Unified multimodal models use a shared Transformer backbone for visual understanding and image generation. Existing post-training signals primarily constrain final outputs, output-side representations, or task-level behavior, providing limited supervision for how visual semantics are organized across intermediate layers. We propose HiSVR, a hierarchical semantic visual reconstruction method for post-training unified multimodal models. A frozen semantic visual tokenizer extracts residual semantic codes from each training image, which supervise image-associated hidden states at different Transformer depths. Shallow layers predict shorter semantic-code prefixes, while deeper layers reconstruct longer prefixes, yielding coarse-to-fine supervision over intermediate representations. To balance the differing optimization difficulty across depths and prefix lengths, we introduce an uncertainty-weighted progressive reconstruction loss. HiSVR is used only during training and removed at inference, preserving the original architecture and inference cost. Experiments on unified multimodal backbones show that HiSVR improves visual understanding while maintaining image generation quality. These results indicate that hierarchical semantic reconstruction effectively regularizes intermediate representations in unified multimodal models.

Recent multimodal large language models (MLLMs) achieve strong performance on visual reasoning benchmarks, yet it remains unclear to what extent such performance reflects reasoning directly grounded in visual evidence. We introduce VisReason, a benchmark for vision-centric reasoning in everyday scenarios where perception and inference are tightly coupled. VisReason contains 1,505 questions across 10 categories spanning perceptual, structural, and conceptual reasoning. Our evaluation shows that VisReason poses a qualitatively different challenge from existing benchmarks, exposing substantial gaps between humans and current MLLMs and revealing limited benefits from test-time reasoning strategies. VisReason offers a focused diagnostic for evaluating vision-centric reasoning beyond language.

Prefilling computational costs pose a significant bottleneck for Large Language Models (LLMs) and Large Multimodal Models (LMMs) in long-context settings. While token pruning reduces sequence length, prior methods rely on heuristics that break compatibility with hardware-efficient kernels like FlashAttention. In this work, we observe that tokens evolve toward semantic fixing points, making further processing redundant. To this end, we introduce Delta Attention Selective Halting (DASH), a training-free policy that monitors the layer-wise update dynamics of the self-attention mechanism to selectively halt stabilized tokens. Extensive evaluation confirms that DASH generalizes across language and vision benchmarks, delivering significant prefill speedups while preserving model accuracy and hardware efficiency. Code will be released at https://github.com/verach3n/DASH.git .