesh Rao retweetledi

Snigdha Chaturvedi@snigdhac25·13 May

Congratulations to Dr. Anvesh Rao Vijjini for successfully defending his PhD thesis on realism and safety of personalized LLMs. Check out his work here: nvshrao.github.io PS: Anvesh is on the job market! @nvshrao @unc_ai_group @unccs

English

0

4

28

2.1K

Anvesh Rao@nvshrao·29 Nis

Heading to Albuquerque! ✈️ I’ll be presenting our work on discovering personalization bias in LLMs. Come hear our talk on Wed, Apr 30, 3pm, Ballroom B (Session EBF.1)! #NAACL2025

Anvesh Rao@nvshrao

(1/4) Excited to present our latest work "𝐄𝐱𝐩𝐥𝐨𝐫𝐢𝐧𝐠 𝐒𝐚𝐟𝐞𝐭𝐲-𝐔𝐭𝐢𝐥𝐢𝐭𝐲 𝐓𝐫𝐚𝐝𝐞-𝐎𝐟𝐟𝐬 𝐢𝐧 𝐏𝐞𝐫𝐬𝐨𝐧𝐚𝐥𝐢𝐳𝐞𝐝 𝐋𝐌𝐬" at #NAACL2025! 🎉 We investigate 𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝑖𝑧𝑎𝑡𝑖𝑜𝑛 𝐵𝑖𝑎𝑠 in LLMs. 📄 Paper: arxiv.org/abs/2406.11107 🧵👇

English

0

3

19

840

Anvesh Rao retweetledi

CoNLL 2026@conll_conf·4 Mar

🚨 Guest Speaker Alert! 🚨 We’re thrilled to announce that #CoNLL2025 will feature: 🥁 Raquel Fernández (@raquel_dmg, University of Amsterdam) & Jean-Rémi King (@JeanRemiKing, CNRS / Meta AI)! 🎤✨ Check out their awesome work!👇

English

1

3

17

2.7K

Anvesh Rao@nvshrao·18 Şub

@snigdhac25 @SomnathBrc (4/4) See you in Albuquerque this May! 🎉 💻 Code: github.com/brcsomnath/per… #NAACL #AI #LLM #Bias #Safety cc: @uncnlp @unccs

English

0

2

145

Anvesh Rao@nvshrao·18 Şub

@snigdhac25 @SomnathBrc (3/4)🚨 Key finding: Despite advancements in alignment, Personalization Bias (PB) remains. Preference tuning shows little to no improvement.

English

1

0

1

129

Anvesh Rao@nvshrao·18 Şub

(1/4) Excited to present our latest work "𝐄𝐱𝐩𝐥𝐨𝐫𝐢𝐧𝐠 𝐒𝐚𝐟𝐞𝐭𝐲-𝐔𝐭𝐢𝐥𝐢𝐭𝐲 𝐓𝐫𝐚𝐝𝐞-𝐎𝐟𝐟𝐬 𝐢𝐧 𝐏𝐞𝐫𝐬𝐨𝐧𝐚𝐥𝐢𝐳𝐞𝐝 𝐋𝐌𝐬" at #NAACL2025! 🎉 We investigate 𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝑖𝑧𝑎𝑡𝑖𝑜𝑛 𝐵𝑖𝑎𝑠 in LLMs. 📄 Paper: arxiv.org/abs/2406.11107 🧵👇

English

2

7

34

4K

Anvesh Rao retweetledi

CoNLL 2026@conll_conf·8 Şub

CoNLL 2025 Call for Papers 😀! #CoNLL2025 conll.org 🔴 Co-located w/ ACL 2025 (July 31 - August 1) ⚪️ This year CoNLL will only accept direct submissions (ddl: March 14 2025) ⚫️ CoNLL will accept both non-archival and archival submissions!

English

1

3

13

4.4K

Anvesh Rao@nvshrao·17 Eki

@emnlpmeeting @rrmenon10 @snigdhac25 @shsriva @unccs @uncnlp @UNC (3/3) We also uncover interesting age and gender biases: for the same actions, models tend to judge Female narrators as more innocent ("NTA/Not-the-A**hole") compared to Male narrators. Code and Data: github.com/nvshrao/social…

English

0

1

2

383

Anvesh Rao@nvshrao·17 Eki

@emnlpmeeting @rrmenon10 @snigdhac25 @shsriva @unccs @uncnlp @UNC (2/3) TL;DR: We find that LLMs widely disagree with human consensus on socially acceptable behavior. However, we can improve alignment by making models consider multiple perspectives before delivering a judgment on the narrator's actions. Paper: arxiv.org/abs/2410.08698…

English

1

0

1

157

Anvesh Rao@nvshrao·17 Eki

(1/3) Check out our latest EMNLP 2024 (Findings) work - “SocialGaze: Improving the Integration of Human Social Norms in Large Language Models”; also to be presented at the 6th WNU @emnlpmeeting ! with @rrmenon10 @snigdhac25 @shsriva at @unccs @uncnlp @UNC

English

1

8

22

4.4K

Anvesh Rao@nvshrao·2 May

(3/3) As a virtual volunteer on May 3rd, I will be available on Gather Town if you wish to connect with me🙂! Virtual attendees, please feel free to let me know of any bugs, missing/broken links, or media on underline (underline.io/events/383/ses…) #eacl2023

English

0

1

96

Anvesh Rao@nvshrao·2 May

(2/3) TL;DR: Our research demonstrates that pre-training the classifier on a task that is easier than and similar to the fine-tuning task can benefit segmenting transcripts. Our proposed pretraining, "Next Conversation Prediction" achieves this goal.

English

1

0

1

137

Anvesh Rao@nvshrao·2 May

(1/3) Check out our paper "Curricular Next Conversation Prediction Pretraining for Transcript Segmentation" at EACL 2023 Findings @eaclmeeting (aclanthology.org/2023.findings-…) with my mentors Hanieh Deilamsalehy, @FranckDernoncou at @AdobeResearch, and @snigdhac25 at @uncnlp @UNC

English

1

2

13

3.3K

Anvesh Rao retweetledi

Snigdha Chaturvedi@snigdhac25·7 Ara