PaperlayerView exact revision on arXiv

Source-linked AI summary

Method-Based Reasoning for Large Language Models: Extraction, Reuse, and Continuous Improvement

Hong Su

arXiv:2508.04289v2cs.AIcs.CLcs.LG

TL;DR

LLMs can struggle with novel problems and logical verification because their behavior is driven largely by statistical patterns. The paper adds an external procedural memory that extracts problem-solution methods from content and interactions, retrieves them for new queries, and ranks them through feedback. Experiments report higher similarity for learned and refined methods, with newly learned methods outperforming earlier ones.

  • Problem

    LLMs may fail on novel problems and leave hallucinations or inconsistent assumptions logically unverified.

    Source: §I · p. 1

  • Method

    The model extracts problem-solution procedures from training content, responses, and user inputs, stores them externally, retrieves them for new queries, and ranks them through feedback.

    Source: §I · p. 1, §V · p. 9

  • Results

    0.8426 average similarity for method2 exceeded 0.4587 for method1, while method1 reached 0.7791 versus 0.4693 for NoMethod.

    Source: C. Learning Methods from Content · p. 8, D. Improvement of Methods Through… · p. 9

  • Takeaways & Limitations

    Newly learned methods can generalize better and improve logical reliability and output alignment through continual refinement.

    Source: §V · p. 9

Abstract

from arXiv · show

Large language models (LLMs) have shown impressive capabilities across a wide range of language tasks. However, their reasoning process is primarily guided by statistical patterns in training data, which limits their ability to handle novel problems and perform consistent logical reasoning. In this paper, we propose a method-based model that enhances LLMs with explicit, reusable procedures extracted from training content, generated responses, and user interactions. Each method is represented as a pair consisting of a problem and its corresponding solution, stored externally and ranked based on feedback. When a new query is received, the system retrieves and applies the most relevant methods to guide the LLM's response. Our model enables continual learning, method reuse, and logical consistency beyond next-token prediction. Experimental results demonstrate that the system improves factual verification and generalization in complex prompts, and that newly learned methods can outperform earlier ones through user-driven refinement.

I. INTRODUCTION

The paper addresses LLMs’ difficulty with novel problems and explicit logical verification by introducing reusable, automatically extracted reasoning methods. It positions these methods as procedural memory that can be retrieved, ranked, and refined through feedback rather than relying only on prompts or output-level adaptation.

I. INTRODUCTION

  • LLMs rely on statistical patterns and often fail to handle out-of-distribution problems or detect hallucinated facts and inconsistent assumptions.Source: §I · p. 1
  • The proposed model represents reusable methods as problem-solution pairs extracted from training materials, generated outputs, and user inputs.Source: §I · p. 1
  • Stored methods are retrieved for new queries, executed by the LLM, and ranked through user or system feedback so higher-ranked procedures are favored over time.Source: §I · p. 1

II. RELATED WORK

  • Unlike CoT and ReAct, which depend on carefully crafted task-specific prompts, the approach automatically extracts generalized reasoning patterns and reuses them across sessions.Source: §I · p. 1, §II · p. 2
  • Unlike RAG’s retrieval of factual documents, the system retrieves abstract procedures that target logical reliability and reusable solution paths.Source: §II · p. 2
  • The method-management module stores and refines procedural knowledge through ranking, replacement, and reuse, combining external memory with user-controlled continual improvement.Source: §II · pp. 2–3

III. METHOD-BASED REASONING FOR LARGE LANGUAGE MODELS

The method-based model treats training material as a repository of reusable procedures rather than only text. It formalizes flexible problem-solution mappings and uses semantic abstraction to retrieve, transfer, and optimize methods.

  • A method is a semantic procedure of one or more actions that solves a specific problem and can transfer to similar problems despite different wording.Source: §III · p. 3
  • Method selection and optimization use logical reasoning or preference mechanisms such as RLHF instead of conventional token-level prediction.Source: §III · p. 3
  • The system extracts actionable procedures from input content as textual prompts or hints, and potentially as executable scripts or API calls.Source: §III · p. 3
  • Explicit method storage provides a more structured and reusable alternative to retaining all training content, which can create bloated memory and obscure relevance.Source: §III · p. 3
  • The model decouples abstract problems and solutions through a many-to-many mapping, allowing multiple solutions per problem and supporting composition and reuse.Source: §III · p. 3
  • A hierarchical tree organizes problem representations and their associated solutions to support semantic retrieval across large method collections.Source: §III · p. 4

A. Storage of Methods

The method storage tree organizes methods by problem similarity, with each node linking a problem to one or more solution variants. It supports general-to-specific retrieval and can be scoped to users or the entire LLM.

1) Method Storage Tree:

  • The method storage tree groups methods according to their associated problems, using problem descriptions to organize similar cases.Source: 1) Method Storage Tree · p. 4
  • Each tree node stores a problem with a list of associated solutions, allowing multiple valid method variants for one problem.Source: 1) Method Storage Tree · p. 4
  • When rankings are unavailable, the LLM can choose among methods along the path from a general root node to a specific leaf.Source: 1) Method Storage Tree · p. 4
  • Storage may be maintained per user for personalized reuse or globally at the LLM level to aggregate methods from training data and interactions.Source: 1) Method Storage Tree · p. 4

2) Methods for Improving Other Methods:

The model supports hierarchical reasoning in which higher-level methods validate or improve lower-level methods, while solutions can also be decomposed into independently reusable parts.

2) Methods for Improving Other Methods:

3) Solution Division and Part-Wise Storage:

  • Solutions can be divided into ordered parts, each assigned a functional role and stored, retrieved, or updated independently.This decomposition enables modular reuse and targeted refinement.Source: 3) Solution Division and Part-Wise… · p. 4
  • Ranking selects updated solution parts at each position, allowing underperforming components to be replaced while effective components are retained.Common parts can also be reused across tasks with overlapping subtasks.Source: 3) Solution Division and Part-Wise… · pp. 4–5

B. Always-On Extraction: Automatic Method Mining

Always-on extraction mines methods from training data, generated outputs, and external inputs, then filters and stores candidates for continual refinement.

C. Measurement of Methods

Method measurement combines semantic relevance with estimated solution effectiveness, using external human rankings and internal LLM selection to maintain a curated repository.

  • External ranking uses human evaluation of candidate solutions, while internal ranking lets the LLM select an appropriate method or offer alternatives.Source: C. Measurement of Methods · p. 5
  • The interaction pipeline first filters candidates externally, then performs internal selection and may present alternatives for user comparison.Source: C. Measurement of Methods · p. 5
  • Dual ranking balances adaptability from human feedback with inference efficiency while maintaining a curated method repository.Source: C. Measurement of Methods · p. 6
  • Method utility combines relevance to the current problem with the estimated effectiveness of the method’s solution.The model selects methods by maximizing this logical utility.Source: C. Measurement of Methods · p. 6
  • Comparing meanings and problem-solving intents rather than surface wording supports transfer and reuse across structurally similar tasks.Source: C. Measurement of Methods · p. 6

D. Brief Proof: Why Methods Generalize to New Problems

Methods generalize by separating abstract problem structure from reusable solution procedures, allowing the same logic to transfer across differently phrased but structurally similar problems.

E. Procedure for Method Extraction and Application

The method extraction module analyzes inputs, retrieves and ranks applicable methods, applies selected candidates through the LLM, and stores validated associations to improve future responses.

F. Method Formats

Methods are modular, executable units that may be implemented as external programs or internal LLM reasoning paths, with hybrid coordination supporting procedural control and natural-language understanding.

  • Extracted methods may be step-by-step procedures, conditional workflows, or validations executed internally or through supporting systems.Source: F. Method Formats · p. 7
  • External methods map method steps to callable functions or APIs for tasks such as data processing, retrieval, visualization, or system I/O.Source: F. Method Formats · p. 7
  • Internal methods use structured prompts or step-wise reasoning paths to generate intermediate results, verify constraints, or simulate planning.Source: F. Method Formats · p. 7
  • Metadata such as descriptors, execution links, or API references coordinates external method invocation within the LLM interaction.Source: F. Method Formats · p. 7
  • Whether implemented as code or structured prompts, methods are reusable and interpretable building blocks beyond token-level prediction.Source: F. Method Formats · p. 7

IV. VERIFICATION

The verification study tests whether method extraction and reuse improve factual checking in complex prompts, using MethodManager to retrieve, rank, and apply learned procedures.

A. Environment

  • MethodManager stores extracted methods, filters inputs, retrieves candidates, appends them to prompts, and supports user ranking of generated outputs.Source: A. Environment · p. 7
  • The verification setup uses GPT-4o1 and is described as model-agnostic, focusing primarily on MethodManager logic.Source: A. Environment · p. 7
  • Methods are stored globally across users and chat sessions after filtering content for qualification as a method.Source: A. Environment · p. 7

B. Verification Objective

  • The system targets hallucinations in procedurally rich prompts by enforcing software-existence checks before providing usage instructions.Source: B. Verification Objective · p. 8

C. Learning Methods from Content

D. Improvement of Methods Through Continuous Learning

Across repeated interaction sessions, the system learned a generalized software-existence-checking method and selected it over an earlier method, producing higher alignment with the intended behavior. The results support continual refinement from narrow procedures toward more general methods.

ics1

  • In ics1, a complex project-duplication prompt was processed using method1 as the baseline, reducing the influence of its software-specific instruction.The prompt concerned verifying parameter impact in the target HongHanKey software.Source: D. Improvement of Methods Through… · p. 9

ics2

  • In ics2, the system received a generalized instruction to check whether target software exists and stored it as method2.The generalized method was intended to apply beyond the specific SuHongKey case.Source: D. Improvement of Methods Through… · p. 9

ics3

  • In ics3, with similar method rankings, the MethodManager delegated selection between method1 and method2 to the LLM before processing the complex prompt.The selected method was passed together with the user prompt.Source: D. Improvement of Methods Through… · p. 9

Experimental setup

Results

Continuous learning

  • Continual interaction lets the MethodManager store, compare, and prefer better or more general strategies, moving from narrow procedures toward abstract methods.The paper links this refinement with improved correctness and generalization.Source: D. Improvement of Methods Through… · p. 9

V. CONCLUSION

The paper concludes that externally stored problem-solution methods can provide reusable procedural memory for LLM reasoning and improve logical validation. Its future work focuses on composing methods hierarchically and integrating them with retrieval- and tool-augmented systems.

Conclusion

  • The method-based architecture extracts, stores, and reuses structured problem-solving procedures as procedural memory across sessions and queries.Methods can be identified from user inputs, model outputs, or training content.Source: §V · p. 9
  • Abstract problem-solution representations support generalization to new tasks with similar logical structures despite different surface forms.User feedback ranks and refines methods over time.Source: §V · p. 9
  • The experiments show improved behavior when logical validation must precede generation, including checking software existence before giving instructions.Newly learned methods also generalize better and yield higher output alignment than earlier stored methods.Source: §V · p. 9

Future work

  • Future work will investigate hierarchical decomposition and multi-method aggregation, alongside integration with retrieval-augmented and tool-augmented systems.The paper identifies software configuration, legal reasoning, and intelligent tutoring as target domains.Source: §V · p. 9
Loading 2508.04289v2…