What Makes Good In-Context Examples for GPT-$3$? (2101.06804v1)

Abstract: GPT-$3$ has attracted lots of attention due to its superior performance across a wide range of NLP tasks, especially with its powerful and versatile in-context few-shot learning ability. Despite its success, we found that the empirical results of GPT-$3$ depend heavily on the choice of in-context examples. In this work, we investigate whether there are more effective strategies for judiciously selecting in-context examples (relative to random sampling) that better leverage GPT-$3$'s few-shot capabilities. Inspired by the recent success of leveraging a retrieval module to augment large-scale neural network models, we propose to retrieve examples that are semantically-similar to a test sample to formulate its corresponding prompt. Intuitively, the in-context examples selected with such a strategy may serve as more informative inputs to unleash GPT-$3$'s extensive knowledge. We evaluate the proposed approach on several natural language understanding and generation benchmarks, where the retrieval-based prompt selection approach consistently outperforms the random baseline. Moreover, it is observed that the sentence encoders fine-tuned on task-related datasets yield even more helpful retrieval results. Notably, significant gains are observed on tasks such as table-to-text generation (41.9% on the ToTTo dataset) and open-domain question answering (45.5% on the NQ dataset). We hope our investigation could help understand the behaviors of GPT-$3$ and large-scale pre-trained LMs in general and enhance their few-shot capabilities.

What Makes Good In-Context Examples for GPT-3?

The paper "What Makes Good In-Context Examples for GPT-3?" addresses an important aspect of leveraging LLMs, specifically focusing on the impact of in-context examples on the performance of GPT-3. The research scrutinizes how in-context examples, which are crucial for few-shot learning, can be selectively chosen to optimize GPT-3's performance across multiple natural language understanding and generation tasks.

Key Contributions and Findings

1. Problem Identification: The paper identifies a critical issue concerning the selection of in-context examples in the few-shot learning paradigm of GPT-3. Prior observations indicated that GPT-3’s performance can vary significantly depending on the chosen examples, thereby implying room for improving the in-context example selection process.
2. Proposed Method - KATE: The authors introduce KATE (k-Nearest Neighbor-Augmented in-conText Example selection), a retrieval-based method designed to select in-context examples that are semantically similar to the test samples. The method leverages pre-trained sentence encoders to determine the semantic similarity.
3. Empirical Validation: KATE is empirically validated on multiple benchmarks, including:
   • Sentiment analysis (between SST-2 and IMDB datasets)
   • Table-to-text generation (using the ToTTo dataset)
   • Open-domain question answering (NQ, WQ, and TriviaQA datasets)
4. Superior Performance: The KATE methodology consistently outperformed random sampling baselines across all evaluated tasks. Notable improvements include:
   • On the SST-2 to IMDB sentiment analysis task, KATE$_{\text{sst-2}}$ achieved an accuracy of 93.43%, substantially higher than the random baseline.
   • For table-to-text generation on the ToTTo dataset, KATE improved BLEU scores to 40.3, compared to 28.4 from random selection.
   • In question answering, KATE's performance gains were significant, reaching an EM score of 41.6 on NQ, compared to 28.6 for random selection.
5. Impact of Sentence Encoders: The research demonstrated that the effectiveness of KATE improves when sentence encoders are fine-tuned on task-related datasets. For instance, encoders fine-tuned on NLI and STS-B datasets resulted in better retrieval quality and improved GPT-3’s performance.
6. Ablation Studies: Detailed ablation studies were conducted to analyze the effect of various factors:
   • The number of in-context examples: Performance improved with a higher number of examples.
   • Size of the retrieval training set: Larger retrieval sets led to better performance.
   • Order of in-context examples: The order had a minor impact, indicating robustness to this factor.

Implications and Future Directions

Practical Implications

• Enhanced Few-Shot Learning: By improving the selection of in-context examples, practical applications relying on GPT-3, such as customer support chatbots, recommendation systems, and information retrieval, can achieve more consistent and higher performance without additional fine-tuning.
• Efficiency: The ability to enhance GPT-3’s performance via better example selection without increasing computational overhead for fine-tuning is particularly beneficial for deployments in resource-constrained environments.

Theoretical Implications

• Understanding Model Behavior: This work contributes to a nuanced understanding of how LLMs like GPT-3 internalize and utilize contextual information. This can pave the way for more refined models that inherently manage context better.
• Retrieval-augmented Models: The findings bolster the efficacy of retrieval-augmented frameworks, suggesting that integrating such approaches can be beneficial for future generations of LLMs.

Speculative Future Developments

• Dynamic In-Context Learning: Future models might dynamically select or generate in-context examples as part of their inference process, thus adapting better to the specific characteristics of each test sample.
• Meta-Learning: Combining KATE with meta-learning paradigms could further enhance adaptability and performance consistency.
• Interactive Model Training: Interactive systems where a model can query a database for the best context in real-time could further optimize performance, particularly in open-domain tasks.

Conclusion

The research presents a substantial advancement in optimizing GPT-3's few-shot learning capability by refining in-context example selection. By demonstrating the practicality and benefits of the KATE methodology across several benchmarks, the paper provides a robust framework for improving the utility of large-scale pre-trained LLMs without the need for additional fine-tuning, thus opening up new avenues for both practical applications and theoretical explorations.