Can language models learn from explanations in context?

By Andrew Kyle Lampinen, Ishita Dasgupta, Stephanie C. Y. Chan, Kory Matthewson, Michael Henry Tessler, Antonia Creswell, James L. McClelland, Jane X. Wang, Felix Hill. 2022

383 citations recorded by the source corpus.

Language Models (LMs) can perform new tasks by adapting to a few in-context examples. For humans, explanations that connect examples to task principles can improve learning. We therefore investigate whether explanations of few-shot examples can help LMs. We annotate questions from 40 challenging tasks with answer explanations, and various matched control explanations. We evaluate how different types of explanations, instructions, and controls affect zero- and few-shot performance. We analyze these results using statistical multilevel modeling techniques that account for the nested dependencies among conditions, tasks, prompts, and models. We find that explanations can improve performance -- even without tuning. Furthermore, explanations hand-tuned for performance on a small validation set offer substantially larger benefits, and building a prompt by selecting examples and explanations together substantially improves performance over selecting examples alone. Finally, even untuned explanations outperform carefully matched controls, suggesting that the benefits are due to the link between an example and its explanation, rather than lower-level features. However, only large models benefit. In summary, explanations can support the in-context learning of large LMs on challenging tasks.

Subfields: Artificial Intelligence, Large Language Models

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 Large Language Models

Citations in the corpus

Emergent Abilities of Large Language Models (2022, positive): “A similar emergence in performance gain was also observed when augmenting few-shot prompting with explanations that came after the final answer (Lampinen et al., 2022).”
Large Language Models Can Self-Improve (2022, neutral): “Majority Voting by answer ···”
Training Language Models with Language Feedback (2022, methodology): “Prior work explores ways of using explanations to train text classiﬁcation models, with mixed re-sults (Camburu et al., 2018; Stacey et al., 2021; Pruthi et al., 2021; Wiegreffe et al., 2021; Hase and Bansal, 2021; Lampinen et al., 2022, inter alia ).”
STaR: Bootstrapping Reasoning With Reasoning (2022, background): “Recent work has shown that explicit intermediate reasoning (“rationales”) can improve large language model (LLM) performance as well [3–8].”
Reframing Human-AI Collaboration for Generating Free-Text Explanations (2021, background): “Others have proposed to generate explanations or clariﬁcations to improve task performance in a supervised (Rajani et al., 2019; Lampinen et al., 2022) or unsupervised (Shwartz et al., 2020) 2021) design hand-crafted templates which they use with mask-inﬁlling to produce contrastive explanations…”
Inner Monologue: Embodied Reasoning through Planning with Language Models (2022, background): “When appropriately conditioned (e.g., prompted), they can even carry out some degree of deduction and respond to questions that appear to require reasoning and inference [8, 9, 10, 11, 12, 13].”
Rationale-Augmented Ensembles in Language Models (2022, background): “…pretrained large language models has been achieved by expanding prompt exemplars with rationales, delivering successes in a variety of natural language reasoning tasks (Wei et al., 2022; Chowdhery et al., 2022; Lampinen et al., 2022; Creswell et al., 2022; Kojima et al., 2022; Jung et al., 2022).”
Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning (2022, background): “[7] show ICL can improve when explanations are provided and [8] use ICL with text retrieved from the web for open-domain question-answering.”
Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering (2022, background): “For example, the learning from instruction paradigm [40, 43, 53, 39, 45, 25], where the task level explanation is provided in the form of instruction, improves model performance significantly.”
Do Prompt-Based Models Really Understand the Meaning of Their Prompts? (2021, background): “In a similar vein, concurrent work by Lampinen et al. (2022) ﬁnds that other components of a prompt such as explanations of priming examples are helpful, but models are indif-ferent to whether the instructions in fact describe their tasks.”
The Unreliability of Explanations in Few-shot Prompting for Textual Reasoning (2022, background): “One appealing way to gain more insight into predictions obtained through in-context learning is to let the language model “explain itself” (Nye et al., 2021; Wei et al., 2022; Chowdhery et al., 2022; Marasović et al., 2022; Lampinen et al., 2022).”
Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking (2024, neutral): “We also visualize the distribution of improvements across tokens in the training set.”
Mind's Eye: Grounded Language Model Reasoning through Simulation (2022, background): “, 2022) is able to elicit LM reasoning in a more explicit way; however, its performance is reported to be sensitive to the quality of human annotated reasoning steps (Lampinen et al., 2022; Dasgupta et al., 2022) (i.”
The Unreliability of Explanations in Few-Shot In-Context Learning (2022, background): “One appealing way to gain more insight into predictions obtained through in-context learning is to let the language model “explain itself” (Nye et al., 2021; Wei et al., 2022; Chowdhery et al., 2022; Marasovi´c et al., 2022; Lampinen et al., 2022).”
Chain of Thought Empowers Transformers to Solve Inherently Serial Problems (2024, methodology): “…reasoning, the technique of including intermediate steps besides the final answers, through prompting (Wei et al., 2022; Chowdhery et al., 2022; Lampinen et al., 2022; Kojima et al., 2022; Zhou et al., 2023) or finetuning (Cobbe et al., 2021; Nye et al., 2021; Chung et al., 2022; Huang et al.,…”