Diagnosing Knowledge Conflict in Multimodal Long-Chain Reasoning
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866915799927619584 |
|---|---|
| author | Tang, Jing Wang, Kun Lu, Haolang Chen, Hongjin Chen, KaiTao Sun, Zhongxiang Li, Qiankun Lyu, Lingjuan Nan, Guoshun Zeng, Zhigang |
| author_facet | Tang, Jing Wang, Kun Lu, Haolang Chen, Hongjin Chen, KaiTao Sun, Zhongxiang Li, Qiankun Lyu, Lingjuan Nan, Guoshun Zeng, Zhigang |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2602_14518 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Diagnosing Knowledge Conflict in Multimodal Long-Chain Reasoning Tang, Jing Wang, Kun Lu, Haolang Chen, Hongjin Chen, KaiTao Sun, Zhongxiang Li, Qiankun Lyu, Lingjuan Nan, Guoshun Zeng, Zhigang Artificial Intelligence 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. |
| title | Diagnosing Knowledge Conflict in Multimodal Long-Chain Reasoning |
| topic | Artificial Intelligence |
| url | https://arxiv.org/abs/2602.14518 |