Conformal prediction sets are a useful way to capture uncertainty for LLMs & deep learning models. But they're data-hungry! We propose a semi-bandit algo to learn these sets online. Check out our @icmlconf paper: https://t.co/EQgTl3Xubc
Work led by @HaosenGe & with @obastani
Excited to see our paper "Generative AI Can Harm Learning" cited in Ch 7 of the 2025 Economic Report of the President: https://t.co/vZmX4Cy7CD
Paper: https://t.co/toKa5BVDLQ, co-authored with @obastani@Alpsungu@HaosenGe Ozge & Rei
Excited to see Philip K.H. Wong Centre @HKUniversity for promoting my working paper! I examined how Open Government Information mobilizes legal resistance in China but suffers a potenial backlash from local governments. Welcome any comments!
3/ To address this issue, we propose a class of new algorithms: compliance robust algorithms. We demonstrate that it is possible to convert any machine learning model into compliance robust models that guarantee improvement in outcome accuracy and fairness.
It's fantastic to see the increased interest in our research! I firmly believe it addresses a crucial question that any organization looking to implement AI to support human decision-making should consider.
.@Wharton's@hamsabastani, Haosen Ge from @WhartonAnalytics, & @obastani explore AI-human collaboration challenges in their latest research featured in @Medium's "How to Improve AI Fairness in an Unfair World."
Check it out here: https://t.co/21wZSgfU63
2/In a nutshell, our research highlights a key issue with current AI decision support systems: the potential for AI-assisted humans to perform worse than either AI operating alone or humans without AI. This contradicts the very purpose of integrating AI systems.
@sid_devic@SimonsInstitute@hamsabastani@obastani Thank you! We are hoping to add that as an extension. In general, we can make an ฯต-compliance-robust guarantee, and the guarantee becomes stronger as we estimate the human decision function more accurately.
Excited to present my work with @hamsabastani and @obastani at the Simons Institute for the Theory of Computing (@SimonsInstitute) !
We delve into the challenges with algorithmic fairness when dealing with human decision makers who can selectively comply.
Nothing to see here.
Just an AI system learning to strategically plan world conquest, negotiate with humans in English and then betray them via the game of Diplomacy.
https://t.co/zkN1EuJnAe
Given a trained #ML model, can we estimate how well it works on new data w/o seeing any labels?
Can we still do it even if both covariates and labels shift in data?
Our #NeurIPS2022 paper shows how to est model perf on new data in this hard case! https://t.co/VCmlNvWLjc 1/2
In our new work - Algorithm Distillation - we show that transformers can improve themselves autonomously through trial and error without ever updating their weights.
No prompting, no finetuning. A single transformer collects its own data and maximizes rewards on new tasks.
1/N