Title: FSA-GRPO: Enabling Auditory LLMs to Leverage Few-Shot Demonstrations

Abstract:

Few-shot prompting serves as a potent strategy for adapting auditory large language models to low-resource scenarios, such as children’s speech recognition. Nevertheless, because most auditory LLMs are not explicitly trained to execute inference within this demonstration-conditioned framework, their capacity to capitalize on few-shot prompting remains constrained. To overcome this bottleneck, we propose Few-Shot Aware GRPO (FSA-GRPO), a post-training methodology grounded in reinforcement learning. This approach employs a novel reward structure designed to motivate the model to effectively utilize few-shot demonstrations, thereby enhancing its few-shot adaptation capabilities.

Remarkably, the training process relies exclusively on high-resource adult Automatic Speech Recognition (ASR) data, yet it significantly boosts the model’s general few-shot adaptation proficiency. This improvement translates into performance gains across multiple domains, including children’s speech recognition, speech translation, and audio understanding. Additionally, we investigate the impact of data selection and auxiliary reward weighting to pinpoint the most effective training configuration. Our experimental results demonstrate that when in-domain data are either inaccessible or unsuitable for training, FSA-GRPO outperforms direct fine-tuning on related out-of-domain datasets.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC