Multimodal long-chain-of-thought (CoT) generation requires models to dynamically route information between visual evidence and an expanding reasoning history. Existing hallucination mitigation methods often strengthen visual grounding or intervene in decoding, but typically treat grounding as a relatively static property of generation. We propose a training-free dynamic attention intervention framework that exploits the head-level sparsity, layer-level stratification, and trajectory-level dynamics of multimodal long-CoT decoding. Our method selects routing-sensitive heads using perturbation-only calibration, assigns visual-evidence and reasoning-history interventions to different layer regions, and continuously rescales selected head outputs according to their instantaneous token-group attention ratios. Across multiple MLLM architectures and multimodal hallucination benchmarks, our method consistently improves over vanilla decoding and visual contrastive decoding, achieving an average performance gain of 4.6%.

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 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. Code is available at [anonymous link](https://anonymous.4open.science/r/Diagnosing-Knowledge-Conflict-C78D/).

Long chain-of-thought (CoT) reasoning improves the performance of large language models, yet hallucinations in such settings often emerge subtly and propagate across reasoning steps. We suggest that hallucination in long CoT reasoning is better understood as an evolving latent state rather than a one-off erroneous event. Accordingly, we treat step-level hallucination judgments as local observations and introduce a cumulative prefix-level hallucination signal that tracks the global evolution of the reasoning state over the entire trajectory. Overall, our approach enables streaming hallucination detection in long CoT reasoning, providing real-time, interpretable evidence.

Large multimodal reasoning models solve challenging visual problems via explicit long-chain inference: they gather visual clues from images and decode clues into textual tokens. Yet this capability also increases hallucinations, where the model generates content that is not supported by the input image or the question. To understand this failure mode, we identify reasoning drift: during clue gathering, the model over-focuses on question-irrelevant entities, diluting focus on task-relevant cues and gradually decoupling the reasoning trace from visual grounding. As a consequence, many inference-time localization or intervention methods developed for non-reasoning models fail to pinpoint the true clues in reasoning settings. Motivated by these insights, we introduce ClueRecall, a metric for assessing visual clue retrieval, and present ClueTracer, a training-free, parameter-free, and architecture-agnostic plugin for hallucination suppression. ClueTracer starts from the question and traces how key clues propagate along the model’s reasoning pathway (question → outputs → visual tokens), thereby localizing task-relevant patches while suppressing spurious attention to irrelevant regions. Remarkably, without any additional training, ClueTracer improves all reasoning architectures (including R1-OneVision, Ocean-R1, MM-Eureka, .) by 1.21 on reasoning benchmarks.