Afra Amini

@afra_amini

Research Scientist @GoogleDeepMind | prev: Ph.D. @ETH_AI_Center, Intern @allen_ai, @GoogleDeepMind

afraamini.github.io

Joined September 2021

307 Following

550 Followers

Replies Reposts Media

Pinned
Afra Amini
@afra_amini
Jul 12, 2024
Best-of-N is a straightforward LLM alignment algorithm: return the highest reward sample of N attempts 👍simple & effective 👎generation throughput decreases by a factor of N. 🤔Can we keep the 👍 while eliminating the 👎? 🧵 w/@ryandcotterell @xtimv arxiv.org/pdf/2407.06057…
4
Afra Amini
@afra_amini
Sep 20, 2025
Excited to share that this paper has been accepted to #NeurIPS2025 main track 🎉!
Afra Amini
@afra_amini
May 6, 2025
Current KL estimation practices in RLHF can generate high variance and even negative values! We propose a provably better estimator that only takes a few lines of code to implement.🧵👇 w/ @xtimv and Ryan Cotterell code: arxiv.org/pdf/2504.10637 paper: github.com/rycolab/kl-rb
Afra Amini
@afra_amini
May 6, 2025
Current KL estimation practices in RLHF can generate high variance and even negative values! We propose a provably better estimator that only takes a few lines of code to implement.🧵👇 w/ @xtimv and Ryan Cotterell code: arxiv.org/pdf/2504.10637 paper: github.com/rycolab/kl-rb
4
Afra Amini
@afra_amini
Mar 4, 2025
Excited to share that this paper has been accepted to ICLR 2025 🎉 We've added more experiments in the camera-ready version: arxiv.org/pdf/2407.06057 Code is available here: github.com/rycolab/vbon
Afra Amini
@afra_amini
Jul 12, 2024
Best-of-N is a straightforward LLM alignment algorithm: return the highest reward sample of N attempts 👍simple & effective 👎generation throughput decreases by a factor of N. 🤔Can we keep the 👍 while eliminating the 👎? 🧵 w/@ryandcotterell @xtimv arxiv.org/pdf/2407.06057…
Afra Amini
@afra_amini
Dec 14, 2024
@xtimv and I will be presenting this work at NeurIPS today! Join us from 16:40h at East Exhibition Hall A!
Afra Amini
@afra_amini
Jul 12, 2024
Best-of-N is a straightforward LLM alignment algorithm: return the highest reward sample of N attempts 👍simple & effective 👎generation throughput decreases by a factor of N. 🤔Can we keep the 👍 while eliminating the 👎? 🧵 w/@ryandcotterell @xtimv arxiv.org/pdf/2407.06057…

Afra Amini

Afra Amini’s posts