ALERT: Adapting Language Models to Reasoning Tasks
Ping Yu♠♠{}^{\spadesuit}start_FLOATSUPERSCRIPT ♠ end_FLOATSUPERSCRIPT Tianlu Wang♠♠{}^{\spadesuit}start_FLOATSUPERSCRIPT ♠ end_FLOATSUPERSCRIPT Olga Golovneva♠♠{}^{\spadesuit}start_FLOATSUPERSCRIPT ♠ end_FLOATSUPERSCRIPT Badr AlKhamissi△△{}^{\triangle}start_FLOATSUPERSCRIPT △ end_FLOATSUPERSCRIPT Siddharth Verma△normal-△{}^{\triangle}start_FLOATSUPERSCRIPT △ end_FLOATSUPERSCRIPT Zhijing Jin‡△normal-‡absentnormal-△{}^{\ddagger\triangle}start_FLOATSUPERSCRIPT ‡ △ end_FLOATSUPERSCRIPT Gargi Ghosh♠normal-♠{}^{\spadesuit}start_FLOATSUPERSCRIPT ♠ end_FLOATSUPERSCRIPT Mona Diab♠normal-♠{}^{\spadesuit}start_FLOATSUPERSCRIPT ♠ end_FLOATSUPERSCRIPT Asli Celikyilmaz♠normal-♠{}^{\spadesuit}start_FLOATSUPERSCRIPT ♠ end_FLOATSUPERSCRIPT ♠♠{}^{\spadesuit}start_FLOATSUPERSCRIPT ♠ end_FLOATSUPERSCRIPTMeta AI △△{}^{\triangle}start_FLOATSUPERSCRIPT △ end_FLOATSUPERSCRIPTWork done at Meta AI ‡‡{}^{\ddagger}start_FLOATSUPERSCRIPT ‡ end_FLOATSUPERSCRIPTMax Planck Institute、ETH
Abstract
Recent advancements in large language models have enabled them to perform well on complex tasks that require step-by-step reasoning with few-shot learning. However, it is unclear whether these models are applying reasoning skills they have learned during pre-training, or if they are simply memorizing their training corpus at finer granularity and have learned to better understand their context. To address this question, we introduce ALERT, a benchmark and suite of analyses for evaluating reasoning skills of language models. ALERT enables comparing pre-trained and finetuned models on complex tasks that require reasoning skills to solve them. Our benchmark provides a test bed to assess any language model on fine-grained reasoning skills, which spans over 20 datasets and covers 10 different reasoning skills. To prove the efficacy of ALERT we investigate the role of finetuning. Our extensive empirical analysis shows that language models acquire reasoning skills such as textual entailment, abductive reasoning, and analogical reasoning during the finetuning stage compared to pretraining stage. Another finding is when language models are finetuned they tend to overfit to the prompt templa
中文速览
大型语言模型（LLM）在推理任务上表现出色，但究竟是真的学会了推理，还是只是在记忆训练数据或特定的提示格式，目前仍不清楚。为此，研究者提出了 ALERT 这套评测基准，覆盖 20 多个数据集、10 种推理技能（如逻辑推理、因果推理、类比推理等），可以系统地对比预训练模型与微调模型在细粒度推理能力上的差异。通过对 OPT 系列模型进行大量实验，研究发现：微调并不是简单地让模型记住训练数据，而是确实能提升部分推理能力（尤其是类比推理和溯因推理），但对常识推理等技能帮助有限；此外，微调会让模型过度依赖特定的提示模板，降低泛化能力，而带有思维链（Chain-of-Thought）的微调方式能在一定程度上缓解这一问题。这项工作为理解 LLM 推理能力的真正来源提供了系统性工具和实证依据，有助于未来更有针对性地提升模型的推理泛化能力。
原文 arXiv:2212.08286;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2212.08286v2