Off-Policy Learning to Reason Works Because It Is More Pessimistic Than You Think
Abstract
Large scale reinforcement learning has become a central tool for improving reasoning in large language models. At this scale, generation is often lagged or asynchronous, so updates are performed on data collected by older policies. This makes learning inherently off-policy. Most existing approaches nevertheless remain rooted in PPO-style trust-region objectives, treating training as approximately on-policy and using importance weights to correct distribution mismatch. These corrections can introduce high variance, destabilize optimization, and accelerate entropy collapse. Recent work suggests an alternative: rather than correcting the mismatch, one can embrace off-policy data and remove importance weights, often yielding stronger algorithms. In this paper, we provide an intuitive construction of off-policy objectives that include successful off-policy objectives and show that their effectiveness can be understood through implicit pessimism: they optimize toward target policies that are more conservative than their nominal objectives suggest. This perspective explains why some particular implementation choices improve stability: they implicitly control the effective target distribution. We then propose a principled modification that stabilize this induced distribution and improve off-policy learning.
Related
- Decision-focused predictions via pessimistic bilevel optimization: complexity and algorithms
- A true single-level reformulation for pessimistic bilevel optimization
- Single-Loop Gradient Algorithms for Pessimistic Bilevel Optimization Problems
- Scalable Distributed Stochastic Optimization via Bidirectional Compression: Beyond Pessimistic Limits
- Reinforcement Learning Under Runtime Confounding: Policy Learning and Off-Policy Evaluation
BibTeX
@misc{sakhi2026offpolicy,
title = {{Off-Policy Learning to Reason Works Because It Is More Pessimistic Than You Think}},
author = {Otmane Sakhi and Aleksei Arzhantsev and Imad Aouali and Flavian Vasile},
howpublished = {HAL},
year = {2026},
month = may,
eprint = {2605.28150},
archivePrefix = {arXiv},
url = {https://hal.science/hal-05734689v1},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Off-Policy Learning to Reason Works Because It Is More Pessimistic Than You Think
AU - Otmane Sakhi
AU - Aleksei Arzhantsev
AU - Imad Aouali
AU - Flavian Vasile
PY - 2026
DA - 2026/05/27
PB - HAL
UR - https://hal.science/hal-05734689v1
AB - Large scale reinforcement learning has become a central tool for improving reasoning in large language models. At this scale, generation is often lagged or asynchronous, so updates are performed on data collected by older policies. This makes learning inherently off-policy. Most existing approaches nevertheless remain rooted in PPO-style trust-region objectives, treating training as approximately on-policy and using importance weights to correct distribution mismatch. These corrections can introduce high variance, destabilize optimization, and accelerate entropy collapse. Recent work suggests an alternative: rather than correcting the mismatch, one can embrace off-policy data and remove importance weights, often yielding stronger algorithms. In this paper, we provide an intuitive construction of off-policy objectives that include successful off-policy objectives and show that their effectiveness can be understood through implicit pessimism: they optimize toward target policies that are more conservative than their nominal objectives suggest. This perspective explains why some particular implementation choices improve stability: they implicitly control the effective target distribution. We then propose a principled modification that stabilize this induced distribution and improve off-policy learning.
ER -