BLEUBERI: BLEU is a surprisingly effective reward for instruction following

TL;DR

This paper demonstrates that BLEU, a simple string-matching metric, can effectively replace complex reward models for aligning large language models with human preferences, reducing training costs while maintaining performance.

Contribution

The authors introduce BLEUBERI, a novel method that uses BLEU as a reward in RL-based alignment, showing competitive results with reward model-guided approaches across multiple benchmarks.

Findings

BLEU matches reward models in agreement with human preferences.

BLEUBERI achieves comparable performance to reward model-based methods.

BLEUBERI outputs are more factually grounded than competing methods.

Abstract

Reward models are central to aligning LLMs with human preferences, but they are costly to train, requiring large-scale human-labeled preference data and powerful pretrained LLM backbones. Meanwhile, the increasing availability of high-quality synthetic instruction-following datasets raises the question: can simpler, reference-based metrics serve as viable alternatives to reward models during RL-based alignment? In this paper, we show first that BLEU, a basic string-matching metric, surprisingly matches strong reward models in agreement with human preferences on general instruction-following datasets. Based on this insight, we develop BLEUBERI, a method that first identifies challenging instructions and then applies Group Relative Policy Optimization (GRPO) using BLEU directly as the reward function. We demonstrate that BLEUBERI-trained models are competitive with models trained via reward model-guided RL across four challenging instruction-following benchmarks and three different base language models. A human evaluation further supports that the quality of BLEUBERI model outputs is on par with those from reward model-aligned models. Moreover, BLEUBERI models generate outputs that are more factually grounded than competing methods. Overall, we show that given access to high-quality reference outputs (easily obtained via existing instruction-following datasets or synthetic data generation), string matching-based metrics are cheap yet effective proxies for reward models during alignment. We release our code and data at https://github.com/lilakk/BLEUBERI.