An Incomplete Loop: Instruction Inference, Instruction Following, and In-context Learning in Language Models

Summary

This paper investigates whether LMs possess the types of reasoning necessary to robustly learn to perform and describe a task in-context. Specifically, the authors propose a reasoning loop consisting of deductive, inductive, and abductive reasoning, and then test whether LMs can perform each type of reasoning. Using coefficient prediction, nonce word translation, and machine translation tasks, the authors test the ability of models to follow instructions (deductive reasoning), generalize from few-shot prompts (inductive reasoning), and induce instructions given labeled examples. The contribution of this paper is clear and original, and it yields a more fine-grained understanding of what kinds of in-context reasoning, exactly, will be more difficult for LLMs to accomplish.

Rating

9

Confidence

4

Ethics flag

1

Reasons to accept

A1. The hypothesized reasoning structure is specific, falsifiable, and broken down into testable components. A2. The results provide evidence that these reasoning mechanisms are separate in LLMs. It is interesting that prediction accuracy is not closely related to the ability to perform instruction induction. A3. Reproducibility seems good. Detailed experimental setup details (including all prompts and hyperparameters) are provided in the Appendix. A4. The writing style and figures are clear. (The paper is also just a joy to read. I especially appreciate the frequent pre-21st-century citations.)

Reasons to reject

R1. Table 3 is opaque. What were the criteria for deciding whether a cell got a checkmark, partial check, or X? Were these qualitative judgments, or based on statistical significance tests? R2. The Kalamang translation results are weak, and don’t seem to provide especially strong evidence for either RQ. That said, this is a very difficult task, and I appreciate its inclusion; I therefore don’t think this is a significant weakness.

Questions to authors

Suggestions: - The Kalamang grammar row of Table 2 looks funny. This task setup is very different from the others, so maybe it makes more sense to simply put this in a separate table where the task setup would be better described using different columns. - In Figure 3, could you reorder the legend rows to match with the order of the bars? Also, I didn’t notice at first that the y-axis ranges were different across subfigures, which made me think that the Llama scores were better than they actually were. Maybe you could use the same [0, 1] range for all subfigures? - When linking to RQ1 and RQ2 (as on p.6), it might look better to exclude the colon. - Inconsistent references to subfigures within Figure 3. First it’s “3A” and “3B”, but in the next paragraph, it’s “3(A)” and “3(B)”.’ Typos: - Add period to end of Footnotes 4 and 10. - Why are “few-shot”, “zero-shot”, and “chain-of-thought” in \texttt{} format on p.3 but nowhere else in the paper? - Add period after “Appendix F” at end of "$\textbf{Hypotheses}$” on p.5. - Remove space before “Settings” at top of p.4.

Summary

This paper evaluates the performance of three styles of prompting, namely, instruction following, few-shot learning, and instruction inference, on four tasks. The main findings are: - instruction inference, which asks an LLM to generate hypotheses given examples and apply the best hypothesis to the input, is effective on simple tasks but not on complex tasks (low-resource translation and grammatical knowledge acquisision), and, - the relationship among the three styles of prompting is not observed.

Rating

6

Confidence

4

Ethics flag

1

Reasons to accept

The experimental results shown in the paper are useful and meaningful to see which style of prompting is effective on which tasks. However, it is difficult to obtain any general insights from the results and to generalize these results to other tasks.

Reasons to reject

As described above, the experimental results are meaningful but no generalized knowledge is obtained from these results. What we can know is that the effectiveness of different prompting styles is mixed, and this is kind of expectable. The main focus of the paper is on three styles of prompting, i.e., instruction following, few-shot learning, and instruction inference, rather than generic logical inference like deductive, inductive, and abductive reasoning. The paper describes these three methods as deductive, inductive, and abductive learning, while they are just a specific instance of these logical inference patterns. It sounds overgeneralization to use these logical terms in the title and the main description of the work. In other words, the paper does not provide any generic observations for deductive/inductive/abductive learning ability of LLMs.

Summary

This paper examines the relationship between instruction following, few-shot prompting, and instruction inference, linking them to deductive reasoning, inductive reasoning, and abductive reasoning, respectively. It explores these concepts across three domains: linear function learning, artificial language translation, and Kalamang translation. The findings suggest that instruction inference can surpass few-shot prompting in simpler tasks; however, its efficacy diminishes in more complex tasks such as low-resource machine translation. Additionally, the paper observes that a model's capability to generate hypotheses is not correlated with its ability to learn from few-shot examples.

Rating

6

Confidence

2

Ethics flag

1

Reasons to accept

The study conducts an extensive examination of the interplay between three prevalent reasoning paradigms, providing empirical insights that could be valuable for future research. This encourages the development of innovative methods for hypothesis verification and correction in natural language processing.

Reasons to reject

While the paper concludes that the ability to generate hypotheses is unrelated to learning from few-shot examples, it fails to suggest future directions or solutions to address this gap. This omission may limit the paper's impact on advancing the field's understanding of integrating these capabilities.

Summary

The paper investigates how modern large language models (LMs) utilize deductive, inductive, and abductive reasoning to learn tasks. The study explores three learning methodologies: instruction following, where tasks are explicitly described; few-shot prompting, where tasks are implied through examples; and instruction inference, where models generate task descriptions from examples before making predictions. Analyzing four LMs from the GPT and Llama families across tasks involving arithmetic functions and machine translation, they find an interesting dissociation in reasoning capabilities. LMs can effectively learn from few-shot examples without being able to articulate the underlying rules, and conversely, they can generate useful task descriptions yet fail to learn from explicit human-generated instructions. The findings suggest that while instruction inference can enhance performance in simpler synthetic setups, the integration of various reasoning types remains challenging in more complex tasks.

Rating

7

Confidence

3

Ethics flag

1

Reasons to accept

The paper is well-organized, and the results are presented clearly. This paper provides a meaningful first step to understanding the three learning methodologies of LLM.

Reasons to reject

The paper primarily focuses on specific simple tasks, linear function The scope of tasks examined in the study—linear function learning, artificial language translation, and Kalamang translation—is somewhat narrow, limiting the broader applicability of the findings. Future studies could benefit from incorporating a wider array of tasks to strengthen the arguments presented. There is room for deeper exploration regarding the limitations of the current methodologies. Discussing potential new domains for future research and outlining prospective directions would provide valuable insights.

Thanks for the responses.

I totally agree that it is meaningful to point out problems and you don't have to present complete solutions. This is still a weakness (given findings are not extremely surprising compared to e.g. adversarial examples), but understandable. I still think using overgeneralized terms is not a good practice (although we see many such papers recently). I prefer the title like "Investing instruction following, few-shot learning, and instruction inference..." that clearly describes what is done in the paper.

Thank you

Thanks for your response. Since I already gave a very high score and these weaknesses are mostly going to be the same (with clarification on R1), and taking the other reviews into account, I'm choosing keeping my score the same. I look forward to reading the final version of this work!

Acknowledgment of Response

After examining your responses as well as the remarks from other reviewers, I agree that the title may be an overstatement. It would be better to consider toning down the title.