![[2602.14518] Diagnosing Knowledge Conflict in Multimodal Long-Chain Reasoning](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.14518] Diagnosing Knowledge Conflict in Multimodal Long-Chain Reasoning
Summary
This paper explores knowledge conflicts in multimodal large language models (MLLMs) during long chain-of-thought reasoning, proposing a framework for diagnosing and addressing these conflicts.
Why It Matters
Understanding knowledge conflicts in MLLMs is crucial for improving their reasoning capabilities. This research provides insights into how these models process conflicting information, which can enhance their reliability and performance in real-world applications.
Key Takeaways
- Knowledge conflicts can be categorized into input-level objective and process-level effective conflicts.
- Different types of conflicts are encoded as linearly separable features in the model's internal representations.
- Conflict signals are primarily located in the mid-to-late layers of the model, indicating a specific processing stage.
- Aggregating noisy token-level signals can effectively recover input-level conflict types.
- Reinforcing a model's implicit source preference under conflict is easier than enforcing the opposite.
Computer Science > Artificial Intelligence arXiv:2602.14518 (cs) [Submitted on 16 Feb 2026] Title:Diagnosing Knowledge Conflict in Multimodal Long-Chain Reasoning Authors:Jing Tang, Kun Wang, Haolang Lu, Hongjin Chen, KaiTao Chen, Zhongxiang Sun, Qiankun Li, Lingjuan Lyu, Guoshun Nan, Zhigang Zeng View a PDF of the paper titled Diagnosing Knowledge Conflict in Multimodal Long-Chain Reasoning, by Jing Tang and 9 other authors View PDF HTML (experimental) 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 ...
