about
Vishal Pallagani
I'm a researcher at Patronus AI. I like questions where a careful measurement changes what people believe.
Does a benchmark measure what it claims? What does a headline number actually depend on? Which failures belong to the model, and which to the test that graded it? Most of my work comes back to questions like these. At the moment that means building environments for training and evaluating AI agents, and finding out why models fail in them.
This site is where I write things up: my own experiments, results I've checked, and opinions on research that catches my attention, wherever it comes from. Each entry tries to say plainly what it found and how sure it is.
path
- nowPatronus AIEnvironments for training and evaluating AI agents, and why models fail in them.
- beforeAIISC, University of South CarolinaPhD research on language models for automated planning, with Biplav Srivastava and Amit Sheth.
- 2022IBM ResearchResearch internship; built Plansformer, a language model that generates valid symbolic plans.
selected papers
- 2022Plansformer: Generating Symbolic Plans using TransformersarXiv
- 2022Epistemic Planning in a Fast and Slow SettingCEUR-WS
- 2022A Rich Recipe Representation as Plan to Support Expressive Multi Modal Queries on Recipe Content and Preparation ProcessarXiv
- 2021A generic dialog agent for information retrieval based on automated planning within a reinforcement learning platformPRL @ ICAPS
- 2021PRUDENT: a dialog agent that mixes automated planning and reinforcement learningICAPS demo
elsewhere
Set in Commit Mono and Newsreader. Built with Astro; figures are hand-written SVG and CSS.