{"paper":{"title":"Conflict-Aware Fusion: Mitigating Logic Inertia in Large Language Models via Structured Cognitive Priors","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A four-stage training pipeline makes LLMs check rules for contradictions before reasoning, fixing their tendency to follow inconsistent premises.","cross_cats":["cs.CL","cs.LG","cs.LO"],"primary_cat":"cs.AI","authors_text":"Michael Witbrock, Qiming Bao, Xiaoxuan Fu","submitted_at":"2025-12-06T10:49:50Z","abstract_excerpt":"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 baseli"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"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).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the four stress tests capture the relevant forms of logical inconsistency that arise in real-world LLM use, and that the symbolic forward-chaining engine (or Lean kernel) supplies rewards that do not embed their own unexamined biases or coverage gaps into the learned policy.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Conflict-Aware Fusion mitigates Logic Inertia in LLMs through a four-stage pipeline of SFT, DPO, logical invariance regularization, and reinforcement learning from a symbolic oracle, saturating four stress tests on rule contradictions.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A four-stage training pipeline makes LLMs check rules for contradictions before reasoning, fixing their tendency to follow inconsistent premises.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"3612cf4edbef4804ebd6e2f8040131b98825c0319488755a5287036837593de6"},"source":{"id":"2512.06393","kind":"arxiv","version":7},"verdict":{"id":"b5e5e1d1-0277-4db1-8f4b-96d3cc140796","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T01:05:09.434154Z","strongest_claim":"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).","one_line_summary":"Conflict-Aware Fusion mitigates Logic Inertia in LLMs through a four-stage pipeline of SFT, DPO, logical invariance regularization, and reinforcement learning from a symbolic oracle, saturating four stress tests on rule contradictions.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the four stress tests capture the relevant forms of logical inconsistency that arise in real-world LLM use, and that the symbolic forward-chaining engine (or Lean kernel) supplies rewards that do not embed their own unexamined biases or coverage gaps into the learned policy.","pith_extraction_headline":"A four-stage training pipeline makes LLMs check rules for contradictions before reasoning, fixing their tendency to follow inconsistent premises."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2512.06393/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"2e79b3fce7d1bc5803778319a7763cc2ee386c5580d081416245cd9eec7c3e83"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}