提出新方法让大模型在逻辑冲突下不僵化，显著提升推理鲁棒性。
Conflict-Aware Fusion: Mitigating Logic Inertia in Large Language Models via Structured Cognitive Priors
- 通过四阶段训练强制模型先验证后推理，构建结构化认知先验
- 在矛盾注入测试中准确率从0.00提升至71.7%，8B模型全测试通过
- 适合需要高可靠性逻辑推理的场景，如法律、医疗、形式化验证
大语言模型在多数推理基准上表现优异，但在规则系统结构扰动下仍显脆弱。我们设计了四项压力测试——冗余规则删除、必要规则删除、矛盾规则注入、多规则叠加——揭示出‘逻辑惯性’现象：生成式LLM（Qwen2/3、TinyLlama、GPT-4o、Gemma-3-4B-IT）和编码器单向的BERT基线，在前提不一致时仍沿既定演绎路径持续输出，导致准确率骤降：未经处理的基线在矛盾注入测试中从1.00降至0.00（实例级精确匹配），GPT-4o仅解决56.0%的矛盾案例。为此提出冲突感知融合（Conflict-Aware Fusion），四阶段训练流程：(i) SFT建立验证前缀；(ii) DPO强化矛盾停顿决策边界；(iii) 逻辑不变性正则化（LIRE）通过对称KL惩罚逻辑等价表述间的偏差；(iv) 基于验证反馈的强化学习（RLVF）使用符号前向链引擎作为确定性奖励源，联合优化不变性与敏感性。该流程使1.5B与8B骨干模型均通过全部四项核心压力测试。进一步验证了第二阶段扩展：以Lean 4内核替代命题型奖励源，在分层抽样的187题中，对105个经典可推导问题（T类）达成99.0%内核一致性（整体跨极性71.7%），为正式验证的强化学习训练提供可靠升级路径。代码与基准：https://github.com/14H034160212/lemo
原文摘要 · Abstract (English)
Large language models (LLMs) achieve high accuracy on many reasoning benchmarks but remain brittle under structural perturbations of rule-based systems. We introduce a diagnostic framework with four stress tests -- redundant vs. essential rule deletion, contradictory-rule injection, logic-preserving rewrites, and multi-law stacking -- and use it to expose Logic Inertia: the tendency of generative LLMs (Qwen2/3, TinyLlama, GPT-4o, Gemma-3-4B-IT) and the encoder-only BERT baseline to persist along learned deductive trajectories under inconsistent premises. The collapse is sharp: untreated baselines fall from accuracy 1.00 on the base task to 0.00 on contradiction injection (instance-level exact match), and GPT-4o resolves only 56.0% of contradiction cases. We propose Conflict-Aware Fusion, a four-stage training pipeline that enforces verification-before-deduction as a learned structural prior: (i) SFT establishes the verification preamble; (ii) DPO sharpens the halt-on-contradiction decision boundary; (iii) Logical Invariance REgularisation (LIRE) penalises divergence between logically equivalent rule formulations via symmetric KL; (iv) Reinforcement Learning from Verification Feedback (RLVF) uses a symbolic forward-chaining engine as a deterministic oracle reward, jointly optimising invariance and sensitivity. The pipeline saturates all four primary stress tests for both 1.5B and 8B backbones. We further validate a Phase 2 extension that replaces the propositional oracle with a Lean 4 kernel, attaining 99.0% kernel agreement on the 105 classically-derivable (T) questions within a stratified 187-question Lean-translated sample (overall 71.7% across both polarities), providing a sound upgrade path to formally verified RL training. Code and benchmark: https://github.com/14H034160212/lemo
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。