Task-Level Natural Language Priors as Learning Signals for Low-Resource LLM Training

2026-09-02

AI Summary

为解决低资源条件下训练大型语言模型时数据模糊或不完整的问题，提出Prior-Guided Tuning方法，通过引入自然语言先验作为辅助学习信号来改善模型性能。

Abstract

Large language models (LLMs) often struggle when low-resource training data are ambiguous or incomplete. Task-level natural-language priors can provide useful guidance in such settings, but existing approaches usually treat these priors as input context rather than as learning signals during training. We propose Prior-Guided Tuning (PGT), a training perspective that incorporates natural-language priors as auxiliary learning signals for low-resource LLM training. Under this perspective, we introduce Contrastive Prior Steering (CPS), which keeps the original supervised objective intact while adding positive and negative prior-conditioned auxiliary losses to encourage task-consistent learning and discourage plausible but misleading alternatives. Experiments on AmbiMath, Jigsaw, and MNLI/HANS show that CPS consistently improves over plain and prompt fine-tuning. On AmbiMath, CPS achieves 97.6% average exact-match accuracy. On Jigsaw, CPS improves average Macro F1 by 9.5 percentage points over standard fine-tuning, and with 1/10 of the experimental training data slightly exceeds full-data plain fine-tuning. On HANS, CPS improves non-entailment accuracy by 8.3 and 5.2 percentage points for LLaMA 3.1 8B and Qwen 2.5 7B, respectively, while maintaining comparable in-domain MNLI accuracy. These results support our central claim: task-level natural-language priors can provide useful guidance as auxiliary learning signals for low-resource LLM training. Our code and data will be publicly available.

Problem

• large language models
• low-resource training data
• natural-language priors
• auxiliary learning signals

Innovation

• Prior-Guided Tuning
• Contrastive Prior Steering
• auxiliary learning signals
• low-resource LLM training