Researchers introduced Latent Frequency Masking, an attack that erases AI-generated image watermarks by manipulating the image's latent representation. The attack preserves image quality and is more efficient than existing methods. This highlights the need for robust watermarking methods and includes latent-frequency manipulation in security evaluations.

MoLE is a new framework for latent visual reasoning in vision-language models, which encourages different latent tokens to extract complementary visual information. It outperforms existing methods on five visual reasoning benchmarks, achieving an average score of 78.6.

This paper examines the behavior of answer candidates and rationales in masked diffusion MLLMs, finding that answers can stabilize before rationales unfold. The study analyzes three visual question-answering benchmarks, revealing differences in answer coverage and observation windows. The findings have implications for the development and optimization of MLLMs.

Researchers introduce ComputerSD, an online self-distillation method for computer-use agents (CUAs) that learns from real-time feedback from GUI transitions. ComputerSD outperforms existing methods by 1.9-4.1 percentage points on two benchmark backbones.

This paper explores the impact of masking-based token pruning on the robustness of CLIP models. It proposes a pre-deployment diagnostic called the Spurious Inversion Metric (SIM) to predict whether masking helps or hurts worst-group robustness. The study finds that masking can have a significant impact on model performance and introduces a batched GPU segmentation routine to mitigate its drawbacks.

This paper introduces ShieldCLIP, a framework for selective safety alignment in multimodal foundation models like CLIP. It conditions safety alignment on the observed safety state of each modality, preserving safe content and redirecting only unsafe content. The authors also introduce ViSUv2, a 195k-quadruplet dataset with independent per-modality safety labels. ShieldCLIP achieves consistent reductions in harmful outputs in various tasks and settings.

A research paper proposes a novel backdoor defense framework for text-to-image (T2I) diffusion models, called Normal Diffusion Dynamics Learning (NDDL), which detects backdoors by learning normal transition dynamics of diffusion trajectories and exploiting deviations in predicted transitions.

Researchers introduce CRAFT, a method for localizing and mitigating failures in medical vision language models. CRAFT identifies two failure modes: arbitration failure and brake failure, and provides a way to excise problematic attention heads to improve model performance.

FedSocket is a recipient-executable knowledge exchange framework for heterogeneous multimodal federated learning. It enables recipient models to execute exchanged knowledge directly, improving accuracy and efficiency. FedSocket combines local and exchanged predictions to achieve better results than independent ensembles or local models alone.

A research paper introduces ENCORE, a system that enables coding agents to discover manipulation strategies using a few demonstrations. ENCORE improves upon prior agentic systems, achieving higher success rates in various tasks, including cube handover and cup inversion.

ThinkingGuard is a guard model for identifying implicit hazards in Multimodal Large Language Models (MLLMs). It uses a step-supervised structured reasoning framework and a step-reward Monte Carlo Tree Search algorithm to detect risks in MLLMs. The authors propose a new dataset, TriggerBench, to train and evaluate ThinkingGuard. The system demonstrates strong performance on both standard and implicit safety benchmarks.

Researchers introduced Video-RSI, a framework for recursive self-improvement in video understanding agents. The framework enables agents to revise their own harness using their language model, leading to improved accuracy and efficiency. The approach involves revisiting original training videos, testing competing failure explanations, and determining cost-effective revisions.