PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Pei Zhou, Jay Pujara, Xiang Ren, Xinyun Chen, Heng-Tze Cheng, Quoc V. Le, Ed H. Chi, Denny Zhou, Swaroop Mishra, Huaixiu Steven Zheng

arXiv:2402.03620v1cs.AIcs.CL

TL;DR

Typical prompting methods use individual reasoning techniques, while tasks may require different intrinsic reasoning structures. SELF-DISCOVER composes such structures from atomic modules and applies them during decoding. It outperforms CoT across challenging benchmarks, while using substantially less inference compute than inference-intensive methods.

  • Problem

    Common prompting techniques use a single atomic reasoning module and thus assume a particular process for tackling a task, although tasks can have distinct intrinsic reasoning structures.

    Source: §1 · p. 1

  • Method

    SELF-DISCOVER selects, adapts, and implements atomic reasoning modules into a task-specific structure, then uses it to solve task instances.

    Source: §1 · p. 1

  • Results

    SELF-DISCOVER outperformed CoT on 21/25 challenging tasks, with gains up to 42%, and required 10–40x fewer inference compute than inference-intensive methods.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The discovered structures transferred across models from PaLM 2-L to GPT-4 and from GPT-4 to Llama-2-70B.

    Source: §1 · p. 2, §7 · p. 8

    The authors observed low-quality structure outputs when trying zero-shot meta-prompting with Llama2.

    Source: §6.1 · p. 8

Abstract

from arXiv · show

We introduce SELF-DISCOVER, a general framework for LLMs to self-discover the task-intrinsic reasoning structures to tackle complex reasoning problems that are challenging for typical prompting methods. Core to the framework 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 for LLMs to follow during decoding. SELF-DISCOVER substantially improves GPT-4 and PaLM 2's performance on challenging reasoning benchmarks such as BigBench-Hard, grounded agent reasoning, and MATH, by as much as 32% compared to Chain of Thought (CoT). Furthermore, SELF-DISCOVER outperforms inference-intensive methods such as CoT-Self-Consistency by more than 20%, while requiring 10-40x fewer inference compute. Finally, we show that the self-discovered reasoning structures are universally applicable across model families: from PaLM 2-L to GPT-4, and from GPT-4 to Llama2, and share commonalities with human reasoning patterns.

1. Introduction

SELF-DISCOVER addresses the assumption that one fixed reasoning module suits every task by composing task-specific structures from atomic modules. The paper reports stronger performance across challenging benchmarks, alongside analyses of task categories, computation errors, and cross-model transfer.

1. Introduction

  • SELF-DISCOVER selects and composes atomic reasoning modules into a task-intrinsic structure, then uses that structure to solve task instances.The framework is motivated by the limitation of treating individual prompting techniques as fixed, task-agnostic modules.Source: §1 · p. 1
  • It requires only three additional task-level inference steps and uses 10–40x fewer inference compute than inference-intensive approaches.The paper also reports performance exceeding methods such as self-consistency and majority voting, and performance on par with or better than OPRO while remaining more interpretable.Source: §1 · p. 2
  • On 21/25 challenging tasks, SELF-DISCOVER outperformed CoT, with gains up to 42%.The tests covered BBH, T4D, and MATH; Figure 1 summarizes comparisons on challenging benchmarks.Source: §1 · p. 2, Figure 1 · p. 2
  • SELF-DISCOVER performs best on world-knowledge tasks, shows moderate gains on algorithmic tasks, and 74.7% of MATH model failures stem from computation errors.The paper also reports transferability from PaLM 2-L to GPT-4 and from GPT-4 to Llama-2-70B.Source: §1 · p. 2

2. Self-Discovering Reasoning Structures for Problem-Solving

SELF-DISCOVER first constructs a task-specific reasoning structure from general problem-solving skills, then applies that structure to task instances. Its process selects useful modules, adapts them to the task, and implements them as an actionable plan.

2. Self-Discovering Reasoning Structures for Problem-Solving

  • The framework has two stages: task-level structure discovery and solving task instances by following the discovered structure.The motivating analogy is that people draw on prior knowledge and skills, apply relevant ones, and connect them to solve a problem.Source: §2 · pp. 2–3

2.1. Stage 1: Self-Discover Task-Specific Structures

  • The model discovers a task-specific structure from unlabeled examples and module descriptions, formatting the structure as interpretable key-value pairs similar to JSON.The method requires no labels or training and is run once at the task level.Source: §2 · pp. 2–3
  • Stage 1 uses three actions: SELECT useful modules, ADAPT their descriptions to the task, and IMPLEMENT them as an actionable reasoning structure.For implementation, the method uses a human-written structure from another task as a demonstration.Source: §2.1 · p. 3

2.2. Stage 2: Tackle Tasks Using Discovered Structures

  • In Stage 2, the model appends the discovered structure to each task instance and follows it step by step to generate an answer.The structure is reused to solve every instance of the given task.Source: §2.2 · p. 3, §2 · p. 3, Figure 2 · p. 3

3. Experiment Setup

The experiments evaluate SELF-DISCOVER on challenging reasoning tasks using several LLMs and compare it with prompting, reasoning, ensemble, and prompt-optimization baselines. The evaluation spans BBH, T4D, and MATH, with some comparisons restricted to task subsets.

3.1. Tasks

  • The evaluation covers 23 BBH tasks across four reasoning categories, grounded social-agent reasoning on T4D, and 200 sampled MATH test examples.MATH uses instance-level reasoning structures generated through a one-shot demonstration; accuracy is the evaluation metric for all three benchmarks.Source: §3.1 · pp. 3–4

3.2. Models

  • The evaluated models are GPT-4, GPT-3.5-turbo, instruction-tuned PaLM 2-L, and Llama2-70B.For MATH, the study uses a PaLM 2-L model with stronger instruction tuning to support following more complex structures.Source: §3.2 · p. 4, §3.3 · p. 4

3.3. Baselines

  • Zero-shot baselines include direct prompting, CoT, and Plan-and-Solve; the latter asks models to plan before solving.SELF-DISCOVER differs from Plan-and-Solve by grounding its explicit key-value structure in atomic reasoning modules.Source: §3.3 · p. 4
  • On a subset of tasks, the study compares SELF-DISCOVER with CoT self-consistency and majority voting across individual reasoning modules.These baselines use repeated queries or module-specific answers, and are included to examine inference-call efficiency and the value of integrating modules into one structure.Source: §3.3 · p. 4
  • The study also compares with OPRO, which optimizes prompts using a training set, to assess transferability and interpretability.The comparison tests whether structures or prompts optimized from one model retain performance gains when applied to another.Source: §3.3 · p. 5

4. Results

The results examine whether SELF-DISCOVER improves reasoning, which task types benefit, and how its efficiency and generated structures compare with other prompting methods.

4.1. Does SELF-DISCOVER Improve LLM Reasoning?

  • On 23 BBH tasks, SELF-DISCOVER gains 7% and 6% absolute over CoT and Plan-and-Solve on PaLM 2-L, and 6% and 8% on GPT-4.Source: §4.1 · p. 5
  • SELF-DISCOVER outperforms direct answering and CoT on over 20/24 tasks in the PaLM 2-L breakdown.Source: §4.1 · p. 5
  • On grounded social agent reasoning, SELF-DISCOVER improves over all baselines by at least 27% on PaLM 2-L and 32% on GPT-4.Source: §4.1 · p. 5

4.2. Which Types of Problems Do SELF-DISCOVER Help the Most?

  • SELF-DISCOVER improves over direct answering and CoT across four task categories, with its strongest gains on tasks requiring diverse world knowledge.Source: §4.2 · p. 5, Figure 4 · p. 5

4.3. How Efficient is SELF-DISCOVER?

  • On GPT-4, SELF-DISCOVER outperforms inference-heavy methods while using 10–40x fewer inference calls; it uses one call per instance plus three task-level calls.Source: §4.3 · p. 6, Figure 5 · p. 6

4.4. Qualitative Examples

  • Examples show task-adapted structures combining reasoning modules, while a geometric-shape example contrasts their correct coordinate reasoning with errors by CoT and Plan-and-Solve.Source: §4.4 · p. 6, Figure 6 · p. 6

5. Deep Diving Into Self-Discovered Reasoning Structures

The analysis tests the contribution of SELF-DISCOVER’s three actions and whether discovered structures transfer across models.

5. Deep Diving Into Self-Discovered Reasoning Structures

  • In a BBH geometric-shape example, SELF-DISCOVER traces coordinates to recognize a closed path, whereas CoT and Plan-and-Solve reach a wrong answer.Source: Figure 7 · p. 7

5.1. Importance of SELF-DISCOVER Actions

  • With each action—SELECT, ADAPT, and IMPLEMENT—GPT-4’s zero-shot reasoning improves consistently across four tasks, with all three actions giving the most gain.Source: §5.1 · p. 7, Figure 8 · p. 7

5.2. Towards Universality of Discovered Reasoning Structures

  • Applying PaLM 2-L-discovered structures to GPT-4 beats OPRO on three of four tasks, although OPRO used 20% data to optimize its prompts and SELF-DISCOVER was zero-shot.Source: §5.2 · pp. 7–8
  • GPT-4-discovered structures outperform CoT on Llama2-70B disambiguation QA (52% vs. 42%) and GPT-3.5-turbo geometry (56% vs. 51%).Source: §5.2 · p. 8

6. Related Work

Related work covers prompting methods for reasoning and planning; SELF-DISCOVER is presented as a way to compose reasoning approaches without task labels or human-annotated plans.

6.1. Prompting Methods

  • The related methods include few-shot and instruction prompting, CoT, decomposition approaches, and search-based methods such as Tree-of-Thought and Graph-of-Thought.Source: §6.1 · p. 8
  • SELF-DISCOVER composes across prompting methods as a program combines basic building blocks, addressing the fact that prompting methods have different strengths and application domains.Source: §6.1 · p. 8

6.2. Reasoning and Planning

  • Unlike related approaches that require human annotation of skills or plans, SELF-DISCOVER uses LLM meta-task reasoning to combine approaches without task labels.Source: §6.2 · p. 8

7. Conclusion

SELF-DISCOVER is presented as an efficient framework for discovering task-specific reasoning structures, with reported improvements on challenging benchmarks and transferability across LLMs. Its prompt details describe how the structure is discovered and then used to solve task instances.

7. Conclusion

  • The conclusion presents SELF-DISCOVER as an efficient framework for discovering reasoning structures from general problem-solving skills, with benchmark improvements up to 30% and transferability across LLMs.The conclusion says ablations demonstrate that the composed structures are universally transferable between LLMs.Source: §7 · p. 8

A. Self-Discover Prompt Details

  • The prompt details list 39 reasoning modules, adopted from Fernando et al. (2023), as cognitive heuristics for problem-solving.Source: A. Self-Discover Prompt Details · p. 12
  • Stage 1 uses three actions to discover a task-level reasoning structure, and Stage 2 has the model follow that structure to answer task instances.The Stage 2 prompt instructs the model to fill in values for the plan's keys by reasoning about the specific task.Source: A. Self-Discover Prompt Details · p. 12
  • The three Stage 1 meta-prompts combine instructions, reasoning-module descriptions, and unlabeled task examples; IMPLEMENT includes a human-written plan for another task.The prompts guide discovery of an intrinsic reasoning structure at the task level.Source: Figure 10 · p. 12, A. Self-Discover Prompt Details · p. 12

B. Evaluation Details

The evaluation uses accuracy and exact matching on BBH, T4D, and MATH, with manual checks to validate answer extraction for a MATH subsample.

B. Evaluation Details

  • Accuracy and exact matching are used on BBH, T4D, and MATH; because MATH answers were difficult to extract accurately, 200 test examples were subsampled and manually checked and annotated.Source: B. Evaluation Details · p. 12

C. BBH Per Task Performance

The section presents per-task BBH performance in Table 3.

C. BBH Per Task Performance

D. Error Analysis

The error analysis distinguishes errors in discovered reasoning structures from errors in intermediate calculations. Most wrong predictions in the analyzed sample are attributed to calculation errors rather than incorrect structures.

D. Error Analysis

  • A reasoning structure counts as correct when a human expert can solve the task by following it.Source: D. Error Analysis · p. 12
  • In the 200-example MATH analysis, 87.5% (175) had correct reasoning structures and 12.5% (25) had incorrect structures leading to prediction errors.The incorrect structures included task misunderstandings, step errors, or unnecessary steps.Source: D. Error Analysis · p. 14
  • Among 99 wrong predictions, 74.7% of errors came from intermediate calculations, while 25.3% were attributed to wrong reasoning structures.The authors suggest improving step-wise calculation accuracy, for example with tools or code generation.Source: D. Error Analysis · p. 14

E. Further Anaysis

The analysis finds that human-written and LLM-discovered reasoning structures share stepwise patterns on BBH navigation. It also presents MATH examples distinguishing errors in reasoning structures from errors in intermediate calculations.

Model-Discovered Reasoning Structures vs. Human Reasoning Patterns

  • Human and LLM reasoning structures share step-wise analysis of instructions on BBH navigation, including mental-noting after each movement.Humans were given three unlabeled task instances and an example structure, then asked to write a structure before solving; Figure 11 compares their structures with LLM-composed ones.Source: E. Further Anaysis · pp. 14, 17, Figure 11 · p. 17
  • Table 4 shows MATH examples of wrong reasoning structures, with the first error highlighted.Source: Table 4 · p. 15
  • Table 5 shows MATH examples of wrong calculations, with the first error in intermediate computations highlighted.Source: Table 5 · p. 16

Related papers

Loading 2402.03620v1…