DeXplicit addresses action-conditioned world modeling for dexterous manipulation by splitting prediction into an explicit metric 3D dynamics stage and a subsequent video rendering stage. Instead of entangling physics, viewpoint, and appearance in a monolithic video generator, it first forecasts structured object motion and deformation, then uses those predictions as geometric scaffolds for efficient causal video synthesis. The core claim is that making physical state explicit improves both geometric accuracy and generation speed.

Stream4D tackles online 4D reconstruction in dynamic scenes, where moving objects break static-scene assumptions and cause pose drift and ghosting artifacts in causal streaming pipelines. The core idea is a training-free plug-in that mines temporal motion cues from the KV cache of causal Transformers, fuses them with depth priors to generate dynamic masks, and suppresses dynamic keys via soft masking to prevent interference during historical aggregation. By operating directly on cached keys without architectural changes or retraining, it aims to make existing streaming backbones like Stream3R and StreamVGGT robust to dynamic content while retaining linear-time scalability.

Future lunar rovers will navigate tens of kilometers using cameras, but the feature detectors and matchers at the heart of visual odometry are benchmarked almost exclusively on urban and driving scenes. This paper evaluates eleven classical and learned detector–matcher combinations on nine synthetic, lunar-testbed, and terrestrial-analogue datasets, measuring how pairwise pose accuracy degrades with viewpoint change and how the resulting correspondences drive open-loop VO error. The central finding is that learned SuperPoint pairings outperform classical SIFT on real low-texture lunar terrain and remain robust across lighting, yet synthetic datasets alone are misleading predictors of real performance and the learned advantage narrows when short-baseline translation direction is scored jointly. The work directly addresses a critical gap for Artemis-era rover autonomy by showing precisely where classical front ends fail and under what conditions learned ones are worth their computational cost.

Transformer attention is commonly viewed as an interacting particle system, yet the mean-field limit on probability densities over $\mathbb{R}^d$ lacks rigorous geometric foundations on non-compact domains. This paper derives accelerated attention blocks from Nesterov-type damped Hamiltonian dynamics on spaces of probability measures, introducing a local-mobility geometry for softmax attention and a quotient Stein geometry for linear attention. The theory yields sharp PDE well-posedness and finite-time blow-up results, plus quantitative links to token-norm growth across discrete layers. Small-scale language-modeling experiments report minor validation-loss improvements at increased computational cost.

GLOW-FM attacks the high inference cost of diffusion-based low-light image enhancement (LLIE) by replacing iterative sampling with a single-step latent flow. The core idea is to run Terminal Velocity Matching (TVM) in the compact latent space of a frozen Tiny-VAE, directly predicting the displacement from a degraded low-light latent to its well-exposed counterpart in one function evaluation. If its quality claims hold up, the result would be a generative LLIE method practical enough for edge-device deployment.

High-dimensional nonparametric density estimation is notoriously plagued by the curse of dimensionality. This paper proposes to factorize the joint distribution autoregressively into a product of one marginal and $d-1$ conditional densities, estimating each factor with parallel gradient boosting (PGB), a fast tree-based multi-quantile regression method. The authors prove that for subadditive divergences the joint convergence rate is governed by the hardest factor, and that when conditional estimators adapt to sparsity the optimal rate relates to the treewidth of an underlying dependence graph. Together, the theory and experiments suggest that flexible autoregressive factorization offers a practical off-the-shelf alternative to kernels, copulas, and random forests for moderate-to-high dimensional continuous data.

FlexPart addresses a key limitation in controllable 3D asset creation: existing image-to-3D part generation systems are typically specialized for a single prompt type, forcing users into rigid workflows. The paper introduces a unified Diffusion Transformer framework that simultaneously accepts sparse point prompts, 2D bounding boxes, and dense masks via Gated Adaptive Layer Normalization, while an Asymmetric Geometric Guidance mechanism uses differentiable optimal transport to let dense spatial priors supervise sparse representations during training. By jointly training across prompt granularities, the method aims to improve part-level geometric fidelity and give users flexible control over how they specify structure.

Motion segmentation is conventionally treated as an object-level property, causing entire instances to be marked moving even when only a small part displaces or when the camera itself moves. This paper reframes motion as a relation between an ordered source–target frame pair and proposes MoRSe, a feed-forward model that predicts moving-region masks without optical flow, point tracks, or a frozen segmentation foundation model at inference. To support this formulation the authors introduce MoRSe-5K, a benchmark with Static, Dynamic-Static, and Dynamic splits that directly measures whether a model can suppress false positives under pure camera motion.

Scientific inference often requires evaluating and optimizing complete trajectories, yet existing methods typically learn local one-step dynamics or train black-box generators that lack a reusable path objective. This paper proposes action learning, the task of recovering a stochastic action from trajectory observations, and introduces LEAP, which learns a path-space correction relative to a tractable reference process via noise-contrastive trajectory comparisons. By constructing a differentiable objective on complete paths, LEAP captures nonadditive temporal dependencies and enables downstream reuse for ranking, constrained inference, and sampling without retraining.

Vision-language models lag at visual in-context learning because they exploit textual shortcuts rather than learning from demonstration images. This work argues the root cause is insufficient training data that forces pure visual induction, and proposes MIME—a synthetic, task-agnostic dataset where the rule is never stated in text—and DTAD, a label-dropout mechanism that shifts attention to visual tokens. The result is an 8.1–18.0 point jump in vICL accuracy across multiple architectures, with an 8B model matching frontier proprietary systems, suggesting vICL-specific data should become a standard ingredient in VLM training.

H-2DGS addresses the need for sharp, photorealistic neurosurgical surface reconstruction that can fuse with MRI and tractography without obscuring internal brain structures. The authors propose a three-stage pipeline that bootstraps appearance with conventional 3D Gaussian Splatting, contracts volumetric density along camera rays using differentiable quantiles, and converts the fused TSDF surface into curved height-field Gaussian patches. By explicitly representing a thin, deformable cortical sheet rather than a diffuse volume, the method targets geometric precision and compactness at the cost of some rendering fidelity.

Multimodal large language models can answer clinical questions about retinal OCT scans, but without spatial grounding clinicians cannot verify whether a model attended to real pathology or relied on spurious shortcuts. This paper introduces RetinaGroundBench, the first benchmark for biomarker visual grounding in retinal OCT, together with $AggIoU$—a metric tailored to MLLMs that fragment lesions into multiple boxes—and RetGrounder, an 8B domain-adapted model that outperforms general-purpose models up to 38B parameters. The core message is that grounding remains largely unsolved: even the best open-source model achieves only 23.7% macro $F_1$, and VQA ability does not imply localization skill. The work matters because trustworthy clinical deployment of MLLMs requires precisely this kind of verifiable spatial evidence.

This paper audits SkinCon, a widely used dermatology concept dataset, for skin-tone-dependent annotation bias.
Within the same disease, erythema is labeled far less often on dark skin (odds ratio 0.17) while pigmentary concepts are labeled more often.
An evidential concept model trained on these annotations reproduces the bias, and three label-error detectors fall below a disease-only baseline at catching potential erythema omissions.
The work warns that concept bottleneck explanations for skin disease diagnosis cannot be trusted until annotations are audited within each subgroup.

Detailed image captioning is usually treated as a global coverage task, but captions are generated as ordered sequences where later tokens can degrade in different ways. This paper introduces a position-aware rubric that decomposes captions into quartile-wise supported, hallucinated, and generic content, exposing two distinct failure modes: mid-caption unsupported specificity and tail generic escape. Motivated by this diagnosis, the authors propose PAGER, a GRPO reward with separate signals for middle evidence, tail density, and length control, achieving strong pairwise wins while preserving competitive global quality metrics.

The paper presents structure-preserving reduced-order models for two high-speed aeroelastic challenge problems—a clamped panel at Mach 1.92 (RC-19) and a cantilever plate at Mach 5.85 (HyMAX). Because the experimental packages lack time-resolved responses or repeated runs, the author proposes a learned model that retains the exact physics-based potential energy, fits generalized aerodynamic and damping forces by least squares, and trains only a dissipation matrix engineered to remain positive definite. An exact discrete-gradient integrator then guarantees an energy balance to machine precision regardless of the trained parameters, aiming to show that physical structure, not large datasets, can carry predictive capability in chaotic flutter regimes.

HiSTEVE tackles the challenge of building Minecraft agents that can perform long-horizon tasks using only native keyboard-and-mouse control, without relying on environment APIs or predefined skills. It proposes a hierarchical VLA agent that pairs a capability-aware harness—equipped with structured memory, a VLM planner, and a policy-coupled completion detector—with a single low-level VLA policy for visuomotor execution. The core idea is to close the loop between high-level planning and raw pixel-to-action control in a partially observable open world, which matters for developing generally capable game-playing agents.

Fine-tuning diffusion models has yielded over 100K community adapters, but merging them interactively creates a high-dimensional, expensive black-box optimization problem: prior work requires live image generation in every round and is limited to a few dozen preselected adapters. HimmBO splits the search into a screening stage that recovers promising adapters via level-set estimation over a precomputed gallery, followed by a refinement stage that optimizes merge coefficients only within that low-dimensional active set. A style-aware Gaussian-process kernel transfers preference information between visually similar adapters, allowing the system to scale to libraries of over one thousand adapters while using under one tenth the image generations of prior work.

Users increasingly rely on AI assistants for complex, multi-turn tasks, but evaluating these systems requires simulators that can both follow intricate intents and behave like real humans. This paper proposes UserLM-2.0, an 8B-parameter simulator trained with multi-turn reinforcement learning that jointly optimizes task fidelity and naturalness through a lightweight, dynamically retrained discriminator. By reaching $46.2\%$ conversation validity—over triple prior state-of-the-art—it demonstrates a practical path forward, though a sizable gap to real human conversations remains.

Sequence compression is a critical bottleneck in efficient language and multimodal models, yet most architectures rely on simple pooling or MLP projections that ignore inter-token correlations. This paper reframes compression as a test-time ridge regression problem: input tokens define a regression whose fitted coefficients become the compressed vector. The resulting TTR-Compressor solves exactly in closed form via a small Cholesky factorization, introduces only $2d$ learnable parameters, and yields strong gains on long-context retrieval and visual token merging.

Multimodal sentiment analysis fuses linguistic, acoustic, and visual signals, yet standard fusion methods entangle modality-specific and cross-modal contributions in opaque joint representations. This paper proposes I2-KAN, a multi-pathway framework that explicitly maintains intra-modal and pairwise inter-modal information in separate pathways, refines each with a Task-Decoupled Linear-Gated Denoising Bottleneck, and fuses them via a structured Kolmogorov–Arnold Network. The core motivation is that a modality's standalone predictive utility can differ sharply from its interaction-conditioned utility, and the work aims to make this reasoning traceable at pathway, feature, and functional-edge levels.