Precision over Diversity: High-Precision Reward Generalizes to Robust Instruction Following

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Zeng, Yirong, Liu, Yufei, Ding, Xiao, Hou, Yutai, Wang, Yuxian, Song, Haonan, Ning, Wu, Tu, Dandan, Zhang, Qixun, Cai, Bibo, He, Yuxiang, Liu, Ting
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866917198109343744
author Zeng, Yirong
Liu, Yufei
Ding, Xiao
Hou, Yutai
Wang, Yuxian
Song, Haonan
Ning, Wu
Tu, Dandan
Zhang, Qixun
Cai, Bibo
He, Yuxiang
Liu, Ting
author_facet Zeng, Yirong
Liu, Yufei
Ding, Xiao
Hou, Yutai
Wang, Yuxian
Song, Haonan
Ning, Wu
Tu, Dandan
Zhang, Qixun
Cai, Bibo
He, Yuxiang
Liu, Ting
contents A central belief in scaling reinforcement learning with verifiable rewards for instruction following (IF) tasks is that, a diverse mixture of verifiable hard and unverifiable soft constraints is essential for generalizing to unseen instructions. In this work, we challenge this prevailing consensus through a systematic empirical investigation. Counter-intuitively, we find that models trained on hard-only constraints consistently outperform those trained on mixed datasets. Extensive experiments reveal that reward precision, rather than constraint diversity, is the primary driver of effective alignment. The LLM judge suffers from a low recall rate in detecting false response, which leads to severe reward hacking, thereby undermining the benefits of diversity. Furthermore, analysis of the attention mechanism reveals that high-precision rewards develop a transferable meta-skill for IF. Motivated by these insights, we propose a simple yet effective data-centric refinement strategy that prioritizes reward precision. Evaluated on five benchmarks, our approach outperforms competitive baselines by 13.4\% in performance while achieving a 58\% reduction in training time, maintaining strong generalization beyond instruction following. Our findings advocate for a paradigm shift: moving away from the indiscriminate pursuit of data diversity toward high-precision rewards.
format Preprint
id arxiv_https___arxiv_org_abs_2601_04954
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Precision over Diversity: High-Precision Reward Generalizes to Robust Instruction Following
Zeng, Yirong
Liu, Yufei
Ding, Xiao
Hou, Yutai
Wang, Yuxian
Song, Haonan
Ning, Wu
Tu, Dandan
Zhang, Qixun
Cai, Bibo
He, Yuxiang
Liu, Ting
Machine Learning
Artificial Intelligence
A central belief in scaling reinforcement learning with verifiable rewards for instruction following (IF) tasks is that, a diverse mixture of verifiable hard and unverifiable soft constraints is essential for generalizing to unseen instructions. In this work, we challenge this prevailing consensus through a systematic empirical investigation. Counter-intuitively, we find that models trained on hard-only constraints consistently outperform those trained on mixed datasets. Extensive experiments reveal that reward precision, rather than constraint diversity, is the primary driver of effective alignment. The LLM judge suffers from a low recall rate in detecting false response, which leads to severe reward hacking, thereby undermining the benefits of diversity. Furthermore, analysis of the attention mechanism reveals that high-precision rewards develop a transferable meta-skill for IF. Motivated by these insights, we propose a simple yet effective data-centric refinement strategy that prioritizes reward precision. Evaluated on five benchmarks, our approach outperforms competitive baselines by 13.4\% in performance while achieving a 58\% reduction in training time, maintaining strong generalization beyond instruction following. Our findings advocate for a paradigm shift: moving away from the indiscriminate pursuit of data diversity toward high-precision rewards.
title Precision over Diversity: High-Precision Reward Generalizes to Robust Instruction Following
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2601.04954