Diagnosing Knowledge Conflict in Multimodal Long-Chain Reasoning

TL;DR

This paper investigates how multimodal large language models handle conflicting knowledge signals during long-chain reasoning, revealing their internal conflict encoding and proposing diagnostic insights.

Contribution

It formalizes knowledge conflicts in multimodal reasoning, uncovers their internal representations, and offers mechanisms for diagnosis and control of model failures.

Findings

Conflict types are linearly separable in internal representations

Conflict signals are concentrated in mid-to-late layers

Aggregating token signals recovers input-level conflict types

Abstract

Multimodal large language models (MLLMs) 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 process-level 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 mid-to-late layers, indicating a distinct processing stage for conflict encoding; (III) Hierarchical Consistency: aggregating noisy token-level 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 mechanism-level view of multimodal reasoning under knowledge conflict and enable principled diagnosis and control of long-CoT failures.