Post Hoc Explanations of Language Models Can Improve Language Models

Large Language Models (LLMs) have demonstrated remarkable capabilities in performing complex tasks. Moreover, recent research has shown that incorporating human-annotated rationales (e.g., Chain-of-Thought prompting) during in-context learning can significantly enhance the performance of these models, particularly on tasks that require reasoning capabilities. However, incorporating such rationales poses challenges in terms of scalability as this requires a high degree of human involvement. In this work, we present a novel framework, Amplifying Model Performance by Leveraging In-Context Learning with Post Hoc Explanations (AMPLIFY), which addresses the aforementioned challenges by automating the process of rationale generation. To this end, we leverage post hoc explanation methods which output attribution scores (explanations) capturing the influence of each of the input features on model predictions. More specifically, we construct automated natural language rationales that embed insights from post hoc explanations to provide corrective signals to LLMs. Extensive experimentation with real-world datasets demonstrates that our framework, AMPLIFY, leads to prediction accuracy improvements of about 10-25% over a wide range of tasks, including those where prior approaches which rely on human-annotated rationales such as Chain-of-Thought prompting fall short. Our work makes one of the first attempts at highlighting the potential of post hoc explanations as valuable tools for enhancing the effectiveness of LLMs. Furthermore, we conduct additional empirical analyses and ablation studies to demonstrate the impact of each of the components of AMPLIFY, which, in turn, leads to critical insights for refining in-context learning.

Paper

Similar papers

Peer review

Reviewer mmJT7/10 · confidence 4/52023-06-30

Summary

This paper proposes an alternative prompting framework to Chain-of-thoughts, by using post-hoc explanations from proxy smaller LLMs. Specifically, given a query question and few-shot examples, the method firstly uses a proxy model to get post-hoc explanations on key input words. Then the key input words are combined with few-shot examples to form in-context prompt. The method is one of the first attempt to utilise post-hoc explanations to boost in-context learning performance. It doesn't require human annotation on reasoning intermediate steps, while outperforming CoT on several challenging tasks from the Big-Bench-Hard benchmark.

Strengths

1. Research on post-hoc explanation have mostly been explored to better understand model prediction. This work potentially opens up a new area of application, as it uses post-hoc explanation to boost in-context learning performance 2. The presentation is clear and easy-to-follow. The author provides informative ablations and comparisons to access the method from multiple aspects, which leads to an optimal settings with sound empirical evidences 3. The improvement with finetuned proxy model, is significant over CoT: 10-25% gain as highlighted in abstract.

Weaknesses

1. Although the improvement over CoT is significant and consistent with fully finetuned proxy model, the improvement becomes less in scale and less consistent with non-finetuned proxy model as shown in Table 2. In fact, if we take a closer look at table 1 & 2, we can see for GPT-3.5 AMPLIFY with non-finetuned proxy model, is worse than CoT on Formal Fallacies(48.3 / 54.6), CommonsenseQA(71.9 / 75.2), Coin Flip (55.4 / 61.0). Besides, such behaviour is not discussed in "Impact of Proxy Model Selection on LLM Performance" paragraph 2. The setting with finetuned proxy model, requires training data for the target task. This breaks the typical assumption of in-context learning where only a handful of annotated examples are available. Besides, it is desirable to show the performance of finetuned proxy model on these tasks (with E=0/10/200), to better assess the benefit of AMPLIFY

Questions

1. When finetuning the proxy model on target task, how many data points did you use? Can we have more information / analysis here with regard to the number of data points used for finetuning vs the number of few-shot examples used to prompt final LLM. It will be good if we can have a fair setting to compare with CoT / AO, where the proxy model is finetuned with the same set of data points as we use to prompt the final LLM. 2. Do you think there are tasks that CoT will work better in principle, like complex / multi-hop reasoning tasks? If so, do you think it's beneficial to combine AMPLIFY with CoT in those cases?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The paper has discussed the limitations properly

Reviewer 4Npw6/10 · confidence 4/52023-07-02

Summary

This paper demonstrates how post hoc explanations through a small LM could assist the performance of LLMs. The process is divided into 4 steps: Proxy Model Selection, Few-shot sample selection, compute explanations, and formatting prompts for LLMs. This technique automatically generates few-shot demonstrations, reducing human-annotation for few-shot in-context learning.

Strengths

- The method is sound and novel. - The paper contains extensive ablation experiments. - The improvement for some tasks is dramatic. (e.g. snarks)

Weaknesses

- This paper does not compare the performance with other baselines such as Auto-CoT that automatically generate demonstrations for in-context learning. Zhang et al (2022) Automatic Chain of Thought Prompting in Large Language Models - Although it is stated that the experimentation is specifically carried out on datasets aimed at evaluating complex linguistic understanding concepts, it omits datasets such as DisambiguationQA, Hyperbaton, and Word Soriting of Big-Bench-Hard benchmark. - When GPT-2 is not fine-tuned (E=0) in Table 2, AMPLIFY outperforms both AO and CoT on 3 out of 7 datasets for both GPT-3 and GPT-3.5, indicating that fine-tuning process is crucial. Also, the performance gets worse compared to AO for datasets such as formal fallacies, commonsenseQA, and ruin names for GPT-3.5.

Questions

- When fine-tuning GPT-2, what training datasets are used? As far as I know, tasks of Big-Bench-Hard do not contain training instances. - Do you expect that AMPLIFY would be still effective for models with less than 100B parameters? Also, do you think the effect of AMPLIFY could generalize to open-source LLMs? - What is the default s value (number of shots) for GPT-3, GPT-3.5 result of Table 1?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors adequately addressed the limitations and potential negative societal impact of their work.

Reviewer LsYX5/10 · confidence 4/52023-07-04

Summary

The paper presents AMPLIFY, an approach that uses post-hoc explanations from a proxy model to improve the prompting performance of large language models. For a given dataset, the approach assumes access to a set of labeled validation data that is used for crafting a prompt. First, the approach selects k examples that are misclassified by LLMs and exhibit high misclassification confidence scores. Next, post-hoc explanation techniques are used to find the most important input features of these selected examples, which are later used to construct template-based rationales. The rationales are used in the final prompts following the chain-of-thought paradigm. The paper conducts experiments on 7 datasets from Big-Bench-Hard. Experimental results suggest good improvements over the few-shot answer-only prompting (AO) baseline and the few-shot CoT baseline. The paper also includes analyses suggesting the importance of example selection strategies as well as the impacts of using different proxy models and explanation techniques.

Strengths

The paper studies an interesting problem that focuses on using post-hoc explanations for improving LLMs’ performance, which is relatively new. The experiments cover 7 datasets and show performance improvements compared to few-shot baselines. The paper provides some useful analyses, especially the analysis of the example selection strategies. The paper is well-written and easy to follow.

Weaknesses

1: The major comparison is unfair in some way The proposed approach uses a validation set (the size of the validation set is also not mentioned in the paper, nor in the appendix) and selects examples from the validation set. By contrast, standard prompting and chain-of-thought prompting typically only use few-shot examples. So part of the improvements might also be attributed to the use of more data in addition to the prompting method proposed in this paper (see weakness 2). In addition, the paper does not compare against approaches that can also use the validation set instead of just a few-shot examples. 2: However, the experiments do not clearly suggest the effectiveness of using post-hoc explanations The proposed approach uses the LLM itself + smaller proxy models to find more informative examples to be used in the prompts (the selection strategy). This active selection strategy does not rely on using post-hoc explanations and can be applied to answer-only prompting or chain-of-thought prompting as well. It would be good to provide AO and CoT performance using the selected examples as well to give a better understanding of the effectiveness of using better examples versus the effectiveness of using post-hoc explanations in prompts. The paper does provide analysis of the impacts of using different selection strategies. As shown in Table 3, using random examples, the performance of AMPLIFY is only 59.3 on GPT3 and 62.0 on GPT-3.5; at the same time, according to Table 1, CoT performance is 58.0 on GPT3 and 62.9 on GPT-3.5. IIUC, this suggests CoT and AMPLIFY is comparable; and selecting better examples contribute the majority of the performance improvements. While it seems that most of the improvements come from the selection strategy, the paper does not compare against other approaches that involve active annotating examples for in-context learning (e.g., Su et al. (2022); Diao et al. (2023)). Also, this proposed selective annotation strategy does not distinguish itself from existing work, which uses similar confidence-based ways to actively annotate examples (Su et al., 2022; Diao et al., 2023). Selective Annotation Makes Language Models Better Few-Shot Learners (Su et al., 2022) Active Prompting with Chain-of-Thought for Large Language Models (Diao et al., 2023) 3: The paper also misses some important details on the approach and the baseline Regarding the proposed approach: * When getting the post-hoc explanations. Does it explain the ground truth label or the predicted label? (it is not clearly explained in line 172) * How does the method get the initial predictions, in order to determine the misclassified examples? Is it obtained using few-shot AO or few-shot CoT? * It seems no information regarding the size of the validation set is provided. Regarding the baselines: * It seems the CoT baselines and AO baselines use few-shot prompting. How is the shot selected? Are they randomly selected? How many shots are used? * How are the CoTs for the CoT baseline written? Based on line 129, the CoTs are taken from Wei et al. (2022), but the original CoT paper does not include prompts for these datasets. Are they from Suzgun et al. (2022)? In addition, it would be also helpful if the appendix can include some prompt examples. 4: The paper uses a subset from big-bench-hard that mainly focuses on understanding linguistic concepts. On these datasets, CoT typically leads to no improvements or minor improvements, compared to AO. While it is understandable that the paper chooses to focus on understanding linguistics concepts, it would still be beneficial to benchmark the effectiveness of this approach on a broader range of multi-step reasoning tasks (e.g., GSM) where CoT shows substantial improvements over AO. And on these tasks, using rationales consisting of just top-k words may be less effective.

Questions

See weakness

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

The paper discusses the limitations in appendix.

Reviewer sqcU5/10 · confidence 4/52023-07-05

Summary

This paper proposes AMPLIFY framework that leverages post-hoc explanations to generate rationales automatically for chain-of-thought prompting. The framework consists of four stages: (1) adopt a light-weight model as the proxy to compute explanations, (2) select few-shot samples misclassified by LLM, (3) use attribution scores to identify the important words as rationales, (4) prompt LLM for predictions. Experimental results show that AMPLIFY outperforms the previous methods by a large margin across diverse tasks.

Strengths

1. The paper is well-written and well-motivated. 2. The idea of leveraging the supervision from the small models to improve LLM is novel. 3. The results on seven tasks demonstrate the effectiveness of the proposed method.

Weaknesses

Overall, the proposed method seems too heavy and its costs outweigh the performance benefits. 1. It requires an extra proxy model and fine-tuning on every target task in most cases, which can be impractical and laborious. 2. The second stage selects misclassified samples from the entire validation set and conducts filtering with a pre-defined metric, which breaks the constraints of "few-shot". 3. The third stage selects the top-k most important words to construct the rationale, which does not take into account the interactions between words that affect the model predictions. 4. The datasets they evaluated on are limited. The authors do not compare their method with CoT on more complex tasks like math problems and multi-hop reasoning. In contrast, CoT performs well under few-shot settings and only requires a few human-annotated rationales without any extra proxies and training costs.

Questions

1. Experiments show that fine-tuning is necessary in most cases (compared with "Random" baseline). What is the minimum size of a validation set to ensure that the explanations provided by the proxy model are reliable? (If it is too small, the proxy model is prone to overfitting.) 2. I'm curious about the performance of the proposed method on math problems compared with CoT. Is it general enough to improve the performance on various tasks?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The authors adequately address the limitations of their work.

Reviewer LsYX2023-08-11

Thank you for the response and clarification. I appreciate the added results for isolating the impacts of selecting better examples. I've raised by score.

Authorsrebuttal2023-08-21

Thank you for taking the time to review our response. We greatly appreciate your recognition of our work and for increasing the score.

Reviewer mmJT2023-08-15

Response to Author

Thank you for the thoughtful and detailed response! I agree that AMPLIFY will work better on tasks requiring "complex language understanding" and it's expected that for multi-step reasoning tasks like GSM8K, AMPLIFY will be less effective to CoT. It's a good practice that the authors share those results, which help us readers to understand better of the applicability of AMPLIFY and have a clearer understanding of its contribution. With the author's response, I perceive AMPLIFY as one of the first attempt on using post-hoc explanation for improving generation performance on tasks requiring "complex language understanding". It serves as a nice alternative prompting method to CoT as it doesn't require any manual design of CoT prompts. I will confidently keep my recommendation score unchanged based on this perception.

Authorsrebuttal2023-08-21

Thank you very much for reviewing our response! We will include the additional results in the final draft.

Reviewer sqcU2023-08-18

Thanks for the clarification. I will raise my score.

Authorsrebuttal2023-08-21

Thank you for taking the time to review our response. We greatly appreciate your recognition of our work and for increasing the score.

Reviewer 4Npw2023-08-21

Thank you for your response and sharing of additional results. I think the additional results would give more valuable insights to the reader. I will keep my score as "weak accept".

Authorsrebuttal2023-08-21

Thank you very much for reviewing our response! We will include the additional results in the final draft.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC