A research paper introduces TraceSampler, a post-training method to correct endpoint errors in low-step sampling for generative models without modifying the pretrained model. It uses a compact representation of sampling trace to predict and correct errors.

A new model, Elastic V-JEPA (EV-JEPA), is introduced for flexible and collaborative inference across heterogeneous devices. It can be reconfigured to optimize compute budgets and supports cross-device execution. EV-JEPA achieves high accuracy and improves upon previous elastic training approaches.

This paper presents an alternative approach to GUI agents by reformulating world modeling to predict action-induced future information in latent space, improving success rates and reducing inference latency.

Researchers introduce Squint Text, a recognition deficiency in multimodal LLMs that miss embedded text in images due to visual bottlenecks. They propose Squint OPSD, a training method that uses a downscaled view and ground-truth transcription to improve reading performance. The method reaches 88.61 ANLS on a benchmark dataset, significantly narrowing the gap between available visual evidence and its generated read-out.

Researchers propose a method for cross-modal alignment of independently trained models without the need for paired data. They introduce a Wasserstein Procrustes method that estimates a single orthogonal map to align representations across different modalities. The method is shown to be effective across various datasets, models, and modalities, and can even outperform pair-based methods in the very few-pair regime.

This research proposes a framework for predicting the performance of vision-language models (VLMs) based on the capabilities of their large language model (LLM) backbones. The framework, called the Capability-Driven Multimodal Scaling Law, uses a low-dimensional capability score extracted from LLM textual benchmarks to predict VLM performance. The authors demonstrate the effectiveness of the framework by training over 150 VLMs on 34 LLMs and evaluating their performance on various benchmarks.

This paper proposes ResOT, a training-free method to mitigate object hallucination in large vision-language models (LVLMs) at inference time. ResOT uses localized distribution alignment to repair representations by projecting hallucinated directions away from the faithful subspace and aligning the hallucinated distribution with the faithful one. This approach improves image caption quality and multimodal performance across multiple benchmarks.

Researchers propose a new post-training framework for CLIP, ComCLIP, which improves contrastive loss and provides a drop-in replacement for the vision encoder. This framework matches self-distillation baselines and improves transferability of visual features without breaking downstream compatibility.

WorldBench is a benchmark and judge for evaluating LLMs on generating interactive 3D worlds using Three.js. It explores the generated world, controlling its clock and navigating to each biome, and checks the code for what it sees. The judge uses a combination of code and visual checks to evaluate the generated world.

A new training framework, Structured Latent Visual Reasoning (SLVR), is proposed to improve multimodal large language models (MLLMs) on visual reasoning tasks. SLVR organizes reasoning into typed latent stages for planning, grounding, evidence selection, and reasoning integration, and shows consistent improvements across multiple benchmarks.

Memento 3 is a model-based recursive self-improvement system for frozen LLM agents. It enables agents to learn explicit world models through external memory, allowing them to revise hypotheses about environment dynamics and improve their performance.