Conceitos Básicos
Self-AMPLIFY is a novel framework that leverages post-hoc explanation methods to automatically generate rationales from small language models, enabling them to self-improve their performance on complex reasoning tasks.
Resumo
The paper introduces Self-AMPLIFY, an extension of the AMPLIFY framework, which aims to improve the performance of small language models (SLMs) through the automatic generation of rationales. Unlike AMPLIFY, Self-AMPLIFY does not rely on an auxiliary proxy model or human-annotated rationales, but instead generates the rationales directly from the SLM itself using various post-hoc explanation methods.
The key steps of Self-AMPLIFY are:
n-shot Sample Selection: Two strategies are implemented to select promising input texts to be included in the final prompt - the "success" strategy selects correctly predicted instances, while the "error" strategy selects misclassified instances.
Rationale Generation: Three types of post-hoc explanation methods are used to generate rationales from the SLM: post-hoc attributions (DeepLift and KernelSHAP), self-topk explanations, and self-natural language explanations.
Prompt Design for SLMs: The final prompt is built by incorporating the generated rationales between the input text and the ground truth answer, following the (x, r, y) template.
The authors evaluate Self-AMPLIFY on five datasets and three SLMs, comparing it to traditional prompting, Auto-CoT, and AMPLIFY. The results show that Self-AMPLIFY achieves good performance gains, often outperforming the competitors, especially when using the self-natural language explanations. The authors also conduct an ablation study on the impact of the different post-hoc explainers, finding that the topk methods generally perform well.
However, the authors note that the performance of Self-AMPLIFY is limited on a smaller 2 billion parameter SLM, suggesting that the reasoning abilities of the model play a key role in the effectiveness of the self-generated rationales.
Estatísticas
"The 7 billion parameters models Mistral and Zephyr achieve state-of-the-art performance among other SLMs in a wide variety of NLP tasks."
"Self-AMPLIFY induces up to a 17.6 points accuracy improvement when applied to Causal Judgment with the error strategy and the DeepLift topk post hoc explainer."
Citações
"Self-AMPLIFY is the first method to generate rationales without the use of any auxiliary side model."
"Self-AMPLIFY is the first approach enriching the prompt without human-annotated rationales or the use of auxiliary models, but only with the SLM itself."