LEMO: Resolving Logical Inertia in Large Language Models via Structured Cognitive Priors

LEMO (Logic Evaluation with Multi-modal Optimization) addresses the logical inertia issue in large language models by proposing a conflict-aware fusion method. Using techniques like synthetic logical reasoning datasets, a two-stage training strategy (basic logic learning + advanced reasoning strategies), and LoRA parameter-efficient fine-tuning, it systematically studies the robustness of models in logical reasoning and reveals their behavioral patterns when facing rule perturbations. The project constructs a reproducible dataset generation framework, multi-stage training process, and comprehensive evaluation system, aiming to mitigate logical inertia and enhance the model's sensitivity to logical conflicts.

Large language models have achieved remarkable results in natural language processing, but they face the problem of logical inertia when dealing with strict logical reasoning tasks: once a model learns a certain reasoning pattern, it is difficult to adjust flexibly even when faced with contradictory new information. For example, after learning "if A then B", the model still reasons according to the original pattern when encountering "if A then not B". This inertia is particularly fatal in scenarios requiring strict formal reasoning, such as mathematical proof, legal reasoning, and medical diagnosis, where users expect models to identify contradictions and adjust reasoning like human experts.

Synthetic Data Generation

Based on the structure of logical syllogisms, generate various variant data (basic reasoning chains, redundant/critical rule removal, contradictory fact injection, logical equivalence transformation) to precisely control variables and locate model weaknesses.

Two-Stage Training

• Stage 1: Use data from critical rule removal and contradictory fact injection for supervised fine-tuning, combined with generative training (predicting missing rules) to establish basic logical understanding.
• Stage 2: Adopt strategies like mixed generative training, DPO preference optimization, Chain of Thought (CoT), fusion training, and RA-CoT to enhance complex reasoning capabilities.

LoRA Parameter-Efficient Fine-Tuning

Freeze most parameters of the pre-trained model, only train a small number of low-rank matrix parameters to improve efficiency, prevent overfitting, and be compatible with multiple models (BERT, Qwen2, LLaMA, etc.).

Baseline Model Comparison

After the first stage of training, the accuracy on standard reasoning tasks is close to 1.0, but performance varies when facing rule perturbations: the accuracy on critical rule removal tests is only 0.25-0.3 (a manifestation of logical inertia), and the accuracy on contradictory fact injection tests is 0 (lack of contradiction detection ability).

Effect of Advanced Training Strategies

• DPO strategy achieves an accuracy of 1.0 on contradictory fact tests, but the accuracy on the baseline test set drops to 0;
• Mixed generative training shows balanced performance (Variant 2: 0.405, Variant 3: 0.973);
• Chain of Thought training improves performance on logical equivalence transformation tasks.

Logical Equivalence Law Tests

The baseline model's accuracy on single logical transformation is close to 1.0, but performance decreases when multiple laws are combined (e.g., Qwen2 is 0.645).

The conflict-aware fusion method handles logical conflicts through explicit mechanisms, including:

• Structured Cognitive Priors: Instill knowledge of logical consistency (contradiction identification, rule completeness check, etc.);
• Dynamic Attention Mechanism: Allocate more attention when potential conflicts are detected;
• Multi-Path Reasoning: Explore multiple reasoning paths and evaluate their rationality;
• Uncertainty Quantification: Assess the uncertainty of reasoning results to avoid forcing incorrect answers. Experiments show that this method significantly improves the model's ability to handle rule perturbations and conflicts.

Limitations

• There is a gap between synthetic data and real-world logical reasoning;
• Experiments are limited to smaller models (e.g., BERT-base, Qwen2-1.5B);
• Generalization ability on real NLP tasks (such as LogicNLI, MNLI) needs to be improved;
• The interpretability of the model's internal decision-making process is insufficient.

Future Directions

• Explore the transfer from synthetic data to real scenarios;
• Validate results on larger-scale models;
• Improve the model's generalization ability on real NLP tasks;
• Enhance the interpretability of the reasoning process.

Through systematic experiments and innovative methods, the LEMO project reveals the strengths and limitations of large language models in logical reasoning, and conflict-aware fusion provides new ideas for solving logical inertia. This research has theoretical value and also provides guidance for practical applications such as automated theorem proving, legal contract analysis, and medical diagnosis assistance. Ensuring AI systems have robust logical reasoning and contradiction detection capabilities is a core challenge in building trustworthy AI, and LEMO has laid a solid foundation for this.