Suggestions
Bhavana Dalvi
Lead Research Scientist@AI2 | ex-Googler | CMU Ph.D, Generative AI, LLMs, Explanations, Machine Learning, Reasoning
Bhavana Dalvi is a Lead Research Scientist at the Allen Institute for AI (AI2), based in the Greater Seattle Area.1 She has over 500 connections on LinkedIn, indicating a strong professional network in her field.1
Dalvi's educational background includes studies at Carnegie Mellon University, which likely contributed to her expertise in artificial intelligence and machine learning.1 Her work at AI2 focuses on various aspects of AI research, particularly in the areas of natural language processing and reasoning.
Some of her recent professional activities and interests include:
-
Natural Language Reasoning and Structured Explanations: Dalvi has been involved in workshops and research related to this topic, demonstrating her expertise in advanced NLP techniques.2
-
Skill Set Optimization: She has shared information about reinforcing language models, indicating her work on improving AI capabilities.3
-
Continual Learning in Language Agents: Dalvi has shown interest in research on continually learning language agents for rapid task adaptation and generalization, suggesting her involvement in cutting-edge AI development.4
Before joining AI2, Dalvi worked at Google, which adds to her impressive background in the tech industry.5 Her journey in machine learning has taken her from Bombay to Seattle, showcasing her international experience and dedication to the field.5
As a lead research scientist at AI2, Bhavana Dalvi is likely involved in shaping the direction of AI research and contributing to significant advancements in the field of artificial intelligence, particularly in areas related to natural language processing and machine learning.
Highlights
Can LLMs go through given datasets analyze them to autonomously generate valid scientific hypotheses? We are thrilled to present 🎉 DiscoveryBench 🎉, a benchmark consisting of 264 discovery tasks from 6 scientific domains, from humanities to biology: https://lnkd.in/gCgnGBss
We benchmark state-of-the-art reasoning frameworks with the most advanced LLMs, but the non-oracle best agent’s performance only peaks at 17% (even with Oracle feedback, the best score is 25%) underscoring the challenging nature of the task and the benchmark. This is a collaborative effort with Bodhisattwa Majumder, Harshit Surana, Dhruv Agarwal, Tushar Khot, Ashish Sabharwal, Peter Clark and others.
Excited to present our recent work on "Skill Set Optimization" a continual learning paradigm for in-context policy abstraction and improvement in goal-oriented LLM actors. 🤩
Thanks to my amazing collaborators: Kolby Nottingham, Bodhisattwa Majumder, Sameer Singh, Peter Clark, Roy Fox
Project & Code: https://lnkd.in/g76DKaAQ Paper: https://lnkd.in/gskRaH7X
