Saved in:
Bibliographic Details
Main Authors: Bagirov, Farid, Arkhipov, Mikhail, Sycheva, Ksenia, Glukhov, Evgeniy, Bogomolov, Egor
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2510.23393
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912672582205440
author Bagirov, Farid
Arkhipov, Mikhail
Sycheva, Ksenia
Glukhov, Evgeniy
Bogomolov, Egor
author_facet Bagirov, Farid
Arkhipov, Mikhail
Sycheva, Ksenia
Glukhov, Evgeniy
Bogomolov, Egor
contents The application of Reinforcement Learning with Verifiable Rewards (RLVR) to mathematical and coding domains has demonstrated significant improvements in the reasoning and problem-solving abilities of Large Language Models. Despite its success in single generation problem solving, the reinforcement learning fine-tuning process may harm the model's exploration ability, as reflected in decreased diversity of generations and a resulting degradation of performance during Best-of-N sampling for large N values. In this work, we focus on optimizing the max@k metric, a continuous generalization of pass@k. We derive an unbiased on-policy gradient estimate for direct optimization of this metric. Furthermore, we extend our derivations to the off-policy updates, a common element in modern RLVR algorithms, that allows better sample efficiency. Empirically, we show that our objective effectively optimizes max@k metric in off-policy scenarios, aligning the model with the Best-of-N inference strategy.
format Preprint
id arxiv_https___arxiv_org_abs_2510_23393
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle The Best of N Worlds: Aligning Reinforcement Learning with Best-of-N Sampling via max@k Optimisation
Bagirov, Farid
Arkhipov, Mikhail
Sycheva, Ksenia
Glukhov, Evgeniy
Bogomolov, Egor
Machine Learning
The application of Reinforcement Learning with Verifiable Rewards (RLVR) to mathematical and coding domains has demonstrated significant improvements in the reasoning and problem-solving abilities of Large Language Models. Despite its success in single generation problem solving, the reinforcement learning fine-tuning process may harm the model's exploration ability, as reflected in decreased diversity of generations and a resulting degradation of performance during Best-of-N sampling for large N values. In this work, we focus on optimizing the max@k metric, a continuous generalization of pass@k. We derive an unbiased on-policy gradient estimate for direct optimization of this metric. Furthermore, we extend our derivations to the off-policy updates, a common element in modern RLVR algorithms, that allows better sample efficiency. Empirically, we show that our objective effectively optimizes max@k metric in off-policy scenarios, aligning the model with the Best-of-N inference strategy.
title The Best of N Worlds: Aligning Reinforcement Learning with Best-of-N Sampling via max@k Optimisation
topic Machine Learning
url https://arxiv.org/abs/2510.23393