Vishvak Murahari

[1/4] Vibe Analyst: the future of data science. No dashboards. No SQL. Just vibe with your data. Powered by Agent Context Protocols (ACPs) — the first protocol for building multi-agent reasoning systems. vibe-analyst.com Paper: arxiv.org/abs/2505.14569

PersonaGym – the first platform to evaluate persona agents. Persona agents are critical to magical AI experiences, but how do we evaluate them? PersonaGym uses dynamic LLM-powered environments for faithful and targeted agent eval. personagym.com

Our EACL workshop on personalization of Generative AI is accepting ARR submissions. Deadline: Jan 27th, 2024 Submission form: forms.gle/pQSbVNEnZ9R9vy… The workshop will be co-located with EACL in Malta. Workshop date: March 21st or March 22nd Website: genpersonalization.github.io

With LLM based Search Engines like Bard & Perplexity rising, how can creators like NYT maintain ownership & visibility? 🤔 We introduce Generative Engine Optimization (GEO)🔍⚙️- a black-box optimization framework for creators in new-era search! 🔗: generative-engines.com 🧵

Given popular demand, we have decided to extend the deadline to January 10, 2024!

Navigating the dynamic landscape of algorithmic, efficiency, and safety challenges in Gen-AI personalization is key. Join us as our workshop sparks principled discussions and fosters collaboration among industry leaders, academics, startups, and policymakers. Save the date! 🗓️🚀

Thanks to my wonderful collaborators: @AmeetDeshpande_, Peter Clark, Tanmay Rajpurohit, Ashish Sabharwal, @karthik_r_n, @ashwinkalyan7 Website: vishvakmurahari.com/QualEval/ Code: github.com/vmurahari3/Qua… Demo: huggingface.co/spaces/vishvak…

QualEval’s insights can be leveraged toward model improvement, enabling unprecedented agility in model iteration. QualEval aids practitioners to accurately discover and precisely improve underperforming sub-domains and sub-tasks, increasing overall performance significantly.📷

QualEval – your LLM data scientist🥼 Quantitative metrics don’t aid in model improvement Our philosophy – Quality over quantity✅ QualEval qualitatively analyzes tons of data🔍, reasons over mistakes⚠️, and improves your model Your model dev copilot✈️ Demo:vishvakmurahari.com/QualEval/

I am excited to share that #PrincetonInnovation has featured our innovations around Data Multiplexing, a novel neural network efficiency paradigm, alongside other cutting-edge research from @Princeton this year.

Learning Transformer Programs We designed a modified Transformer that can be trained to solve a task and then automatically converted into a discrete, human-readable program. With @_awettig and @danqi_chen. Paper: arxiv.org/abs/2306.01128 Code: github.com/princeton-nlp/… [1/12]

Glad to see the direct impact of our research on OpenAI. Th recent call for democratic inputs recognizes studying the effects of personalization on model behavior as a key direction. Link to our work which showed adverse effects of assigning a persona: tinyurl.com/545pakrk

Semantic textual similarity is an important task in NLP. But sentence similarity can be ambiguous and depends on the aspect of interest. We propose Conditional-STS, which resolves this ambiguity through conditions. Even SOTA models find it challenging! 📜tinyurl.com/4tvxdkj9

Congrats to #ICLR2023 Best Paper Awardees @arimorcos, @DhruvBatraDB , and @erikwijmans, pictured on May 1 in Kilagi. Wijmans presented the work at a special award ceremony. @iclr_conf awarded only 4 best papers. (Coauthors not pictured: Manolis Savva, @irrfaan, and @stefmlee)

"Extremely" happy to share that our work which pushes the boundaries of extreme classification has been accepted at ICML 2023. We show that semantic information can be smartly leveraged to improve classification over millions of labels, a pertinent real-world scenario.

🔥 🔥 Our alternative to CLIP that captures visual-semantic hierarchies 🥳, is more efficient for smaller embeddings ✅, has geometric insights for multimodal learning 🤝🔥🔥 .. led by @kdexd (ex-intern, FAIR)

"Depending on the persona assigned to ChatGPT, its toxicity can increase up to [6 times]. This may be ... harmful to an unsuspecting user." @thomasgermain reports on new work from AI2 student researcher @AmeetDeshpande_ & team: gizmodo.com/chatgpt-ai-ope…

New research from @AmeetDeshpande_, @VishvakM and @karthik_r_n featured in @TechCrunch. Giving ChatGPT a certain "persona" can make it speak like a "bad person," or even Mao Zedong. The responses, researchers found, can become consistently toxic. bit.ly/3GP9IRg

Lowkey goated when #NaturalLanguageProcessing is the vibe! 🤩 Check out this dope paper by @VishvakM and co. to learn more: deepai.org/publication/to…