推动自然语言推理的边界
摘要
自然语言推理（NLI）是自然语言理解的核心任务，具有 fact-checking、question answering 和 information retrieval 等应用。尽管其重要性，current NLI systems 仍 heavy 依赖 supervised learning，datasets often contain annotation artifacts and biases，limiting generalization and real-world applicability。In this work, we apply a reinforcement learning-based approach using Group Relative Policy Optimization（GRPO） for Chain-of-Thought（CoT） learning in NLI，eliminating the need for labeled rationales and enabling this type of training on more challenging datasets such as ANLI。我们使用 parameter-efficient techniques（LoRA 和 QLoRA） fine-tune 7B、14B 和 32B language models，demonstrating strong performance across standard and adversarial NLI benchmarks。我们的 32B AWQ-quantized model 在 7 out of 11 adversarial sets——或在 considering our replication——within a 22GB memory footprint 中超越了 state-of-the-art results，showing robust reasoning can be retained under aggressive quantization。This work provides a scalable and practical framework for building robust NLI systems without sacrificing inference quality。
引用
@article{arxiv.2504.18376,
title = {Pushing the boundary on Natural Language Inference},
author = {Pablo Miralles-González and Javier Huertas-Tato and Alejandro Martín and David Camacho},
journal= {arXiv preprint arXiv:2504.18376},
year = {2025}
}