TAPS: Target-Aware Prefix Tree Selection for Diffusion-Drafted Speculative Decoding

Using a diffusion model for parallel drafting is a promising approach for speculative decoding. By predicting tokens at multiple future positions in a single forward pass, diffusion drafters substantially reduce drafting latency. However, this shifts the bottleneck to verification: verifying a single sequence limits acceptance length, while verifying large draft trees incurs excessive target-model latency. We identify a key mismatch in existing draft-tree methods: existing diffusion-tree methods rank nodes by the marginal probability, ignoring that verification is prefix-conditioned. As a result, they may verify unreachable descendants of rejected prefixes, increasing latency with limited acceptance gains. To address this, we propose TAPS, a target-aware prefix selection method that turns diffusion marginals into path-conditioned acceptance estimates. TAPS then selects a compact prefix-closed subtree under a fixed verification budget, improving the acceptance-cost tradeoff rather than simply expanding the draft tree. Experiments across diverse datasets and model families demonstrate that TAPS achieves up to 7.9x lossless end-to-end speedup over vanilla autoregressive decoding, outperforming state-of-the-art DFlash and DDTree by 1.36x and 1.74x respectively. Our work is available at https://anonymous.4open.science/r/TAPS-EMNLP2026-53DD

SimInsert: Seamless Video Object Insertion via Regional Sparse Attention Fusion

Video object insertion requires ensuring spatio-temporal coherence and interactive realism, extending far beyond simple content placement. However, current approaches are often hindered by a reliance on explicit motion engineering or resource-intensive retraining, restricting their flexibility and generalization. To bridge this gap, we present \textit{SimInsert}, a training-free paradigm that efficiently decouples the task into intuitive single-frame editing and semantic motion description. By harnessing the robust generative priors of image-to-video diffusion models, SimInsert propagates edits temporally, strictly preserving background invariance while enabling plausible, text-driven interactions between the inserted object and the dynamic environment. Our approach hinges on non-invasive guidance mechanisms that enforce structural consistency, facilitate seamless boundary fusion, and counteract the fidelity drift that typically accumulates during the denoising trajectory. Extensive quantitative experiments validate our efficacy: SimInsert surpasses state-of-the-art methods with an 18.8\% gain in PSNR, 20.1\% in SSIM, and a 44.1\% decrease in LPIPS, offering a streamlined solution for high-fidelity video editing.

DEFLECT: Delay-Robust Execution via Flow-matching Likelihood-Estimated Counterfactual Tuning for VLA Policies

Vision-Language-Action (VLA) policies are typically deployed with asynchronous inference: the robot executes a previously predicted action chunk while the model computes the next one. This creates a prediction-execution misalignment: the chunk is conditioned on the observation taken before inference began, but executes in a physical state that has already drifted forward by several control steps; naive asynchronous rollover collapses from 89% to under 1% on Kinetix as the inference cycle covers up to seven control steps. We introduce DEFLECT, a fully offline post-training refinement that applies as a near drop-in upgrade to existing async-VLA stacks by converting latency itself into a label-free preference signal: counterfactual fresh/stale action pairs are constructed from a frozen reference policy and scored under the deployment-time conditioning via an implicit flow-matching likelihood-ratio surrogate, with no human labels, reward models, or online rollouts. DEFLECT substantially extends the usable delay envelope of async VLA control, with +6.4 success-rate gain in the high-latency regime (5-7 control steps), +4.6 when transferred to a real-scale VLA at the longest delay, and consistent improvements on two real-robot tasks (a bimanual conveyor pick-and-place and a reactive whack-a-mole).

EMA: Efficient Model Adaptation for Learning-based Systems

Machine learning (ML) is increasingly applied to optimize system performance in tasks such as resource management and network simulation. Unlike traditional ML tasks (e.g., image classification), networked systems often operate in heterogeneous, long-running, and dynamic environment states, where input conditions (e.g., network loads) and operational objectives can shift over time and across settings. Existing learning-based systems offer little support for adaptation, resulting in costly model training, extensive data collection, degraded system performance, and slow responsiveness. This paper presents EMA, the first model adaptation system supporting learning-based systems to adapt to evolving environments with minimal operational overhead. EMA takes a system-driven, data-centric approach that accommodates diverse system and model designs while addressing two key deployment challenges. First, it reduces expensive model training by introducing state transformers that align the input state of a new environment with previously similar states, allowing models to warm-start adaptation. Second, it addresses the often-overlooked yet costly process of data labeling--collecting ground truth for exploring and training on various system decisions--by prioritizing labeling high-utility data while balancing the tradeoff between training and labeling cost. Evaluations on eight representative learning-based systems show that EMA reduces adaptation costs (e.g., GPU training time) by 14.9-42.4% while improving system performance (e.g., network throughput) by 6.9-31.3%.

FashionStylist: An Expert Knowledge-enhanced Multimodal Dataset for Fashion Understanding

Fashion understanding requires both visual perception and expert-level reasoning about style, occasion, compatibility, and outfit rationale. However, existing fashion datasets remain fragmented and task-specific, often focusing on item attributes, outfit co-occurrence, or weak textual supervision, and thus provide limited support for holistic outfit understanding. In this paper, we introduce FashionStylist, an expert-annotated benchmark for holistic and expert-level fashion understanding. Constructed through a dedicated fashion-expert annotation pipeline, FashionStylist provides professionally grounded annotations at both the item and outfit levels. It supports three representative tasks: outfit-to-item grounding, outfit completion, and outfit evaluation. These tasks cover realistic item recovery from complex outfits with layering and accessories, compatibility-aware composition beyond co-occurrence matching, and expert-level assessment of style, season, occasion, and overall coherence. Experimental results show that FashionStylist serves not only as a unified benchmark for multiple fashion tasks, but also as an effective training resource for improving grounding, completion, and outfit-level semantic evaluation in MLLM-based fashion systems.

A Layer-wise Analysis of Supervised Fine-Tuning

While critical for alignment, Supervised Fine-Tuning (SFT) incurs the risk of catastrophic forgetting, yet the layer-wise emergence of instruction-following capabilities remains elusive. We investigate this mechanism via a comprehensive analysis utilizing information-theoretic, geometric, and optimization metrics across model scales (1B-32B). Our experiments reveal a distinct depth-dependent pattern: middle layers (20\%-80\%) are stable, whereas final layers exhibit high sensitivity. Leveraging this insight, we propose Mid-Block Efficient Tuning, which selectively updates these critical intermediate layers. Empirically, our method outperforms standard LoRA up to 10.2\% on GSM8K (OLMo2-7B) with reduced parameter overhead, demonstrating that effective alignment is architecturally localized rather than distributed. The code is publicly available at https://anonymous.4open.science/r/base_sft.

INSPATIO-WORLD: A Real-Time 4D World Simulator via Spatiotemporal Autoregressive Modeling

Building world models with spatial consistency and real-time interactivity remains a fundamental challenge in computer vision. Current video generation paradigms often struggle with a lack of spatial persistence and insufficient visual realism, making it difficult to support seamless navigation in complex environments. To address these challenges, we propose INSPATIO-WORLD, a novel real-time framework capable of recovering and generating high-fidelity, dynamic interactive scenes from a single reference video. At the core of our approach is a Spatiotemporal Autoregressive (STAR) architecture, which enables consistent and controllable scene evolution through two tightly coupled components: Implicit Spatiotemporal Cache aggregates reference and historical observations into a latent world representation, ensuring global consistency during long-horizon navigation; Explicit Spatial Constraint Module enforces geometric structure and translates user interactions into precise and physically plausible camera trajectories. Furthermore, we introduce Joint Distribution Matching Distillation (JDMD). By using real-world data distributions as a regularizing guide, JDMD effectively overcomes the fidelity degradation typically caused by over-reliance on synthetic data. Extensive experiments demonstrate that INSPATIO-WORLD significantly outperforms existing state-of-the-art (SOTA) models in spatial consistency and interaction precision, ranking first among real-time interactive methods on the WorldScore-Dynamic benchmark, and establishing a practical pipeline for navigating 4D environments reconstructed from monocular videos.

MSVBench: Towards Human-Level Evaluation of Multi-Shot Video Generation

The evolution of video generation toward complex, multi-shot narratives has exposed a critical deficit in current evaluation methods. Existing benchmarks remain anchored to single-shot paradigms, lacking the comprehensive story assets and cross-shot metrics required to assess long-form coherence and appeal. To bridge this gap, we introduce MSVBench, the first comprehensive benchmark featuring hierarchical scripts and reference images tailored for Multi-Shot Video generation. We propose a hybrid evaluation framework that synergizes the high-level semantic reasoning of Large Multimodal Models (LMMs) with the fine-grained perceptual rigor of domain-specific expert models. Evaluating 20 video generation methods across diverse paradigms, we find that current models--despite strong visual fidelity--primarily behave as visual interpolators rather than true world models. We further validate the reliability of our benchmark by demonstrating a state-of-the-art Spearman's rank correlation of 94.4% with human judgments. Finally, MSVBench extends beyond evaluation by providing a scalable supervisory signal. Fine-tuning a lightweight model on its pipeline-refined reasoning traces yields human-aligned performance comparable to commercial models like Gemini-2.5-Flash.

Uni-MoE-2.0-Omni: Scaling Language-Centric Omnimodal Large Model with Advanced MoE, Training and Data

We present Uni-MoE 2.0 from the Lychee family. As a fully open-source omnimodal large model (OLM), it substantially advances Lychee's Uni-MoE series in language-centric multimodal understanding, reasoning, and generating. Based on the Qwen2.5-7B dense architecture, we build Uni-MoE-2.0-Omni from scratch through three core contributions: dynamic-capacity Mixture-of-Experts (MoE) design, a progressive training strategy enhanced with an iterative reinforcement strategy, and a carefully curated multimodal data matching technique. It is capable of omnimodal understanding, as well as generating images, text, and speech. Architecturally, our new MoE framework balances computational efficiency and capability for 10 cross-modal inputs using shared, routed, and null experts, while our Omni-Modality 3D RoPE ensures spatio-temporal cross-modality alignment in the self-attention layer. For training, following cross-modal pretraining, we use a progressive supervised fine-tuning strategy that activates modality-specific experts and is enhanced by balanced data composition and an iterative GSPO-DPO method to stabilise RL training and improve reasoning. Data-wise, the base model, trained on approximately 75B tokens of open-source multimodal data, is equipped with special speech and image generation tokens, allowing it to learn these generative tasks by conditioning its outputs on linguistic cues. Extensive evaluation across 85 benchmarks demonstrates that our model achieves SOTA or highly competitive performance against leading OLMs, surpassing Qwen2.5-Omni (trained with 1.2T tokens) on over 50 of 76 benchmarks. Key strengths include video understanding (+7% avg. of 8), omnimodallity understanding (+7% avg. of 4), and audiovisual reasoning (+4%). It also advances long-form speech processing (reducing WER by 4.2%) and leads in low-level image processing and controllable generation across 5 metrics.

Medical Referring Image Segmentation via Next-Token Mask Prediction

Medical Referring Image Segmentation (MRIS) involves segmenting target regions in medical images based on natural language descriptions. While achieving promising results, recent approaches usually involve complex design of multimodal fusion or multi-stage decoders. In this work, we propose NTP-MRISeg, a novel framework that reformulates MRIS as an autoregressive next-token prediction task over a unified multimodal sequence of tokenized image, text, and mask representations. This formulation streamlines model design by eliminating the need for modality-specific fusion and external segmentation models, supports a unified architecture for end-to-end training. It also enables the use of pretrained tokenizers from emerging large-scale multimodal models, enhancing generalization and adaptability. More importantly, to address challenges under this formulation-such as exposure bias, long-tail token distributions, and fine-grained lesion edges-we propose three novel strategies: (1) a Next-k Token Prediction (NkTP) scheme to reduce cumulative prediction errors, (2) Token-level Contrastive Learning (TCL) to enhance boundary sensitivity and mitigate long-tail distribution effects, and (3) a memory-based Hard Error Token (HET) optimization strategy that emphasizes difficult tokens during training. Extensive experiments on the QaTa-COV19 and MosMedData+ datasets demonstrate that NTP-MRISeg achieves new state-of-the-art performance, offering a streamlined and effective alternative to traditional MRIS pipelines.