Source-linked AI summary
Explanations from Large Language Models Make Small Reasoners Better
Shiyang Li, Jianshu Chen, Yelong Shen, Zhiyu Chen, Xinlu Zhang, Zekun Li, Hong Wang, Jing Qian, Baolin Peng, Yi Mao, Wenhu Chen, Xifeng Yan
TL;DR
The paper addresses how to use LLM-generated explanations to improve low-cost small reasoners, whose reasoning performance can lag behind LLMs. It explores three explanation-generation approaches within multi-task learning and reports consistent gains over finetuning baselines, including up to 9.5% higher accuracy than finetuning or prompting GPT-3 (175B) on CommonsenseQA.
Problem
The paper asks how explanations generated by large language models can improve small reasoners, which are more favorable for deployment because of their lower storage and computation costs.
Method
The method explores three LLM explanation-generation approaches and uses multi-task learning to train small models for both reasoning and explanation generation.
Results
Across multiple reasoning tasks and settings, the method consistently and significantly outperforms finetuning baselines and exceeds finetuning or prompting GPT-3 (175B) by up to 9.5% in accuracy on CommonsenseQA.
Takeaways & Limitations
The approach improves small-model reasoning while also producing high-quality explanations that justify predictions according to human evaluation.
Takeaways & Limitations
The work is limited to textual reasoning tasks and requires multi-task finetuning plus development-set tuning of weight α, adding hyper-parameter-tuning compute.
Abstract
from arXiv · showhide
Integrating free-text explanations to in-context learning of large language models (LLM) is shown to elicit strong reasoning capabilities along with reasonable explanations. In this paper, we consider the problem of leveraging the explanations generated by LLM to improve the training of small reasoners, which are more favorable in real-production deployment due to their low cost. We systematically explore three explanation generation approaches from LLM and utilize a multi-task learning framework to facilitate small models to acquire strong reasoning power together with explanation generation capabilities. Experiments on multiple reasoning tasks show that our method can consistently and significantly outperform finetuning baselines across different settings, and even perform better than finetuning/prompting a 60x larger GPT-3 (175B) model by up to 9.5% in accuracy. As a side benefit, human evaluation further shows that our method can generate high-quality explanations to justify its predictions, moving towards the goal of explainable AI.
1 Introduction
The paper asks whether explanations generated by large language models can improve small reasoners, then proposes explanation-generation and multi-task learning strategies to do so. Experiments report consistent gains over finetuning baselines, including a comparison with a much larger GPT-3 model.
- Approach: The paper uses explanations generated by large language models to improve the reasoning capabilities of small language models.The framework generates training explanations from LLM demonstrations and transfers them to small reasoners.
- Motivation: Small language models are attractive for deployment because their storage and computation costs are lower, but their reasoning performance can lag behind large language models.The paper frames closing this gap, especially on complicated reasoning tasks and in few-shot settings, as an open question.
- Approach: The authors systematically explore three approaches for generating explanations from large language models and combine them with multi-task learning.Multi-task learning supports partially generated explanations and avoids the weaker performance reported for self-rationalizing targets.
- Results: 8.1% in accuracy is the largest reported improvement over the standard finetuning baseline across the evaluated settings.The result is reported across explanation-generation and multi-task setups, with gains over standard finetuning baselines.
- Results: 9.5% in accuracy is the largest reported advantage over finetuning or prompting GPT-3 (175B) on CommonsenseQA.Human evaluation also finds that the method generates high-quality explanations to justify predictions.
2 Related Work
Prior work studies explanation-augmented prompting and learning with explanations, but this paper focuses on using noisy LLM-generated explanations to improve small reasoners without human filtering. It reports gains for several explanation-aware multi-task setups over standard finetuning.
- Prompting with Explanations: Chain-of-thought and related prompting methods add intermediate reasoning or explanations to improve large language model performance on reasoning tasks.The related work covers arithmetic, symbolic, numerical, commonsense, and few-shot reasoning settings.
- Prompting with Explanations: Earlier studies also use large language models to annotate explanations, sometimes with human filtering or calibration because generated explanations can be less factual.The paper distinguishes its setting by studying noisy explanations generated without human-involved filtering.
- Learning with Explanations: MT-Ra was previously applied to explanation learning, but earlier work did not observe gains over single-task finetuning in its reported setting.The paper compares this prior result with its own use of MT-Re, MT-Ra, and MT-CoT using LLM-generated explanations.
- Learning with Explanations: MT-Re, MT-Ra, and MT-CoT with LLM-generated explanations consistently and significantly outperform standard finetuning baselines without an accuracy-explanation trade-off.This result extends explanation-learning comparisons across the paper's multi-task setups.
3 Explanation Generation from LLM
The paper generates explanations for training instances with LLM prompting, using filtered chain-of-thought, rationalization, and a hybrid backup strategy. These explanations are then intended to improve small-model reasoning.
- Problem setup: The problem setup uses training pairs (x_i, y_i), a small set of human-written explanation examples, and generated explanations e_i for the training set.The stated goal is to use the generated explanations to improve small language model reasoning.
- COTE: Chain-of-thought prompting generates intermediate reasoning steps before the answer, using human-written examples as demonstrations for the LLM.The method greedily decodes an explanation for each training instance and then parses the decoded text as the explanation.
- COTE: COTE keeps a chain-of-thought explanation only when the LLM prediction matches the gold answer and otherwise sets the explanation to none.This filtering is intended to avoid retaining explanations associated with incorrect predictions.
- RP: Rationalization prompting generates explanations for every training instance while conditioning on the gold answer, without filtering the decoded explanation.The prompt concatenates demonstration examples with the current problem and answer before decoding.
- CROP: CROP uses rationalization prompting as a backup when COTE produces no explanation, combining prediction-based filtering with coverage of otherwise rejected instances.The hybrid is motivated by COTE's potentially higher-quality explanations and RP's ability to produce explanations for all instances.
4 Multi-task Learning with Explanations
The paper uses multi-task learning to train small models to answer questions and generate explanations, with MT-CoT explicitly linking answers to intermediate reasoning steps. MT-CoT supplements prior approaches and improves reasoning capability across two datasets.
- MT-Re and MT-Ra: Multi-task learning combines question-to-answer and question-to-reason tasks so small models learn predictions alongside explanations.MT-Re generates predictions and explanations in separate tasks, while MT-Ra conditions rationalizations on predictions.
- Training objective: The training objective mixes the question-to-answer and question-to-reason losses, with α weighting the two losses and tuned on a development set.The same objective is used for MT-Re, MT-Ra, and MT-CoT.
- MT-CoT: MT-CoT trains models to generate answers and chain-of-thought reasoning, explicitly modeling how answers are derived.This addresses limitations of MT-Re's lack of explicit answer-explanation interaction and MT-Ra's inability to access explanations when generating answers.
- MT-CoT: MT-CoT exposes models to intermediate reasoning steps before they produce answers, supplementing MT-Re and MT-Ra.The paper reports that MT-CoT consistently improves small-language-model reasoning capability and outperforms MT-Re and MT-Ra on two datasets.
5 Experiments
Experiments across three reasoning tasks evaluate explanation-based multi-task learning against single-task finetuning, few-shot settings, model sizes, larger LLMs, and explanation quality. The method consistently improves reasoning accuracy, including gains with limited data and competitive performance against much larger GPT-3 models.
- Experimental setup: Experiments cover CommonsenseQA, StrategyQA, and OpenbookQA, with results reported on development or test splits because some test sets are unavailable.CommonsenseQA requires commonsense reasoning, StrategyQA requires implicit multi-hop reasoning, and OpenbookQA combines open-book facts with broad knowledge and multi-hop reasoning.
- Main results: Three explanation-generation methods combined with multi-task learning consistently and significantly outperform single-task finetuning baselines.The best result for two of three datasets uses CROP, while two of the three dataset-best results use MT-CoT.
- Few-shot learning results: 6%-8% absolute accuracy gains over single-task finetuning occur on CommonsenseQA with 50, 100, or 200 training examples.On OpenbookQA, multi-task learning improves over single-task finetuning by about 4%-6% for 100, 200, or 400 examples.
- Results across model sizes: 0.7%-1.8% absolute accuracy gains on CommonsenseQA and 1.4%-3.0% on OpenbookQA persist across T5-small through T5-3B.Even T5-3B improves over its single-task counterpart by 2% absolute accuracy.
- Comparison with Large Language Models: 9.5% absolute accuracy is the maximum reported advantage over 60x larger GPT-3 finetuning and prompting on CommonsenseQA.The method still outperforms GPT-3 chain-of-thought prompting by 6% absolute accuracy on OpenbookQA, although it underperforms two other GPT-3 prompting baselines there.
- Human evaluation on generated explanations: 58% of T5 explanations are judged better than or competitive with GPT-3 explanations in the head-to-head human evaluation.The comparison uses 100 CommonsenseQA examples answered correctly by both models and three human annotators per example.
6 Conclusion
The paper uses LLM-generated explanations within multi-task learning to improve small reasoners across multiple reasoning tasks. The method outperforms finetuning baselines and can exceed a much larger GPT-3 model on CommonsenseQA while producing high-quality explanations.
- The method leverages LLM-generated explanations to improve small reasoners through multi-task learning.
- The approach consistently and significantly outperforms single-task finetuning baselines across explanation-generation methods, multi-task setups, training samples, and reasoner sizes.
- 9.5% in accuracy is the maximum improvement over finetuning or prompting a 60x larger GPT-3 on CommonsenseQA.
- Human evaluation finds that the model generates high-quality explanations compared with GPT-3, supporting more explainable AI.
Limitations
The approach has computational and scope limitations. It requires multi-task finetuning and development-set tuning, and the experiments are limited to textual reasoning problems.
- The approach requires multi-task learning finetuning to integrate explanations into small language models.
- Tuning weight α on a development set increases computation during hyper-parameter tuning.
- The work is constrained to textual reasoning problems and does not explore symbolic or arithmetic reasoning tasks.
Ethics Statement
Because the method incorporates explanations generated by LLMs, it may transfer their documented social biases into small models. Multi-task learning can disable explanation generation while retaining direct-answer performance gains.
- LLM-generated explanations may contain gender, race, and religion biases that can be integrated into small models during finetuning.
- Small models may expose these biases when generating explanations to justify predictions.
- The multi-task framework can disable explanation generation while retaining performance gains from direct answer prediction.
A Prompt details
The appendix presents prompts and example explanations used for CommonsenseQA, StrategyQA, and OpenbookQA. The examples show explanations that connect answer choices to commonsense, factual, temporal, physical, or definitional reasoning.
- Prompt sources: CommonsenseQA and StrategyQA prompts are based on prior work, while OpenbookQA explanations use science facts from the dataset repository.
- Prompt formats: The appendix includes standard, chain-of-thought, and rationalization prompts for CommonsenseQA, plus a chain-of-thought prompt for StrategyQA.
- CommonsenseQA: The prompt examples include reasoning about cable-dependent equipment, animal habitats, populated places, grocery carts, atlases, and ink absorption.
- CommonsenseQA: CommonsenseQA examples explain answers by identifying the relevant property or purpose and selecting the matching option.
- StrategyQA: StrategyQA examples answer binary questions by chaining factual or temporal premises before deriving yes-or-no conclusions.
B Explanation examples
Figure 4 presents three examples where both T5 and GPT-3 make incorrect predictions. In two examples, their predictions and explanations remain plausible despite differing from the gold labels.
- Figure 4 provides three examples in which both T5 and GPT-3 predict incorrectly.
- In examples (b) and (c), both models produce plausible predictions and explanations.
- The predictions in examples (b) and (c) differ from the gold labels despite their plausibility.