Medical image parsing requires a single model to produce heterogeneous outputs—categorical labels, bounding boxes, counts, numeric estimates, and free-text reports—across disparate imaging modalities. This work studies task-specific adaptation of a shared 4-bit Qwen3.5-9B backbone using lightweight LoRA adapters, where a question-only GLiClass router selects the adapter, image budget, and prompt for each case on the FLARE 2026 Task 3 benchmark. The core finding is that adaptation recipes should vary by task—rank, resolution, and prompts allocated according to demand—rather than applied uniformly.

PAWA tackles multi-agent world modeling in shared environments like Minecraft, where one player's actions change what every other player sees—not just in their local field of view, but in directions they are not currently looking. The core idea is to jointly forecast each player's point-of-view (POV) video alongside a world-stabilized panoramic video, conditioned on all players' actions. The model uses a dual-tower architecture with ray-aligned cross-attention that links tokens viewing the same world direction regardless of whether they appear in rectilinear POV or equirectangular panorama coordinates, plus a quadratic-phase player encoding that scales identities as $m^2$ instead of $m+1$. The work also introduces a scalable dataset pipeline that records gameplay without graphics clients and renders replays offline, supporting more than 64 simultaneous players.

This paper asks whether releasing learned embeddings from pretrained brain MRI encoders—instead of raw scans—actually protects patient privacy. The authors audit five diverse encoders on two longitudinal datasets, measuring patient linkage, re-identification, and sex-attribute inference under an explicit attacker model. They find that several embeddings retain enough subject-specific structure to enable rank-1 re-identification above 90% on both datasets, and that sex can be inferred with balanced accuracy up to 95.5%. The core takeaway is that embedding release alone is insufficient to establish anonymity.

StateCraft tackles the long-horizon inconsistency problem in video world models by explicitly decoupling a persistent, structured abstract state from the visual rendering pipeline. A finetuned vision-language model maintains and updates a graph-structured abstract state in response to actions, including off-screen changes, while a video diffusion model generates each frame conditioned on that state. This enables consistent revisits, direct state editing, and large empirical gains on a new GUI-world-model benchmark measuring memory and controllability.

LiFi targets the task of generating metric LiDAR point clouds from cheap onboard multi-view camera images for autonomous driving. The core idea is a latent diffusion model guided by two decoupled branches: a geometry branch that aligns dense monocular depth estimates to the target LiDAR scanning pattern via a reverse sampling algorithm and uncertainty-modulated encoding, and a semantic branch that aggregates global cross-view context. If the approach holds up, it could enable large-scale data augmentation, cross-modal simulation, and sensor completion without requiring expensive LiDAR capture at deployment time.

SceneAE proposes a sparse 3D-anchored latent representation for scene reconstruction and camera-controllable generation. Instead of viewpoint-aligned video latents, each latent couples an explicit 3D point with an implicit feature, enabling queryable rendering from arbitrary cameras. A two-stage diffusion model decouples geometry generation from content synthesis, improving camera control and cutting computational cost roughly in half compared with dense-grid baselines.

RGB vision-language-action models struggle to distinguish waste objects that appear identical in color but differ spectrally. SPECTRA-VLA learns a compact, base-orthogonal spectral residual that enriches frozen visual tokens at their native width without retraining the visual encoder or action decoder. On the SpectralWaste dataset, the method improves category balanced accuracy across 86 held-out acquisition sessions while retaining within-category spectral shape better than category-level codes.

Arctic coastal polynyas drive critical ocean-atmosphere heat exchange and sea-ice production, yet mapping them from passive microwave sea ice concentration (SIC) is fraught with ambiguity because narrow coastal openings share SIC ranges with the marginal ice zone and open ocean. The authors develop a daily mapping framework that couples hierarchical SIC partitioning, connected-component analysis, and differentiated region growing to separate enclosed low-concentration polynyas from ocean-connected water, then apply it to September–May seasons from 2002/03 to 2024/25. Their central finding is that absolute polynya extent is highly sensitive to both the upper SIC threshold and the input product—varying by up to roughly 65% across thresholds and showing 68% MAPE across products—while broad temporal patterns remain comparatively stable. This makes a clear case that reproducible long-term monitoring requires a fixed, explicitly documented configuration rather than ad hoc threshold choices.

HyrCap addresses inaccurate temporal localization and low-quality proposals in long untrimmed videos by treating dense query-based predictions as an over-complete temporal event hypothesis space rather than final detections. It hybridizes complementary video representations—InternVideo2 for strong proposal quality and V-JEPA2 for localization reliability—and recalibrates proposal confidence via cross-representation consensus, localization quality estimation, and redundancy suppression. This matters because reliable temporal proposals are framed as a critical interface between detection and LLM-driven semantic understanding, motivating a new two-stage evaluation protocol that decouples localization from segment-level semantics.

The paper tackles whether iterative image generation with flow matching can actually perform deep sequential reasoning, or whether the many sampling steps are merely an illusion of compute. It proves that vanilla flow matching collapses to constant-depth circuits because each ODE step changes state by only $O(1/K)$, but shows that self-conditioning—feeding back an endpoint prediction as an internal scratchpad—breaks this barrier and can reach $\mathsf{P/poly}$ and even $\mathsf{PSPACE/poly}$. The result is a sharp theoretical and empirical argument that how information flows between sampling steps matters more than how many steps you take.

Cameras and foundation models recognize what a 3D scene contains, but not what it is made of: surfaces of the same colour can be different materials. Spectral World Renderer tackles this by unifying world generation and hyperspectral unmixing inside 3D Gaussian splatting, so that each Gaussian carries material abundances decoded before ray compositing. The resulting SWR-Bench provides eight synthetic worlds with exact per-pixel material labels, enabling verifiable decomposition benchmarks and real-time material queries.

Holi-Spatial 360 targets a key data bottleneck for spatial intelligence: creating large-scale, photorealistic, semantically grounded 3D indoor environments without costly LiDAR or dense video capture. The pipeline converts sparse real-estate panoramas into coherent 3D Gaussian Splatting scenes by anchoring optimization with cross-view correspondences from a large reconstruction model, then densifies trajectories with adapted image and video generative priors. The output is Holi360, a dataset of over 1,000 multi-room scenes with dense RGB-D, 3D instance annotations, and 2 million spatial question-answer pairs, plus a benchmark for VLM spatial reasoning.

Active visual recognition fixes both field of view and observation count, yet images vary in the scale and amount of evidence they need. This paper extends the public AdaptiveNN-DeiT-S pipeline with a controllable observation interface that jointly adapts fixation geometry via a learned scale action, local evidence quality via scale-gated foveation and a residual detail-token branch, and observation count via label-grounded confidence-conditioned stop heads. The resulting single checkpoint exposes threshold-controlled accuracy-compute trade-offs without retraining, directly addressing the mismatch between uniform observation budgets and heterogeneous image complexity.

Single-image reconstruction of standalone garments remains under-supported by existing benchmarks, which mostly target clothed humans or rigid objects. MagicGarment introduces a dataset of 1,518 synthetic assets across 17 composite types, paired with aligned multimodal annotations and a unified evaluation protocol. The authors also propose a garment-aware baseline that adapts a frozen Hunyuan3D backbone via input-level semantic refinement (GSR) and a dynamic timestep- and routing-conditioned LoRA (FTR-LoRA), demonstrating that domain-specific adaptation improves over generic 3D priors on this benchmark.

Robota-Embed tackles the problem of retrieving relevant physical experience for robots by learning a single embedding space that jointly encodes vision, language, and action from both the current state and a future trajectory segment. The core insight is that a useful retrieval key must capture not just task intent and visual scene, but also the robot's ongoing motion and execution stage, because the goal is to retrieve an appropriate continuation rather than simply a similar observation. A new multi-source benchmark and two downstream applications—in-context policy conditioning and memory evolution around a frozen large policy—show that this unified, action-aware representation significantly outperforms vision-language-only baselines and improves both success rates and token efficiency.

End-to-end accuracy on multi-image vision-language tasks hides whether failures stem from visual retrieval, schema understanding, or downstream composition. This paper introduces a within-item $2 \times 2$ factorial Oracle design in which a model performs a 16-image schema-application task under four matched conditions, receiving verified textual schema (Oracle-A), verified ordered target appearances (Oracle-B), both, or neither. The results localize the dominant recoverable deficit to upstream ordered instance recovery—Oracle-B rescues strict accuracy by 54.6 percentage points—while supplying the schema alone barely helps, suggesting that current VLMs often lose the relevant visual facts before any codebook reasoning begins.

Industrial anomaly detection is bottlenecked by scarce anomalous samples, which limits supervised training and defect understanding. This paper proposes CORE, a training-free framework that leverages Large Multimodal Models to synthesize realistic anomalies in both zero-shot and one-shot regimes through structured prompt optimization. The key idea is to decompose prompts into BASE, VARIANT, and NEGATIVE components and refine them via a self-supervised contrast-and-reflect mechanism without human annotations or task-specific fine-tuning. If effective, this could remove the need for costly real defect collections and enable open-world adaptation.

Real-time monocular human pose estimation breaks down when the lower body is occluded or the subject steps out of frame. This paper introduces a part-decoupled temporal fusion framework that augments monocular video with sparse plantar-pressure and foot-mounted acceleration from smart insoles. The core insight is that cheap, wearable foot sensors provide complementary physical constraints exactly where monocular vision is weakest, enabling robust pose and gait recovery under severe visual degradation.

LED-BENCH targets a realistic but under-evaluated setting: answering engineering questions from complete, large-format drawings in mechanical, architectural, and transportation domains. The benchmark contains 107 sheets and 377 expert-verified questions that require models to localize fine-grained evidence and integrate information across distant regions or views. Its central finding is that current multimodal models remain far from solving this problem—the best evaluated model achieves only 52.65% accuracy, and performance consistently degrades on harder reasoning tasks.

ActICL attacks the problem of suboptimal multimodal in-context learning by training an MLLM agent to iteratively acquire, inspect, and curate demonstrations through tool-augmented reasoning. Its core idea replaces static similarity retrieval with an agentic loop of search, read, curate, and solve actions, learned via a two-stage post-training pipeline of rejection-sampling SFT and joint reinforcement learning. This matters because it shows that even a modestly sized trained agent can construct contexts that substantially improve its own accuracy and also boost larger frontier answering models.