Abstract
Logical reasoning is a fundamental capability of large language models. However, existing studies often overlook the interaction between logical complexity and semantic complexity, leading to systems that struggle with abstract propositions, ambiguous contexts, and conflicting stances that are central to human reasoning. We propose LogicAgent, a semiotic-square-guided framework that jointly addresses these two axes of difficulty. The semiotic square provides a principled structure for multi-perspective semantic analysis, and LogicAgent integrates automated deduction with reflective verification to manage logical complexity across deeper reasoning chains. To support evaluation under these conditions, we introduce RepublicQA, a benchmark that couples semantic complexity with logical depth. RepublicQA reaches college-level semantic difficulty (FKGL 11.94), contains philosophically grounded abstract propositions with systematically constructed contrary and contradictory forms, and offers a semantically rich setting for assessing logical reasoning in large language models. Experiments show that LogicAgent achieves state-of-the-art performance on RepublicQA with a 6.25 percent average improvement over strong baselines, and generalizes effectively to mainstream logical reasoning benchmarks including ProntoQA, ProofWriter, FOLIO, and ProverQA, achieving an additional 7.05 percent average gain.
Introduction
While Large Language Models (LLMs) have demonstrated significant progress in logical tasks, they frequently struggle with the inherent "messiness" of human language. This difficulty arises from two distinct axes: logical complexity (the depth of the reasoning chain) and semantic complexity (abstract concepts, ambiguous contexts, and conflicting stances). Most existing benchmarks rely on simple, template-based predicates that fail to capture the nuance of real-world reasoning found in fields like law or philosophy. LogicAgent addresses this gap by grounding reasoning in structuralist semiotics, allowing for a more robust navigation of abstract propositions.
Methodology: The LogicAgent Framework
LogicAgent utilizes a three-stage pipeline grounded in Greimas’ Semiotic Square, a tool that maps logical oppositions between concepts to ensure a comprehensive evaluation of any given proposition.
-
Semantic Structuring Stage: Instead of evaluating a single proposition ($S_1$) in isolation, the agent generates its contradictory ($ eg S_1$), its contrary ($S_2$, e.g., "just" vs. "unjust"), and the contradictory of the contrary ($ eg S_2$). This four-point map ensures the agent considers the full semantic space.
-
Logical Reasoning Stage: The agent translates these four perspectives into First-Order Logic (FOL). A "Planner" outlines a reasoning blueprint, which a "Solver" then executes via symbolic deduction.
- Reflective Verification Stage: If different perspectives yield conflicting results, the agent enters a "Deep Reflection" mode. It uses the structural implications of the semiotic square to adjudicate which path is logically sound, emulating human-like self-correction.
RepublicQA: Bridging Semantics and Logic
To test the intersection of semantic and logical complexity, we introduce RepublicQA. Derived from Plato’s Republic, this benchmark features abstract philosophical propositions regarding justice and morality. It reaches a college-level reading difficulty (FKGL 11.94) and includes 600 samples across 61 unique topics. Our analysis shows that semantic complexity acts as a "force multiplier" for logical difficulty; as reasoning depth increases, the accuracy of standard LLMs drops significantly faster on RepublicQA than on simpler, template-based datasets.
Experimental Results
LogicAgent achieves state-of-the-art performance across multiple benchmarks. On RepublicQA, it outperformed strong baselines, including GPT-4o with Chain-of-Thought (CoT), by an average of 6.25%. Furthermore, the framework demonstrated robust generalization, achieving a 7.05% average gain on mainstream benchmarks such as FOLIO and ProofWriter.
Key insights from our experiments include the importance of the contrary perspective. By forcing the LLM to consider opposites (e.g., proving something is "unjust" to validate it is "just"), the system catches errors that single-path reasoning often misses. Additionally, ablation studies confirmed that a dual-modality approach—combining natural language for contextual grounding and FOL for inferential precision—is essential for peak performance.
Conclusion and Future Work
LogicAgent demonstrates that semiotic-grounded multi-perspective reasoning significantly enhances logical performance in LLMs. However, challenges remain, such as the "reasoning complexity threshold" where overly long trajectories can impair performance, and the computational overhead of running parallel reasoning paths. Future work will focus on selectively activating high-cost components only when high ambiguity is detected and further refining the reflective verification stage to eliminate rare reasoning hallucinations.
Summary updated