Rationale-Augmented Ensembles in Language Models

By Xuezhi Wang, Jason Wei, Dale Schuurmans, Dale Schuurmans, Quoc V. Le, Ed H. Chi, Denny Zhou. 2022

145 citations recorded by the source corpus.

Recent research has shown that rationales, or step-by-step chains of thought, can be used to improve performance in multi-step reasoning tasks. We reconsider rationale-augmented prompting for few-shot in-context learning, where (input ->output) prompts are expanded to (input, rationale ->output) prompts. For rationale-augmented prompting we demonstrate how existing approaches, which rely on manual prompt engineering, are subject to sub-optimal rationales that may harm performance. To mitigate this brittleness, we propose a unified framework of rationale-augmented ensembles, where we identify rationale sampling in the output space as the key component to robustly improve performance. This framework is general and can easily be extended to common natural language processing tasks, even those that do not traditionally leverage intermediate steps, such as question answering, word sense disambiguation, and sentiment analysis. We demonstrate that rationale-augmented ensembles achieve more accurate and interpretable results than existing prompting approaches--including standard prompting without rationales and rationale-based chain-of-thought prompting--while simultaneously improving interpretability of model predictions through the associated rationales.

Subfields: Artificial Intelligence, Prompting & Reasoning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Prompting & Reasoning

Citations in the corpus

ReAct: Synergizing Reasoning and Acting in Language Models (2022, neutral): “…arithmetic reasoning (“1844 < 1989”), guide 2We find more examples do not improve performance. a HotpotQA EM is 27.1, 28.9, 33.8 for Standard, CoT, CoT-SC in Wang et al. (2022b). b (Zhu et al., 2021; Lewis et al., 2020) search reformulation (“maybe I can search/look up x instead”), and synthesize…”
Large Language Models Can Self-Improve (2022, neutral): “More importantly, new capabilities have emerged from LLMs as they are scaled to hundreds of billions of parameters (Wei et al., 2022b): in-context few-shot learning (Brown et al., 2020) makes it possible for an LLM to perform well on a task it never trained on with only a handful of examples;…”
PAL: Program-aided Language Models (2022, methodology): “This approach of prompting the model to first generate a reasoning process t test improves the accuracy of the answer y test across a wide range of tasks (Wang et al., 2022a; Wei et al., 2022; Zhou et al., 2022; Wang et al., 2022b).”
GLM-130B: An Open Bilingual Pre-trained Model (2022, methodology): “Fortunately, one of the attempts on Post-LN initialized with the newly-proposed DeepNorm (Wang et al., 2022b) generates promising training stability.”
Multimodal Chain-of-Thought Reasoning in Language Models (2023, neutral): “A soft object changes shape when you squeeze it.”
Language Models are Multilingual Chain-of-Thought Reasoners (2022, background): “io/xlwic/ (7)One potential reason is that our prompts are not necessarily optimal (Wang et al., 2022) and may benefit from a broader investigation of other prompt formats.”
Decomposed Prompting: A Modular Approach for Solving Complex Tasks (2022, background): “This includes, for example, diverse decoding and ensembling to improve self-consistency (Wang et al., 2022b), more complex and contrasting rationales (Marasović et al., 2021; Wang et al., 2022a), generating sub-questions and answering them sequentially in least-to-most complexity order (Zhou et…”
Large Language Models Can Be Easily Distracted by Irrelevant Context (2023, background): “This means that in addition to improving model performance on clean arithmetic reasoning tasks (Wang et al., 2022c), SC also substantially reduces the distractibility of large language models to irrelevant context.”
A Comprehensive Overview of Large Language Models (2023, unknown): “Many of the multimodal tool augmentation systems employ multimodal LLMs [271], [272], [270], [261], while others utilize single modality LLMs and generate a plan on using different modality tools to solve multimodal queries [273].”
Teaching Algorithmic Reasoning via In-context Learning (2022, background): “…models (LLMs) have shown impressive progress in recent years, driven by the scaling up of models and training data sizes (Kaplan et al., 2020; Wei et al., 2022a; Hoffmann et al., 2022) that has led to improved performance and sample efﬁciency (Brown et al., 2020; Chen et al., 2021; Chowdhery…”
Complexity-Based Prompting for Multi-Step Reasoning (2022, methodology): “Handcrafted CoT constructed originally by Wei et al. (2022b) then reused in following-up works (Wang et al., 2022b; Kojima et al., 2022; Wang et al., 2022a).”
Measuring Faithfulness in Chain-of-Thought Reasoning (2023, background): “Wei et al. (2022) test three hypotheses for why CoT prompting provides a performance boost: that it produces an equation to be evaluated, that it provides additional test-time compute, and that it the CoT better enables the model to access relevant information from pretraining.”
A Survey of Large Language Models (2023, unknown): “Further, the authors in [253] expand the selfconsistency strategy to a more general ensemble framework (extending to ensemble on the prompts), and they find that diverse reasoning paths are the key to the performance improvement in CoT reasoning.”
“According to . . . ”: Prompting Language Models Improves Quoting from Pre-Training Data (2023, methodology): “B.1 Sources for SOTA Performance in Table 3 SOTA zero-shot results are from LLaMA 33B and 65B (Touvron et al., 2023), PaLM 540B (Wang et al., 2022a), and BART (Su et al., 2022) respectively.”