Conflict-Aware Fusion: Mitigating Logic Inertia in LLMs
A novel diagnostic framework has uncovered Logic Inertia in large language models (LLMs), where these models continue along established deductive paths despite facing inconsistent premises. This framework comprises four stress tests: essential vs. redundant rule removal, injection of contradictory rules, logic-preserving rewrites, and stacking multiple laws. Evaluated on generative LLMs (Qwen2/3, TinyLlama, GPT-4o, Gemma-3-4B-IT) alongside the BERT encoder-only baseline, accuracy plummets from 1.00 on the base task to 0.00 when contradiction injection is applied (instance-level exact match). GPT-4o addresses only 56.0% of contradiction scenarios. The suggested remedy, Conflict-Aware Fusion, is a four-phase training pipeline that prioritizes verification before deduction, beginning with SFT to create the verification preamble.
Key facts
- Logic Inertia is the tendency of LLMs to persist along learned deductive trajectories under inconsistent premises.
- Diagnostic framework includes four stress tests: redundant vs. essential rule deletion, contradictory-rule injection, logic-preserving rewrites, and multi-law stacking.
- Tested models: Qwen2/3, TinyLlama, GPT-4o, Gemma-3-4B-IT, and BERT baseline.
- Accuracy drops from 1.00 to 0.00 on contradiction injection (instance-level exact match).
- GPT-4o resolves only 56.0% of contradiction cases.
- Conflict-Aware Fusion is a four-stage training pipeline enforcing verification-before-deduction.
- First stage: SFT establishes the verification preamble.
- The framework is designed for rule-based systems.
Entities
—