Diagnosing Knowledge Conflict in Multimodal Long-Chain Reasoning

Multimodal large language models in long chain-of-thought reasoning often fail when different knowledge sources provide conflicting signals.We formalize these failures under a unified notion of knowledge conflict, distinguishing input-level objective conflict from processlevel effective conflict.Through probing internal representations, we reveal that: (I) Linear Separability: different conflict types are explicitly encoded as linearly separable features rather than entangled; (II) Depth Localization: conflict signals concentrate in midto-late layers, indicating a distinct processing stage for conflict encoding; (III) Hierarchical Consistency: aggregating noisy tokenlevel signals along trajectories robustly recovers input-level conflict types; and (IV) Directional Asymmetry: reinforcing the model's implicit source preference under conflict is far easier than enforcing the opposite source.Our findings provide a representation-level characterization of multimodal reasoning under knowledge conflict and enable principled diagnosis and control of long-CoT failures.Code: github.com/Diagnosing