Introduction
Large Language Models (LLMs) have demonstrated impressive capabilities in natural language processing, yet they frequently falter when faced with complex logical deduction tasks. Traditional approaches to improving LLM reasoning often impose rigid structures, such as Chain-of-Thought (CoT) or Tree-of-Thought (ToT), which may not align with the fluid nature of human problem-solving. This paper introduces DetermLR, a novel framework designed to enhance logical reasoning by treating deduction as a dynamic transition from "indeterminacy" to "determinacy."
The DetermLR Framework
The core philosophy of DetermLR is that logical reasoning is essentially the process of transforming complex, conditional rules (indeterminate premises) into simple, factual insights (determinate premises). The framework operates through an iterative process consisting of four main components:
- Premise Identification: The system first classifies input information into determinate premises (facts) and indeterminate premises (rules or hypotheses).
- Premise Prioritization and Exploration: To avoid the combinatorial explosion of possible deductions, DetermLR employs a two-stage scoring system. A relevance scorer identifies premises most pertinent to the target conclusion, while a supplement scorer evaluates how effectively different premises can be combined to derive new information.
- Three-Fold Verification: Every newly generated proposition undergoes a rigorous check. It must be logically valid, contribute toward the final goal, and be non-redundant relative to existing knowledge.
- Reasoning Memorization: A dynamic memory module stores both successful derivations and failed attempts. This "reasoning memory" prevents the model from repeating errors and allows it to backtrack or pivot based on past experiences.
Experimental Evaluation
DetermLR was rigorously tested against five major logical reasoning benchmarks: LogiQA, ProofWriter, FOLIO, PrOntoQA, and LogicalDeduction. These datasets cover a wide range of logical challenges, from civil service exam questions to synthetic first-order logic puzzles. Additionally, the framework's generalizability was tested on multi-variable mathematical equations, treating variable elimination as a form of logical reduction.
Key Results
The experimental results demonstrate that DetermLR significantly outperforms existing state-of-the-art methods. On the LogiQA dataset, using GPT-4, DetermLR achieved an accuracy of 54.19%, a substantial improvement over the 45.25% achieved by Cumulative Reasoning (CR).
Beyond accuracy, DetermLR exhibits remarkable efficiency. By prioritizing relevant premises and utilizing its memory module, the framework reaches conclusions in fewer steps. For instance, on LogiQA, DetermLR averaged 11.74 steps per problem, whereas Tree-of-Thoughts required nearly 20.
As shown in the performance metrics, there is a clear positive correlation between the number of generated determinate premises and the model's final accuracy. This validates the core hypothesis of the framework: reducing indeterminacy is the key to successful logical deduction.
Conclusion
DetermLR represents a shift away from static reasoning templates toward a more adaptive, memory-augmented approach. By explicitly managing the transition from indeterminate rules to determinate facts, the framework provides LLMs with a clearer path through complex logical landscapes.
Despite its success, challenges remain in handling implicit constraints and deeply nested logical structures. Future work will focus on improving the model's ability to infer "unspoken" conditions and manage increasingly verbose real-world descriptions. Ultimately, DetermLR provides a robust foundation for more human-like, efficient, and accurate automated reasoning.
Summary updated