Diagnosing Knowledge Conflict in Multimodal Long-Chain Reasoning

2026-02-16

AI Summary

Multimodal large language models often fail in long-chain reasoning due to signal conflicts arising from heterogeneous knowledge sources. This work systematically investigates the encoding mechanisms and processing stages of such knowledge conflicts through internal representation probing, distinguishing between objective conflicts at the input layer and effective conflicts emerging during reasoning. The authors propose a trajectory-aggregation-based diagnostic method and find that conflicting information is explicitly encoded in linearly separable features, with conflict signals predominantly localized in the middle-to-late layers of the network. Trajectory aggregation robustly identifies conflict types, and the model exhibits directionally asymmetric preferences toward conflicting sources. These findings offer mechanistic insights and principled tools for understanding, diagnosing, and mitigating failures in multimodal reasoning.

Technology Category

• Knowledge Representation and Reasoning: Diagnosis and Abductive Reasoning
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

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.

Problem

• knowledge conflict
• multimodal reasoning
• long chain-of-thought
• reasoning failure
• conflicting signals

Innovation

• knowledge conflict
• multimodal reasoning
• linear separability
• depth localization
• long chain-of-thought