Apriel-1.5-OpenReasoner: RL Post-Training for General-Purpose and Efficient Reasoning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Pardinas, Rafael, Kamalloo, Ehsan, Vazquez, David, Drouin, Alexandre
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914444653625344
author Pardinas, Rafael
Kamalloo, Ehsan
Vazquez, David
Drouin, Alexandre
author_facet Pardinas, Rafael
Kamalloo, Ehsan
Vazquez, David
Drouin, Alexandre
contents Building general-purpose reasoning models using reinforcement learning with verifiable rewards (RLVR) across diverse domains has been widely adopted by frontier open-weight models. However, their training recipes and domain mixtures are often not disclosed. Joint optimization across domains poses significant challenges: domains vary widely in rollout length, problem difficulty and sample efficiency. Further, models with long chain-of-thought traces increase inference cost and latency, making efficiency critical for practical deployment. We present Apriel-1.5-OpenReasoner, trained with a fully reproducible multi-domain RL post-training recipe on Apriel-Base, a 15B-parameter open-weight LLM, across five domains using public datasets: mathematics, code generation, instruction following, logical puzzles and function calling. We introduce an adaptive domain sampling mechanism that preserves target domain ratios despite heterogeneous rollout dynamics, and a difficulty-aware extension of the standard length penalty that, with no additional training overhead, encourages longer reasoning for difficult problems and shorter traces for easy ones. Trained with a strict 16K-token output budget, Apriel-1.5-OpenReasoner generalizes to 32K tokens at inference and improves over Apriel-Base on AIME 2025, GPQA, MMLU-Pro, and LiveCodeBench while producing 30-50% shorter reasoning traces. It matches strong open-weight models of similar size at lower token cost, thereby pushing the Pareto frontier of accuracy versus token budget.
format Preprint
id arxiv_https___arxiv_org_abs_2604_02007
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Apriel-1.5-OpenReasoner: RL Post-Training for General-Purpose and Efficient Reasoning
Pardinas, Rafael
Kamalloo, Ehsan
Vazquez, David
Drouin, Alexandre
Machine Learning
Building general-purpose reasoning models using reinforcement learning with verifiable rewards (RLVR) across diverse domains has been widely adopted by frontier open-weight models. However, their training recipes and domain mixtures are often not disclosed. Joint optimization across domains poses significant challenges: domains vary widely in rollout length, problem difficulty and sample efficiency. Further, models with long chain-of-thought traces increase inference cost and latency, making efficiency critical for practical deployment. We present Apriel-1.5-OpenReasoner, trained with a fully reproducible multi-domain RL post-training recipe on Apriel-Base, a 15B-parameter open-weight LLM, across five domains using public datasets: mathematics, code generation, instruction following, logical puzzles and function calling. We introduce an adaptive domain sampling mechanism that preserves target domain ratios despite heterogeneous rollout dynamics, and a difficulty-aware extension of the standard length penalty that, with no additional training overhead, encourages longer reasoning for difficult problems and shorter traces for easy ones. Trained with a strict 16K-token output budget, Apriel-1.5-OpenReasoner generalizes to 32K tokens at inference and improves over Apriel-Base on AIME 2025, GPQA, MMLU-Pro, and LiveCodeBench while producing 30-50% shorter reasoning traces. It matches strong open-weight models of similar size at lower token cost, thereby pushing the Pareto frontier of accuracy versus token budget.
title Apriel-1.5-OpenReasoner: RL Post-Training for General-Purpose and Efficient Reasoning
topic Machine Learning
url https://arxiv.org/abs/2604.02007