A Validation-Based Framework for Reasoning Reliability in Large Language Models Without Model Retraining

Sammanfattning

Introduction: Large Language Models (LLMs) generate fluent responses but often produce reasoning errors due to unsupported inference. Existing approaches mainly improve answer correctness but do not ensure that reasoning is supported by available information. This study focuses on improving reasoning reliability through explicit validation. Research Question This study investigates: “Can the reasoning reliability of LLMs be improved, without retraining the model, by introducing external validation mechanisms that regulate the reasoning process?” and “How can such mechanisms be designed and integrated into the reasoning process to ensure that generated reasoning is supported by available information?” Method: A quantitative experimental design evaluates a validation-based reasoning control framework. The framework consists of structured representation, reasoning generation, and validation-based control. Input and reasoning are represented as graphs, and unsupported reasoning triggers revision or refusal. Experiments use the NeuLR and CLUTRR datasets with GPT-5, evaluated by correctness, error, and refusal rates. Results: The framework reduces the answer error rate across tasks. Correctness improves for inductive and abductive reasoning but decreases for deductive reasoning and CLUTRR. The refusal rate increases, indicating that unsupported answers are filtered. Discussion: External validation improves reasoning reliability by reducing unsupported answers without retraining. However, it does not consistently improve accuracy and is more effective for structurally regular tasks. Performance is limited for tasks requiring implicit reasoning, and the multi-stage pipeline introduces significant computational cost.