Self-Discover: Large Language Models Self-Compose Reasoning Structures

Summary

The paper introduces a new prompting method, Self-Discover, that enhances the complex reasoning abilities of Large Language Models. Self-Discover consists of two stages. At the first one, the method develops a task-adapted prompt by performing 3 steps: SELECT, when the LLM selects the reasoning modules (RMs) from a pre-defined list, ADAPT, when the model adapts the chosen reasoning modules to the task, and IMPLEMENT, when the model, conditioned on the adapted reasoning modules, generates the reasoning structure as a JSON. Each step is performed in a zero-shot manner and controlled by a dedicated meta-prompt. The resulting reasoning structure is used in the final second stage, as part of the prompt to solve each task and produce the output. For each task instance, the LLM first reasons by completing the JSON value fields, and then outputs the answer. Self-Discover is compared to several baselines: Direct Prompting, CoT (+Self-Consistency), Plan-and-Solve, Majority voting of RMs, and Best of each RM, as well as Graph-of-Thought (GoT) and Tree-of-Thought (ToT). Evaluation is performed on BIG-Bench Hard (BBH), Thinking for Doing (T4D), and (subsampled) MATH benchmarks. The results demonstrate the superiority of Self-Discover against the baselines above in terms of performance while maintaining comparatively low computational cost, expressed in the number of necessary inference calls. The ablation study highlights that each step of Stage 1 significantly contributes to the overall performance. The generated reasoning structures appear to be general and transferable between different LLMs, compared to Optimized Prompts (OPRO).

Strengths

- The paper proposes a simple yet effective approach that greatly improves the reasoning abilities of LLMs. This is especially emphasized by the low computational costs of the method of only 3 extra inference calls per task. The method unifies multiple prompting-reasoning techniques and thus is scalable and adaptable. - The evaluations are purposeful, and every major claim of the paper is supported by empirical evidence, including all the necessary ablation studies justifying the steps of Stage 1. A thorough analysis of the methods' performance, showing tasks where the method excels more, is included in the paper.

Weaknesses

- The paper doesn't emphasize enough that the method requires task adaptation. This is a crucial detail, as, unlike CoT or ToT, Self-Discover cannot be deployed as a general zero-shot agent, but requires some specialized prompting with task examples. This limitation is addressed barely in the paper. Though this is not a huge disadvantage, it places the method right between zero-shot methods (CoT / ToT / GoT / PS) and task-adaptive methods (OPRO / DSPy). Thus, more comparisons with the prompt adaptation techniques should be present in the paper, and the efficiency of Self-Discover should be emphasized when comparing to them. - The authors claim that their method "composes atomic reasoning models into a reasoning structure", but this seems misleading. Figure 4 demonstrates that the method only slightly outperforms the best RM baseline, suggesting that no actual "composition" is happening, but only a selection and adaptation of the best fitting reasoning technique. This requires further investigation, for example, by analyzing which reasoning modules were picked for a task, and ablating the performance on these modules. This would reveal whether the model indeed uses multiple reasoning modules or only one in its prediction.

Questions

- How robust is the method to changes in the list of reasoning modules? An ablation study on them, or at least the list of chosen modules for each task would reveal this information. Also, how robust is the method to the set of chosen task examples? What was their number in the experiments, and does varying their number or themselves affect the performance? - What were the parameters of ToT & GoT in the experiments? The value of the breadth parameter can largely affect the performance of these methods. - What modifications can be applied to the set of RM to increase the performance on the algorithmic tasks? I assume some specialized algorithmic prompting can be useful. What you also include the relationship between reasoning modules and the tasks in which they are useful, and assert that the model chooses the correct reasoning module?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

As mentioned above, though the method is very effective, it is dependent on the set of chosen task examples, and so is task-adaptable, and not completely zero-shot. This limitation is not emphasized enough in the paper. However, overall, it is a nice paper with promising results, which once again underscores the power of optimal prompting and alignment.

Summary

The paper introduces SELF-DISCOVER, a framework enabling large language models (LLMs) to autonomously identify and utilize intrinsic reasoning structures to tackle various reasoning tasks. The framework is applicable across different model families and aligns with human reasoning patterns.

Strengths

1. The paper introduces a novel framework, SELF-DISCOVER, that allows LLMs to autonomously identify and use task-specific reasoning structures. 2. The method is effective across various benchmarks.

Weaknesses

1. The framework's effectiveness is heavily dependent on the quality and comprehensiveness of the atomic reasoning modules available, which may require significant human effort to define. Additionally, there is no guideline provided on how to develop these modules. 2. The main results are reported using a very limited range of models, specifically GPT-4 and PaLM-2. A systematic evaluation across a broader range of models, such as GPT-3.5, Llama2, and Llama3, is necessary, as merely applying reasoning structures generated by GPT-4 to these models is insufficient. 3. The compared baselines, such as zero-shot CoT and Plan-and-Solve, are not strong enough. Automatic prompting engineering methods, like LLM as Optimizers [1] and StrategyLLM [2], should be included in comparisons and discussions. [1] Large Language Models as Optimizers. https://arxiv.org/abs/2309.03409 [2] StrategyLLM: Large Language Models as Strategy Generators, Executors, Optimizers, and Evaluators for Problem Solving. https://arxiv.org/abs/2311.08803

Questions

NA

Rating

4

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes

Summary

This paper introduces SELF-DISCOVER, a framework designed to enhance the reasoning capabilities of Large Language Models (LLMs) such as GPT-4 and PaLM 2. The framework enables LLMs to self-discover and compose intrinsic reasoning structures tailored to specific tasks, improving performance on complex reasoning benchmarks.

Strengths

- By enabling LLMs to self-compose reasoning structures, this method significantly improves performance on complex reasoning tasks, demonstrating the ability to handle intricate problems more effectively than traditional methods. - SELF-DISCOVER achieves superior results while requiring substantially fewer inference steps (10-40x fewer) compared to inference-intensive approaches like CoT-Self-Consistency. - The reasoning structures discovered by the SELF-DISCOVER method show strong transferability across different LLM families, highlighting its broad applicability and robustness.

Weaknesses

1. While atomic reasoning structures form the foundation of the SELF-DISCOVER method, their inherent simplicity can restrict the overall reasoning performance. These basic units, although useful, may not fully capture the complexity needed for certain advanced reasoning tasks, potentially limiting the depth and flexibility of the reasoning process. 2. The proposed method is significantly reliant on the performance of the underlying LLM. Factors such as the model's ability to follow instructions accurately and the limitations imposed by context window sizes can pose significant challenges. For instance, if the base LLM struggles with instruction-following or is constrained by a limited context window, the effectiveness of the SELF-DISCOVER method may be compromised, limiting its potential for broader application and scalability. 3. The iterative process of selecting, adapting, and implementing reasoning modules introduces extra computational costs. Each step in this process requires additional queries to the LLM, which can accumulate and result in increased computational demands. This overhead may offset some of the efficiency gains achieved through reduced inference steps, particularly in scenarios requiring frequent adaptation and customization of reasoning structures.

Questions

1. What is the additional cost of implementing this method, such as the extra tokens generated or the increase in inference time? 2. Please refer to the correct section number in the Appendix in lines 83-84. 3. The comparison in Figure 5 may not be fair. The self-discovery method requires fewer calls but could lead to more token generations. 4. The design of the "select, adapt, and implement" steps seems ad-hoc. Why are these three steps proposed separately? For instance, why can't "select" and "adapt" be combined into a single step since their prompts are similar, as shown in Figure 9? 5. What are the commonly selected reasoning structures? Are all 39 reasoning modules useful? It would be better to show the density of the selected reasoning modules to ensure that the selection of 39 modules is meaningful. 6. Does the model perform consistently in the selection process? 7. Do different models, such as GPT-4 and Palm-2, perform consistently? Do they discover similar reasoning structures?

Rating

5

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

This paper does not sufficiently illustrate the limitations of the proposed method. The checklist on lines 542-543 states, "We present an extensive analysis of the limitations of Self-Discover in both Sections 4 and 5 with an extended error analysis on MATH in Appendix E." However, Sections 4 and 5 are primarily general evaluations of the proposed method and do not contain sufficient and clear illustrations of its limitations. This issue needs to be strictly addressed.

Summary

The paper introduces SELF-DISCOVER, a novel framework that enables large language models (LLMs) to self-discover and compose reasoning structures for tackling complex reasoning tasks. The core of SELF-DISCOVER is a self-discovery process where LLMs select multiple atomic reasoning modules, such as critical thinking and step-by-step thinking, and compose them into an explicit reasoning structure to follow during decoding.

Strengths

1) SELF-DISCOVER significantly improves the performance of state-of-the-art LLMs on complex reasoning tasks, showcasing substantial gains over traditional prompting methods like Chain of Thought (CoT). 2) The discovered reasoning structures are not only effective but also transferable across different model families, indicating a level of universality in the approach. SELF-DISCOVER provides a more interpretable way to understand the reasoning process of LLMs by generating explicit reasoning structures. 3) The paper includes a thorough empirical evaluation on a diverse set of reasoning tasks, demonstrating the framework's effectiveness across various domains.

Weaknesses

1) The paper does not specify the exact number of reasoning structures that can be discovered by the SELF-DISCOVER framework. It's unclear whether the framework can generate a wide variety of structures or if it tends to converge on a limited set of common structures. 2) It is not detailed whether different examples within the same task utilize the same reasoning structure or if the framework can adapt the structure to suit the nuances of individual examples. 3) The paper's comparative analysis may be considered basic, as it does not delve into a comprehensive set of existing methods or include recent state-of-the-art approaches for comparison. 4) Although the paper mentions a similarity to the "Meta Reasoning for Large Language Models" approach, it does not provide a direct comparative experiment or analysis to highlight differences and potential advantages of SELF-DISCOVER. 5) The paper might not fully account for or cite the most recent literature, specifically works from 2024 and beyond, which could provide additional context and comparison points for the SELF-DISCOVER framework.

Questions

See weaknesses.

Rating

4

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

See weaknesses.

Summary

This paper proposes a prompt engineering scheme: Given a task and 39 prompts for solving tasks, an LLM is prompted to select which of the 39 prompts are suitable for the task. The selected prompts are then rephrased to be more specific to the task, and reformatted in json. The paper uses GPT-4 Turbo, PaLM 2-L, and Llama2-70B and compares direct prompting, chain of thought, plan and solve, chain of though with self-consistency, majority vote of the 39 prompts, and the proposed method.

Strengths

1. This prompting scheme performs best on tasks that require diversity, and the work demonstrates the value of prompt diversity. 2. Ablation experiments show the increased accuracy for each of the three steps of selecting prompts, rephrasing them, and reformatting (or implementation), over chain of thought.

Weaknesses

1. As a prompt engineering method, it is unclear if the choice of the specific 39 prompts will withstand the test of time across LLMs and tasks. 2. Recent prompt engineering tools, such as Anthropic's Claude 3.5 Prompt Engineer, receive a task as input and output a detailed prompt best suited for the task and LLM, which may supersede this approach. 3. The claim of efficiency should be more rigorously validated and consider both the number of calls and the token lengths of the prompts and responses of each call.

Questions

When comparing prompting schemes, it is important to give equal compute time to different methods. When measuring efficiency did the work consider the lengths of the prompts and responses?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes