WorldCast is a benchmark that tests multimodal generative models on multi-step world state prediction, requiring them to forecast successive future states from multimodal observations and diverse references. The study evaluates 12 models across 320 manually designed cases in eight categories, revealing that advanced models like GPT-Image-2.5 achieve high scores while open-source models lag behind, especially on complex context and reasoning tasks. The results underscore the importance of integrating dynamic planning, reasoning, and unified multimodal capabilities for realistic world simulation.
By Wenhao Yan, Haodong Li, Xiaotong Li, Yaxin Luo, Zhengrong Yue, Jialv Zou, Xu Huang, Haojin Yang, Nan Jia, Yan Feng, Peng Pei, Xunliang Cai
The paper introduces a benchmark of 400 synthetic healthcare workflow episodes to evaluate large language model (LLM) agents’ reliability in escalating care when evidence is missing or conflicting. It finds that while clean decision accuracy is high (0.97), models miss a large proportion of necessary escalations (32–97%) under uncertainty, and even the largest 9.7B model’s performance drops to 0.78 in conflict scenarios. A workflow-aware controller improves utility and reduces missed escalations, but its effectiveness is limited by lexical generalization failures.
By Zihao Zheng (Washington University in St. Louis), Junyi Yao (Washington University in St. Louis), Baichuan Li (Southern Methodist University), Jiayu Long (Washington University in St. Louis)
Medical vision-language models (VLMs) for visual question answering (VQA) often pass long visual-token sequences to the language model, making visual-token pruning attractive but risky for medical ima...
As assistants move from reactive answering toward proactive intervention, memory changes role: it no longer only supplies the answer to a query, it determines whether and when to speak at all. Prior w...
Retrieval is a fundamental task in multimodal systems: given an image or text query, the goal is to find the most relevant images or texts in a collection, often using a shared embedding space. Reliab...
Gen4P is a unified framework that transforms a pretrained video generator into a feed‑forward 4D perceiver, enabling recovery of 4D geometry and motion from monocular video without iterative denoising. It processes clean video latents in a single encoder pass and uses a lightweight cross‑attention decoder to predict requested 4D attributes, supporting flexible spatial sampling and continuous‑time queries for temporal flexibility. By incorporating a frozen pretrained video‑depth LoRA as a geometry prior, Gen4P improves performance on 3D point tracking, video depth, and camera pose estimation, achieving state‑of‑the‑art or competitive results on these tasks and demonstrating robust tracking of complex motion in in‑the‑wild videos.
PathFM is a trajectory-based forecasting model that uses flow matching to generate complete future paths for multivariate time series. By distilling the model to one or two steps and sharing a history attention cache, it can produce multiple paths in the time of a single quantile query and optimize calibration loss directly. The paper demonstrates competitive marginal accuracy on fev-bench and TIME, and superior aggregate path scores on the new fps-bench, while maintaining long‑run variability in chaotic systems.
AVC-Reward introduces an event‑centric framework for evaluating audio‑video consistency in joint generation models. It separates audio and visual events, assesses them independently, cross‑validates cross‑modal relations, and aggregates results into a three‑level metric covering global narrative harmony, event‑level synchronization, and fine‑grained attribute alignment. The resulting AVC‑Reward model aligns closely with human judgments, offers interpretable failure localization, and enhances consistency when used as a reinforcement‑learning signal.
EOS-Story is a training‑free framework that improves visual storytelling by balancing identity consistency with scene‑specific adaptation. It separates narrative prompts into identity and scene sub‑prompts, then uses Cross‑Contextual Bridging to inject appearance cues into the scene embedding. In the vision space, Adaptive Identity Preservation aligns identity features to the scene scale and employs Identity‑Aware Self‑Attention to avoid structural artifacts, achieving strong results on the Consistory+_+ benchmark.
ARC‑SORT is a real‑time, training‑free multi‑object tracker designed for fast, direction‑changing athletes and dynamic camera motion in sports broadcasts. It uses an Interacting Multiple Model Kalman filter to adaptively size search regions based on per‑track uncertainty, and applies a single corrective identity swap after track crossings when appearance evidence indicates a mistake. The method achieves state‑of‑the‑art HOTA scores on SportsMOT and DanceTrack while running at ~50 fps on a laptop GPU.
DropKV introduces a new key-value cache eviction strategy for large language models that decouples the joint eviction decision into independent per-token scoring, thereby avoiding the NP-hard combinatorial problem of minimizing output perturbation. The method offers a provable constant-factor approximation guarantee under a cone condition verified on three long-context LLMs, and its high-performance implementation using fused Triton kernels achieves a 19.8× speedup in scoring and up to 9.9% end-to-end prefill speedup. Evaluations on RULER and LongBench show that DropKV outperforms state-of-the-art baselines at equivalent cache budgets, delivering the lowest inference latency among eviction methods while remaining faster than the dynamic cache baseline.
Existing feature attribution methods that account for feature dependencies do not scale to high dimensions. We propose causal integrated gradients (CIG) and causal expected gradients (CEG), which exte...
Multimodal large language models (MLLMs) can act unsafely when individually benign images and texts are interpreted jointly. Beyond implicit reasoning risk, where reasoning drifts toward harm, we iden...
Given a state and a set of options, such as intents, tools or interface elements, System One decision models return a distribution over the options in one forward pass, without generated reasoning. We...
Discrete image generators learn the distribution of tokens produced by an image tokenizer. These token representations should be easy for generators to model, not merely enable faithful image reconstr...
Sheaf neural networks generalize graph neural networks by assigning each edge a learned linear restriction map that specifies how the local representations of adjacent nodes relate to one another, the...
Grid foundation models (GridFMs) are graph transformers pretrained on a masked power-flow objective, and evaluate power flow orders of magnitude faster than an iterative solver. Their size, however, t...
A fixed discount factor in Q-learning imposes a single planning horizon throughout training, despite the differing demands of early exploration and long-term credit assignment. We propose Q($γ_H$), a...
Introducing Timestral offers state‑of‑the‑art probabilistic forecasting across four model sizes ranging from 5 million to 209 million parameters. The models are designed to be CPU‑efficient and ready for zero‑shot use, enabling rapid deployment without extensive fine‑tuning.
Vision Transformer architectures have showcased impressive performance on semantic segmentation, yet their quadratic attention complexity hinders practical applications. Recent works have focused on m...