Multimodal large language models excel at vision-language tasks but suffer from hallucinations that undermine their reliability. Existing mitigation methods apply uniform heuristics at the output or attention level, overlooking the heterogeneous causal roles of attention heads and their dynamic behaviors during inference. We revisit hallucination from a causal perspective and identify it as competition between two pathways: a direct path from visual evidence to outputs, and an indirect path dominated by knowledge priors. When visual evidence weakens, MLLMs increasingly rely on parametric knowledge priors—such as spurious correlations learned during pretraining. This over-reliance causes attention heads responsible for visual grounding to drift toward language-driven patterns, resulting in generations consistent with the language model but unfaithful to the image. We propose Causal Head Gating, an inference-time framework that separates visual evidence from knowledge priors by modeling their causal contributions at the attention-head level. Under this framework, we introduce CHERI (Causal Head Evidence Reweighting and Intervention), which identifies attention heads whose visual contributions weaken under hallucination and dynamically restores their influence. Specifically, CHERI reinforces heads that propagate visual evidence while suppressing those dominated by knowledge priors, strengthening the direct causal effect of vision on generation and attenuating spurious indirect effects. Experiments across multiple benchmarks show that CHERI consistently reduces hallucination rates across diverse task formats while preserving overall generation quality and task performance.

Multimodal large language models excel at vision-language tasks but remain prone to hallucinations that undermine their reliability. Existing approaches predominantly treat hallucinations as classification errors, overlooking the heterogeneous behaviors of attention heads and their dynamic influences during inference. We revisit MLLM reasoning from an energy perspective and identify that hallucinations stem from imbalances between visual potential and language prior potential: when visual information is ambiguous or language priors dominate, attention heads tend to be driven by linguistic statistical patterns, generating content inconsistent with visual evidence. We propose AHEAD, a framework that quantifies the energetic properties of each attention head during object generation through two potential networks—the Visual Grounding Potential Network and the Language Prior Potential Network—and dynamically adjusts their contributions at inference time. Specifically, we amplify attention heads with strong visual grounding capacity while suppressing those overly reliant on language priors. Experiments across multiple benchmarks demonstrate that AHEAD significantly reduces hallucination rates without fine-tuning the base MLLM while maintaining generation quality.

Visualization-of-thought (VoT) has substantially advanced multimodal spatial reasoning by interleaving visual thinking with verbal reasoning. However, multi-hop VoT reasoning introduces a fundamental scalability bottleneck: the autoregressive accumulation of interleaved visual and textual tokens causes KV cache memory to grow rapidly across successive hops. We present ScoutKV, a framework that recasts KV cache eviction as a reinforcement learning problem. ScoutKV trains a dedicated lightweight policy agent per KV head, conditioned on key–value representations, cross-hop positional encodings, and modality embeddings, to predict each token’s long-term utility across the full reasoning trajectory. Training combines offline supervised fine-tuning with future-query attention targets and online GRPO optimization with a multimodal process reward targeting low-entropy text anchors and modality-specific spatial anchors. ScoutKV consistently reduces FLOPs, latency, and peak memory while maintaining or improving accuracy across multimodal spatial reasoning benchmarks, and generalizes robustly to broader vision-language tasks.