Recent advances in 3D scene editing using NeRF and 3DGS enable high-quality static scene editing. In contrast, dynamic scene editing remains challenging, as methods that directly extend 2D diffusion models to 4D often produce motion artifacts, temporal flickering, and inconsistent style propagation. We introduce Catalyst4D, a framework that transfers high-quality 3D edits to dynamic 4D Gaussian scenes while maintaining spatial and temporal coherence. At its core, Anchor-based Motion Guidance (AMG) builds a set of structurally stable and spatially representative anchors from both original and edited Gaussians. These anchors serve as robust region-level references, and their correspondences are established via optimal transport to enable consistent deformation propagation without cross-region interference or motion drift. Complementarily, Color Uncertainty-guided Appearance Refinement (CUAR) preserves temporal appearance consistency by estimating per-Gaussian color uncertainty and selectively refining regions prone to occlusion-induced artifacts. Extensive experiments demonstrate that Catalyst4D achieves temporally stable, high-fidelity dynamic scene editing and outperforms existing methods in both visual quality and motion coherence.

Mobile Graphical User Interface (GUI) agents powered by multimodal large language models have demonstrated promising capabilities in automating complex smartphone tasks. However, existing approaches face two critical limitations: the scarcity of high-quality multilingual datasets, particularly for non-English ecosystems, and inefficient history representation methods. To address these challenges, we present SecAgent, an efficient mobile GUI agent at 3B scale. We first construct a human-verified Chinese mobile GUI dataset with 18k grounding samples and 121k navigation steps across 44 applications, along with a Chinese navigation benchmark featuring multi-choice action annotations. Building upon this dataset, we propose a semantic context mechanism that distills history screenshots and actions into concise, natural language summaries, significantly reducing computational costs while preserving task-relevant information. Through supervised and reinforcement fine-tuning, SecAgent outperforms similar-scale baselines and achieves performance comparable to 7B-8B models on our and public navigation benchmarks. Our dataset is available at https://huggingface.co/datasets/alibabagroup/CMGUI.

Cross-modal image-to-image translation among Electro-Optical (EO), Infrared (IR), and Synthetic Aperture Radar (SAR) sensors is essential for comprehensive multi-modal aerial-view analysis. However, translating between these modalities is notoriously difficult due to their distinct electromagnetic signatures and geometric characteristics. This paper presents \textbf{EarthBridge}, a high-fidelity translation framework developed for the 4th Multi-modal Aerial View Image Challenge -- Translation (MAVIC-T). We explore two distinct methodologies: \textbf{Diffusion Bridge Implicit Models (DBIM)}, which we generalize using non-Markovian bridge processes for high-quality deterministic sampling, and \textbf{Contrastive Unpaired Translation (CUT)}, which utilizes contrastive learning for structural consistency. Our EarthBridge framework employs a channel-concatenated UNet denoiser trained with Karras-weighted bridge scalings and a specialized "booting noise" initialization to handle the inherent ambiguity in cross-modal mappings. We evaluate these methods across all four challenge tasks (SAR$\rightarrow$EO, SAR$\rightarrow$RGB, SAR$\rightarrow$IR, RGB$\rightarrow$IR), achieving superior spatial detail and spectral accuracy. Our solution achieved a composite score of 0.38, securing the second position on the MAVIC-T leaderboard. Code is available at https://github.com/Bili-Sakura/EarthBridge-Preview.

Video world models have shown immense promise for interactive simulation and entertainment, but current systems still struggle with two important aspects of interactivity: user control over the environment for reproducible, editable experiences, and shared inference where players hold influence over a common world. To address these limitations, we introduce an explicit external memory into the system, a persistent state operating independent of the model's context window, that is continually updated by user actions and queried throughout the generation roll-out. Unlike conventional diffusion game engines that operate as next-frame predictors, our approach decomposes generation into Memory, Observation, and Dynamics modules. This design gives users direct, editable control over environment structure via an editable memory representation, and it naturally extends to real-time multiplayer rollouts with coherent viewpoints and consistent cross-player interactions.

Egocentric video understanding is inherently complex due to the dynamic 4D nature of the environment, where camera motion and object displacements necessitate a continuous re-evaluation of spatial relations. In this work, we target a suite of under-explored egocentric 4D reasoning tasks, including fixture interaction counting, viewpoint-relative fixture location, object movement itinerary tracking, and stationary object localization, that require fundamentally different cognitive operations: spatial anchoring, temporal tracking, and duration reasoning. We observe that these structural differences make task-agnostic approaches insufficient: generic Chain-of-Thought methods lack task-appropriate reasoning primitives, and uniform reinforcement learning actively destabilizes performance on spatial tasks. To address this, we propose EgoReasoner, a two-stage framework that aligns both the reasoning scaffold and the reward signal to each task's cognitive structure. In the first stage, Task-Adaptive Thinking Templates guide the synthesis of structured CoT traces that teach the model to reason adaptively across task types via supervised fine-tuning. In the second stage, task-aware reward functions verify entity grounding, temporal alignment, and task-adaptive logical consistency, selectively strengthening each reasoning pathway via reinforcement fine-tuning with GRPO. Our 3B-parameter model, trained on only 16K samples, achieves 37.5% average accuracy on the challenging HD-EPIC benchmark, surpassing Qwen2.5-VL-7B (25.7%) by over 10 points.

Reinforcement learning with verifiable rewards (RLVR) and Rubrics as Rewards (RaR) have driven strong gains in domains with clear correctness signals and even in subjective domains by synthesizing evaluation criteria from ideal reference answers. But many real-world tasks admit multiple valid outputs and lack the single ideal answer that rubric generation depends on. We identify this reference-free setting as a gap in current post-training methods and propose Implicit Error Counting (IEC) to fill it. Instead of checking what a response gets right against a rubric, IEC enumerates what it gets wrong, applying severity-weighted scores across task-relevant axes and converting them into calibrated per-aspect rewards. We show that naïve explicit enumeration is too noisy for stable optimization, and that two design choices: implicit score emission and group calibration are necessary to make error counting a reliable reward. As a case study, we validate IEC on virtual try-on (VTO), a domain that is simultaneously too constrained for holistic scoring and too permissive for rubric-based evaluation: subtle garment errors are unacceptable, yet many output variations are correct. We introduce Cascaded Error Counting (CEC) as an evaluation metric, which tracks human preferences well (60% top-1 vs. 30% others), and curate Mismatch-DressCode (MDressBench), a benchmark with maximal attribute mismatch to stress-test reward designs. On MDressBench, IEC outperforms RaR across all metrics (CEC: 5.31 vs. 5.60 on flat references; 5.20 vs. 5.53 on non-flat). On VITON-HD and DressCode, IEC matches or surpasses six baselines on 6 of 8 perceptual metrics. These results suggest that when ideal answers are unavailable, counting errors provide a stronger signal than constructing rubrics.

Safe highway autonomy for heavy trucks remains an open and unsolved challenge: due to long braking distances, scene understanding of hundreds of meters is required for anticipatory planning and to allow safe braking margins. However, existing driving datasets primarily cover urban scenes, with perception effectively limited to short ranges of only up to 100 meters. To address this gap, we introduce TruckDrive, a highway-scale multimodal driving dataset, captured with a sensor suite purpose-built for long range sensing: seven long-range FMCW LiDARs measuring range and radial velocity, three high-resolution short-range LiDARs, eleven 8MP surround cameras with varying focal lengths and ten 4D FMCW radars. The dataset offers 475 thousands samples with 165 thousands densely annotated frames for driving perception benchmarking up to 1,000 meters for 2D detection and 400 meters for 3D detection, depth estimation, tracking, planning and end to end driving over 20 seconds sequences at highway speeds. We find that state-of-the-art autonomous driving models do not generalize to ranges beyond 150 meters, with drops between 31% and 99% in 3D perception tasks, exposing a systematic long-range gap that current architectures and training signals cannot close.

Accurate rejection of sensitive or harmful visual content, i.e., harmful image guardrail, is critical in many application scenarios. This task must continuously adapt to the evolving safety policies and content across various domains and over time. However, traditional classifiers, confined to fixed categories, require frequent retraining when new policies are introduced. Vision-language models (VLMs) offer a more adaptable and generalizable foundation for dynamic safety guardrails. Despite this potential, existing VLM-based safeguarding methods are typically trained and evaluated under only a fixed safety policy. We find that these models are heavily overfitted to the seen policy, fail to generalize to unseen policies, and even lose the basic instruction-following ability and general knowledge. To address this issue, in this paper we make two key contributions. First, we benchmark the cross-policy generalization performance of existing VLMs with SafeEditBench, a new evaluation suite. SafeEditBench leverages image-editing models to convert unsafe images into safe counterparts, producing policy-aligned datasets where each safe-unsafe image pair remains visually similar except for localized regions violating specific safety rules. Human annotators then provide accurate safe/unsafe labels under five distinct policies, enabling fine-grained assessment of policy-aware generalization. Second, we introduce SafeGuard-VL, a reinforcement learning-based method with verifiable rewards (RLVR) for robust unsafe-image guardrails. Instead of relying solely on supervised fine-tuning (SFT) under fixed policies, SafeGuard-VL explicitly optimizes the model with policy-grounded rewards, promoting verifiable adaptation across evolving policies. Extensive experiments verify the effectiveness of our method for unsafe image guardrails across various policies.

Creating interactive digital environments for gaming, robotics, and simulation relies on articulated 3D objects whose functionality emerges from their part geometry and kinematic structure. However, existing approaches remain fundamentally limited: optimization-based reconstruction methods require slow, per-object joint fitting and typically handle only simple, single-joint objects, while retrieval-based methods assemble parts from a fixed library, leading to repetitive geometry and poor generalization. To address these challenges, we introduce ArtLLM, a novel framework for generating high-quality articulated assets directly from complete 3D meshes. At its core is a 3D multimodal large language model trained on a large-scale articulation dataset curated from both existing articulation datasets and procedurally generated objects. Unlike prior work, ArtLLM autoregressively predicts a variable number of parts and joints, inferring their kinematic structure in a unified manner from the object's point cloud. This articulation-aware layout then conditions a 3D generative model to synthesize high-fidelity part geometries. Experiments on the PartNet-Mobility dataset show that ArtLLM significantly outperforms state-of-the-art methods in both part layout accuracy and joint prediction, while generalizing robustly to real-world objects. Finally, we demonstrate its utility in constructing digital twins, highlighting its potential for scalable robot learning.

As Large Language Models (LLMs) are increasingly integrated into healthcare to address complex inquiries, ensuring their reliability remains a critical challenge. Recent studies have highlighted that generic LLMs often struggle in clinical contexts, occasionally producing misleading guidance. To mitigate these risks, this research focuses on the domain-specific adaptation of \textbf{Llama-2-7B} using the \textbf{Low-Rank Adaptation (LoRA)} technique. By injecting trainable low-rank matrices into the Transformer layers, we efficiently adapted the model using authentic patient-physician transcripts while preserving the foundational knowledge of the base model. Our objective was to enhance precision and contextual relevance in responding to medical queries by capturing the specialized nuances of clinical discourse. Due to the resource-intensive nature of large-scale human validation, the model's performance was evaluated through a dual-track framework: \textbf{Track A} utilized traditional lexical similarity metrics (e.g., BLEU, ROUGE), while \textbf{Track B} employed an "LLM-as-a-Judge" paradigm using GPT-4 for semantic assessment. Our results demonstrate that while the LoRA-enhanced model achieved significant improvements across all quantitative lexical dimensions, a profound disagreement surfaced in the GPT-4 evaluation, which marginally favored the baseline model's conversational flow. This metric divergence underscores a pivotal finding: traditional automated scores may not fully reflect clinical utility. Consequently, we propose that while automated metrics and LLM judges serve as valuable developmental proxies, rigorous validation by human medical experts remains an indispensable requirement for the safe deployment of LLMs in healthcare settings.

Understanding sources of uncertainty is fundamental to trustworthy three-dimensional scene modeling. While recent advances in neural radiance fields (NeRFs) achieve impressive accuracy in scene reconstruction and novel view synthesis, the lack of uncertainty estimation significantly limits their deployment in safety-critical settings. Existing uncertainty quantification methods for NeRFs fail to separately capture both aleatoric and epistemic uncertainties. Among those that do quantify one or the other, many of them either compromise rendering quality or incur significant computational overhead to obtain uncertainty estimates. To address these issues, we introduce Evidential Neural Radiance Fields, a probabilistic approach that seamlessly integrates with the NeRF rendering process, enabling direct quantification of both aleatoric and epistemic uncertainties from a single forward pass. We compare multiple uncertainty quantification methods on three standardized benchmarks, where our approach demonstrates state-of-the-art scene reconstruction fidelity and uncertainty estimation quality. Code is available at https://github.com/KerryDRX/EvidentialNeRF.

Recommendation model performance is intrinsically tied to the quality, volume, and relevance of their training data. To address common challenges like data sparsity and cold start, recent researchs have leveraged data from multiple auxiliary domains to enrich information within the target domain. However, inherent domain gaps can degrade the quality of mixed-domain data, leading to negative transfer and diminished model performance. Existing prevailing \emph{model-centric} paradigm -- which relies on complex, customized architectures -- struggles to capture the subtle, non-structural sequence dependencies across domains, leading to poor generalization and high demands on computational resources. To address these shortcomings, we propose \textsc{Taesar}, a \emph{data-centric} framework for \textbf{t}arget-\textbf{a}lign\textbf{e}d \textbf{s}equenti\textbf{a}l \textbf{r}egeneration, which employs a contrastive decoding mechanism to adaptively encode cross-domain context into target-domain sequences. It employs contrastive decoding to encode cross-domain context into target sequences, enabling standard models to learn intricate dependencies without complex fusion architectures. Experiments show \textsc{Taesar} outperforms model-centric solutions and generalizes to various sequential models. By generating enriched datasets, \textsc{Taesar} effectively combines the strengths of data- and model-centric paradigms. The code accompanying this paper is available at~ \textcolor{blue}{https://github.com/USTC-StarTeam/Taesar}.

Generalized Category Discovery (GCD) aims to identify both known and unknown categories, with only partial labels given for the known categories, posing a challenging open-set recognition problem. State-of-the-art approaches for GCD task are usually built on multi-modality representation learning, which is heavily dependent upon inter-modality alignment. However, few of them cast a proper intra-modality alignment to generate a desired underlying structure of representation distributions. In this paper, we propose a novel and effective multi-modal representation learning framework for GCD via Semi-Supervised Rate Reduction, called SSR$^2$-GCD, to learn cross-modality representations with desired structural properties based on emphasizing to properly align intra-modality relationships. Moreover, to boost knowledge transfer, we integrate prompt candidates by leveraging the inter-modal alignment offered by Vision Language Models. We conduct extensive experiments on generic and fine-grained benchmark datasets demonstrating superior performance of our approach.

Reinforcement learning fine-tuning has proven effective for steering generative diffusion models toward desired properties in image and molecular domains. Graph diffusion models have similarly been applied to combinatorial structure generation, including neural architecture search (NAS). However, neural architectures are directed acyclic graphs (DAGs) where edge direction encodes functional semantics such as data flow-information that existing graph diffusion methods, designed for undirected structures, discard. We propose Directed Graph Policy Optimization (DGPO), which extends reinforcement learning fine-tuning of discrete graph diffusion models to DAGs via topological node ordering and positional encoding. Validated on NAS-Bench-101 and NAS-Bench-201, DGPO matches the benchmark optimum on all three NAS-Bench-201 tasks (91.61%, 73.49%, 46.77%). The central finding is that the model learns transferable structural priors: pretrained on only 7% of the search space, it generates near-oracle architectures after fine-tuning, within 0.32 percentage points of the full-data model and extrapolating 7.3 percentage points beyond its training ceiling. Bidirectional control experiments confirm genuine reward-driven steering, with inverse optimization reaching near random-chance accuracy (9.5%). These results demonstrate that reinforcement learning-steered discrete diffusion, once extended to handle directionality, provides a controllable generative framework for directed combinatorial structures.

Humans can infer the three-dimensional structure of objects from two-dimensional visual inputs. Modeling this ability has been a longstanding goal for the science and engineering of visual intelligence, yet decades of computational methods have fallen short of human performance. Here we develop a modeling framework that predicts human 3D shape inferences for arbitrary objects, directly from experimental stimuli. We achieve this with a novel class of neural networks trained using a visual-spatial objective over naturalistic sensory data; given a set of images taken from different locations within a natural scene, these models learn to predict spatial information related to these images, such as camera location and visual depth, without relying on any object-related inductive biases. Notably, these visual-spatial signals are analogous to sensory cues readily available to humans. We design a zero-shot evaluation approach to determine the performance of these 'multi-view' models on a well established 3D perception task, then compare model and human behavior. Our modeling framework is the first to match human accuracy on 3D shape inferences, even without task-specific training or fine-tuning. Remarkably, independent readouts of model responses predict fine-grained measures of human behavior, including error patterns and reaction times, revealing a natural correspondence between model dynamics and human perception. Taken together, our findings indicate that human-level 3D perception can emerge from a simple, scalable learning objective over naturalistic visual-spatial data. Code, images, and human data needed to reproduce all analyses can be found at https://tzler.github.io/human_multiview/

We present the first comprehensive, large-scale study of training long-context vision language models up to 344K context, targeting long-document visual question answering with measured transfer to long-context text. While several such strong are open-weight, namely Qwen3 VL and GLM 4.5/6V, their training recipes and data pipelines are not reproducible. We systematically study continued pretraining, supervised finetuning, and preference optimization for 24B and 32B parameter models, backed by extensive LC evaluations and ablations to bridge this gap, and achieve state-of-the-art performance on MMLongBenchDoc for both parameter scales. In addition to this, our key findings include: (i) training on context lengths that match evaluation context lengths outperforms training on longer contexts, (ii) training and evaluating with page indices provides a simple, high-impact boost to long-document performance, (iii) our synthetic data pipelines enable self-improvement via continued pretraining and supervised finetuning, and (iv) we extend the known text-to-visual long context transfer to the reverse, showing that visual long context training transfers to long-context text performance. We also release MMLBD-C, a manually corrected version of MMLongBenchDoc to reduce erroneous and low quality examples in the benchmark.

Inference-Time-Compute (ITC) methods like Best-of-N and Tree-of-Thoughts are meant to produce output candidates that are both high-quality and diverse, but their use of high-temperature sampling often fails to achieve meaningful output diversity. Moreover, existing ITC methods offer limited control over how to perform reasoning, which in turn limits their interpretability. We present STATe Of Thoughts (STATe), an interpretable ITC method that searches over high-level reasoning patterns. STATe replaces stochastic sampling with discrete and interpretable textual interventions: a controller selects actions encoding high-level reasoning choices; a generator produces reasoning steps conditioned on those choices; and an evaluator scores candidates to guide search. This structured approach yields three main advantages. First, action-guided textual interventions reliably influence LLM generations and produce greater response diversity than temperature-based sampling. Second, in a case study on argument generation, STATe's explicit action sequences capture interpretable features that are highly predictive of output quality. Third, estimating the association between performance and action choices allows us to identify promising yet unexplored regions of the action space and steer generation toward them. Together, these results establish STATe as both a practical framework for diverse and controllable text generation, and as a tool for understanding the reasoning patterns that drive performance.

Text-driven image and video editing can be naturally cast as inpainting problems, where masked regions are reconstructed to remain consistent with both the observed content and the editing prompt. Recent advances in test-time guidance for diffusion and flow models provide a principled framework for this task; however, existing methods rely on costly vector--Jacobian product (VJP) computations to approximate the intractable guidance term, limiting their practical applicability. Building upon the recent work of Moufad et al. (2025), we provide theoretical insights into their VJP-free approximation and substantially extend their empirical evaluation to large-scale image and video editing benchmarks. Our results demonstrate that test-time guidance alone can achieve performance comparable to, and in some cases surpass, training-based methods.

Arabic Language Models (LMs) are pretrained predominately on Modern Standard Arabic (MSA) and are expected to transfer to its dialects. While MSA as the standard written variety is commonly used in formal settings, people speak and write online in various dialects that are spread across the Arab region. This poses limitations for Arabic LMs, since its dialects vary in their similarity to MSA. In this work we study cross-lingual transfer of Arabic models using probing on 3 Natural Language Processing (NLP) Tasks, and representational similarity. Our results indicate that transfer is possible but disproportionate across dialects, which we find to be partially explained by their geographic proximity. Furthermore, we find evidence for negative interference in models trained to support all Arabic dialects. This questions their degree of similarity, and raises concerns for cross-lingual transfer in Arabic models.

Deploying Large Language Models (LLMs) for discriminative workloads is often limited by inference latency, compute, and API costs at scale. Active distillation reduces these costs by querying an LLM oracle to train compact discriminative students, but most pipelines distill only final labels, discarding intermediate reasoning signals and offering limited diagnostics of what reasoning is missing and where errors arise. We propose Graph of Concept Predictors (GCP), a reasoning-aware active distillation framework that externalizes the teacher's decision process as a directed acyclic graph and mirrors it with modular concept predictors in the student. GCP enhances sample efficiency through a graph-aware acquisition strategy that targets uncertainty and disagreement at critical reasoning nodes. Additionally, it improves training stability and efficiency by performing targeted sub-module retraining, which attributes downstream loss to specific concept predictors and updates only the most influential modules. Experiments on eight NLP classification benchmarks demonstrate that GCP enhances performance under limited annotation budgets while yielding more interpretable and controllable training dynamics. Code is available at: https://github.com/Ziyang-Yu/GCP.

Extracting structured data from the web is often a trade-off between the brittle nature of manual heuristics and the prohibitive cost of Large Language Models. We introduce AXE (Adaptive X-Path Extractor), a pipeline that rethinks this process by treating the HTML DOM as a tree that needs pruning rather than just a wall of text to be read. AXE uses a specialized "pruning" mechanism to strip away boilerplate and irrelevant nodes, leaving behind a distilled, high-density context that allows a tiny 0.6B LLM to generate precise, structured outputs. To keep the model honest, we implement Grounded XPath Resolution (GXR), ensuring every extraction is physically traceable to a source node. Despite its low footprint, AXE achieves state-of-the-art zero-shot performance, outperforming several much larger, fully-trained alternatives with an F1 score of 88.1% on the SWDE dataset. By releasing our specialized adaptors, we aim to provide a practical, cost-effective path for large-scale web information extraction. Our code and adaptors are publicly available at https://github.com/abdo-Mansour/axetract.

Existing video avatar models have demonstrated impressive capabilities in scenarios such as talking, public speaking, and singing. However, the majority of these methods exhibit limited alignment with respect to text instructions, particularly when the prompts involve complex elements including large full-body movement, dynamic camera trajectory, background transitions, or human-object interactions. To break out this limitation, we present JoyAvatar, a framework capable of generating long duration avatar videos, featuring two key technical innovations. Firstly, we introduce a twin-teacher enhanced training algorithm that enables the model to transfer inherent text-controllability from the foundation model while simultaneously learning audio-visual synchronization. Secondly, during training, we dynamically modulate the strength of multi-modal conditions (e.g., audio and text) based on the distinct denoising timestep, aiming to mitigate conflicts between the heterogeneous conditioning signals. These two key designs serve to substantially expand the avatar model's capacity to generate natural, temporally coherent full-body motions and dynamic camera movements as well as preserve the basic avatar capabilities, such as accurate lip-sync and identity consistency. GSB evaluation results demonstrate that our JoyStreamer model outperforms the state-of-the-art models such as Omnihuman-1.5 and KlingAvatar 2.0. Moreover, our approach enables complex applications including multi-person dialogues and non-human subjects role-playing. Some video samples are provided on https://joystreamer.github.io/.

Large Vision-Language Models (VLMs) often answer classic visual illusions "correctly" on original images, yet persist with the same responses when illusion factors are inverted, even though the visual change is obvious to humans. This raises a fundamental question: do VLMs perceive visual changes or merely recall memorized patterns? While several studies have noted this phenomenon, the underlying causes remain unclear. To move from observations to systematic understanding, this paper introduces VI-Probe, a controllable visual-illusion framework with graded perturbations and matched visual controls (without illusion inducer) that disentangles visually grounded perception from language-driven recall. Unlike prior work that focuses on averaged accuracy, we measure stability and sensitivity using Polarity-Flip Consistency, Template Fixation Index, and an illusion multiplier normalized against matched controls. Experiments across different families reveal that response persistence arises from heterogeneous causes rather than a single mechanism. For instance, GPT-5 exhibits memory override, Claude-Opus-4.1 shows perception-memory competition, while Qwen variants suggest visual-processing limits. Our findings challenge single-cause views and motivate probing-based evaluation that measures both knowledge and sensitivity to controlled visual change. Data and code are available at https://sites.google.com/view/vi-probe/

Large Language Models (LLMs) generate fluent text, yet whether they truly understand the world or merely produce plausible texts about it remains contested. We propose an architectural principle, the mouth is not the brain, that explicitly separates world models from language models. Our architecture comprises three components: a DBM that captures domain structure as an energy-based world model, an adapter that projects latent belief states into embedding space, and a frozen GPT-2 that provides linguistic competence without domain knowledge. We instantiate this framework in the consumer review domain using Amazon smartphone reviews. Experiments demonstrate that (1) world model conditioning achieves lower cross-entropy loss and higher semantic similarity than architectural baselines including direct projection and full fine-tuning, while qualitative analysis reveals that soft prompt conditioning resolves a trade-off that prompt-based approaches cannot: simple prompts lack expressiveness while detailed prompts cause output collapse in small LLMs; (2) the DBM's energy function distinguishes coherent from incoherent market configurations, assigning higher energy to implausible brand-price combinations; and (3) interventions on specific attributes propagate causally to generated text with intervened outputs exhibiting distributions statistically consistent with naturally occurring samples sharing the target configuration. These findings suggest that even small-scale language models can achieve consistent, controllable generation when connected to an appropriate world model, providing empirical support for separating linguistic competence from world understanding.

Diffusion model alignment aims to bridge the gap between generated outputs and human preferences by enhancing both semantic consistency with textual prompts and overall visual quality. Existing alignment methods face a challenging trade-off: test-time approaches enable input-specific adaptability but introduce significant computational overhead and tend to under-optimize, while fine-tuning approaches risk reward over-optimization and loss of generation diversity. To bridge this gap, we propose HyperAlign, a framework that trains a hypernetwork for efficient and effective test-time alignment. Instead of modifying latent states directly, HyperAlign dynamically generates input-and-state-conditioned low-rank adaptation weights to modulate the denoising trajectory toward target rewards. We introduce multiple HyperAlign variants of varying granularity to balance alignment quality and computational efficiency. The hypernetwork is optimized with a reward objective regularized by preference data to mitigate reward hacking. We evaluate HyperAlign across multiple generative paradigms, including Stable Diffusion and FLUX, where it significantly outperforms existing alignment methods in semantic consistency and visual quality.

Scale does not uniformly improve reasoning - it restructures it. Analyzing 25,000+ chain-of-thought trajectories across four domains (Law, Science, Code, Math) and two scales (8B, 70B parameters), we discover that neural scaling laws trigger domain-specific phase transitions rather than uniform capability gains. Legal reasoning undergoes Crystallization: 45% collapse in representational dimensionality (d95: 501 -> 274), 31% increase in trajectory alignment, and 10x manifold untangling. Scientific and mathematical reasoning remain Liquid - geometrically invariant despite 9x parameter increase. Code reasoning forms a discrete Lattice of strategic modes (silhouette: 0.13 -> 0.42). This geometry predicts learnability. We introduce Neural Reasoning Operators - learned mappings from initial to terminal hidden states. In crystalline legal reasoning, our operator achieves 63.6% accuracy on held-out tasks via probe decoding, predicting reasoning endpoints without traversing intermediate states. We further identify a universal oscillatory signature (coherence ~ -0.4) invariant across domains and scales, suggesting attention and feedforward layers drive reasoning through opposing dynamics. These findings establish that the cost of thought is determined not by task difficulty but by manifold geometry - offering a blueprint for inference acceleration where topology permits.

Large language models (LLMs) face significant deployment challenges due to their massive computational demands. % While pruning offers a promising compression solution, existing methods suffer from two critical limitations: (1) They neglect activation distribution shifts between calibration data and test data, resulting in inaccurate error estimations; (2) They overlook the long-tail distribution characteristics of activations in the attention module. To address these limitations, this paper proposes a novel pruning method, $D^2Prune$. First, we propose a dual Taylor expansion-based method that jointly models weight and activation perturbations for precise error estimation, leading to precise pruning mask selection and weight updating and facilitating error minimization during pruning. % Second, we propose an attention-aware dynamic update strategy that preserves the long-tail attention pattern by jointly minimizing the KL divergence of attention distributions and the reconstruction error. Extensive experiments show that $D^2Prune$ consistently outperforms SOTA methods across various LLMs (e.g., OPT-125M, LLaMA2/3, and Qwen3). Moreover, the dynamic attention update mechanism also generalizes well to ViT-based vision models like DeiT, achieving superior accuracy on ImageNet-1K.

Most existing 3D referring expression segmentation (3DRES) methods rely on dense, high-quality point clouds, while real-world agents such as robots and mobile phones operate with only a few sparse RGB views and strict latency constraints. We introduce Multi-view 3D Referring Expression Segmentation (MV-3DRES), where the model must recover scene structure and segment the referred object directly from sparse multi-view images. Traditional two-stage pipelines, which first reconstruct a point cloud and then perform segmentation, often yield low-quality geometry, produce coarse or degraded target regions, and run slowly. We propose the Multimodal Visual Geometry Grounded Transformer (MVGGT), an efficient end-to-end framework that integrates language information into sparse-view geometric reasoning through a dual-branch design. Training in this setting exposes a critical optimization barrier, termed Foreground Gradient Dilution (FGD), where sparse 3D signals lead to weak supervision. To resolve this, we introduce Per-view No-target Suppression Optimization (PVSO), which provides stronger and more balanced gradients across views, enabling stable and efficient learning. To support consistent evaluation, we build MVRefer, a benchmark that defines standardized settings and metrics for MV-3DRES. Experiments show that MVGGT establishes the first strong baseline and achieves both high accuracy and fast inference, outperforming existing alternatives. The code is available at https://mvggt.github.io/.

Understanding real-world videos such as movies requires integrating visual and dialogue cues. Yet existing VideoQA benchmarks struggle to capture this multimodal reasoning and, given the difficulty of evaluating free-form answers, largely resort to simple multiple choice questions. We introduce a novel open-ended multimodal VideoQA benchmark, MovieRecapsQA, created using movie recap videos -- a distinctive type of YouTube content that summarizes a film via a voiceover description of key clips from the movie (recap video). From the transcribed voiceover (recap summary) of 60 recap videos, we generate $\approx$8.2K questions along with the necessary ``facts'' expected in each answer; the former facilitates the creation of questions that require mutimodal reasoning and the latter allow the construction of a reference-free evaluation metric that can be applied to open-ended responses. To our knowledge, this is the first reference-free open-ended VideoQA benchmark. The benchmark allows each question to be evaluated in different input video settings: given (a) the full-length movie, (b) the full ($\approx$11 min) recap video (visual only), (c) $\approx$14 min of aligned movie scenes, i.e, movie scenes relevant to the question, and (d) $\approx$1.2 min of aligned recap video scenes. In all cases, the text of any associated movie dialogue is provided. Each question is categorized by the modality required to answer it -- visual, dialogue, or both -- enabling fine-grained evaluation of multimodal capabilities. We benchmark (setting (d)) seven state-of-the-art MLLMs and find that (i) only our reference-free metric produces meaningful human-aligned model separation; (ii) vision-centric questions yield the lowest scores across all models; (iii) removing visual input often \textit{improves} model factuality; and (iv) the primary bottleneck is visual perception, not visual reasoning.

The diffusion model presents a powerful ability to capture the entire (conditional) data distribution. However, due to the lack of sufficient training and data to learn to cover low-probability areas, the model will be penalized for failing to generate high-quality images corresponding to these areas. To achieve better generation quality, guidance strategies such as classifier free guidance (CFG) can guide the samples to the high-probability areas during the sampling stage. However, the standard CFG often leads to over-simplified or distorted samples. On the other hand, the alternative line of guiding diffusion model with its bad version is limited by carefully designed degradation strategies, extra training and additional sampling steps. In this paper, we proposed a simple yet effective strategy Internal Guidance (IG), which introduces an auxiliary supervision on the intermediate layer during training process and extrapolates the intermediate and deep layer's outputs to obtain generative results during sampling process. This simple strategy yields significant improvements in both training efficiency and generation quality on various baselines. On ImageNet 256x256, SiT-XL/2+IG achieves FID=5.31 and FID=1.75 at 80 and 800 epochs. More impressively, LightningDiT-XL/1+IG achieves FID=1.34 which achieves a large margin between all of these methods. Combined with CFG, LightningDiT-XL/1+IG achieves the current state-of-the-art FID of 1.19.