Summary
The paper presents a new perspective to understand the in-context learning behavior of large language models (LLMs) from the angle of latent concept learning resembling topic models. Based on the generative process defined by input data, latent concepts, and labels, the authors propose a two-stage algorithm to first learn the latent concept and then use it to select the best-performing demonstrations that boost in-context learning performance. Experiments on eight datasets show that the method is able to consistently outperform random selection and selection based on semantic similarity.
Strengths
* Originality: Although the latent concept learning perspective is largely inspired by Xie et al., the goal and method proposed in this paper for selecting effective demonstrations is still sufficiently interesting and different from Xie et al., which seem novel to me.
* Quality: The paper first states its generative assumptions defined by input data, latent concepts, and labels, based on which it then derives a two-stage method to first learn the concept and then select demonstrations. The theories and methods seem solid to me.
* Clarity: The paper is overall clear and nicely written.
* Significance: While it's nice to see that the method is able to consistently outperform simple baselines across the tasks selected, I generally feel that the setting considered in this paper is somewhat artificial, and that the evaluation is not comprehensive enough to testify the generalization ability of the method. These concerns weaken the significance of the paper. See weaknesses below for details.
Weaknesses
* Problem setting: The paper assumes access to a (relatively) large training set from which a few demonstrations can be drawn. With this amount of training data (e.g., 100), there could be better alternative choices than in-context learning. For example, one can tune an LLM with parameter-efficient methods and easily outperform in-context learning (shown in Liu et al.) without even having to consider how to select the best demonstrations. Of course, one can argue that in-context learning is applicable to non-open-source LLMs while training-based methods are not. However, given the recent growth in the availability of open-source LLMs such as LLaMA and Falcon, I believe it would be generally better to consider parameter-efficient tuning than in-context learning if the authors assume ~100 training samples are available. In summary, I feel that the problem setting of "carefully selecting good demonstrations from a larger training set" is somewhat artificial on its own from the very beginning.
* Evaluation tasks: The eight downstream tasks selected for evaluation appear to be too easy for nowadays LLMs, and it's unclear given the current evaluation how generalizable the method is to more challenging tasks like MMLU and reasoning, potentially when combined with chain-of-thought prompting.
* Typo: Line 104: "semantic analysis" -> "sentiment analysis"
Reference:
Liu et al. "Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning." NeurIPS (2022).
Questions
* What is the role of $\boldsymbol{\epsilon}$ in the generative process (Line 91)? It seems like it is not taken into account in the current modeling process.
* How does the method work for instruction-tuned models (e.g., ChatGPT, Alpaca)?
* Why are the results of LLaMA that bad (Figure 3a)? LLaMA is generally recognized as a better-performing model than OPT series under similar sizes.
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.
Limitations
Please see the Weaknesses section.