This research paper introduces a new approach to adapting deep learning models for underwater synthetic aperture sonar Automatic Target Recognition (ATR) using a three-stage framework that leverages Low-Rank Adaptation (LoRA) and Supervised Contrastive Learning (SupCon). The method improves performance by 379% on a mission-level evaluation, with a significant reduction in training data required.

Researchers propose Test-Time Self-Evolving via Reflection (TTSR), a self-evolving framework for adapting large language models during inference using unlabeled test inputs. TTSR improves upon existing methods by addressing bottlenecks in learnable samples and efficient exploration. The framework alternates between a Student and Teacher role, with the Teacher analyzing failed trajectories and synthesizing targeted questions. Experiments show consistent test-time improvements and strong cross-backbone generalization.

Researchers designed Phase Attacks to target the phase information of complex-valued neural networks, which are gaining popularity. They also derived complex-valued versions of common adversarial attacks. The study shows that CVNNs are more robust in some scenarios but still vulnerable to phase changes, highlighting the need for robustness analysis in AI model development.

Researchers propose a method for multi-resolution attribution in adaptive hierarchical systems, such as LLMs, Census, and telecom networks, by leveraging the learned routing state. This approach allows for consistent explanations at various levels of the system without requiring a separate attribution model.

Researchers proposed CleanVideo, a selective erasure framework for text-to-video diffusion models. It uses a tri-modal gating mechanism to control low-dimensional subspace intervention, jointly processing visual features, timestep signals, and textual semantics to erase target concepts while preserving non-target content.

A research paper debates the effectiveness of gradient-based data attribution methods for large language models, concluding that these methods track format similarity more than task semantics. This has implications for data selection and analysis in AI training, and highlights the need for robustness and reliability in these methods.

Researchers introduced a model-agnostic method, TC-LIA, to detect mirage reasoning in vision-language models, a failure mode where VLMs produce confident answers without relevant visual evidence. TC-LIA is an unsupervised technique that tracks question-image alignment across layers of a frozen CLIP encoder, with promising results on state-of-the-art VLMs.

Researchers introduced Generalist-Specialist-Mixture-of-Experts (GS-MoE) architecture for multimodal medical imaging, balancing modality-specific specialization with cross-modal shared representations. GS-MoE performed well on a dataset of 1.35M images, recovering detection of rare pathologies with per-class gains up to +0.60 F1. This may inform the development of AI agents for medical imaging tasks.

Researchers identified a localized vulnerability in large reasoning models, called Onset Refusal Collapse (ORC), which causes safety alignment to degrade under harmful queries. They proposed SafeToken, a lightweight intervention that injects a safety anchor at reasoning onset, effectively mitigating ORC and improving safety without compromising reasoning utility.

Researchers introduce Dual Randomized Smoothing, an improved technique for certifying neural network robustness against adversarial perturbations. This method allows for input-dependent noise variances, breaking through a limitation of the standard Randomized Smoothing formulation. Experiments show significant performance gains on CIFAR-10 and ImageNet datasets.

Researchers developed ProxiDex, a dynamics-guided proximity policy framework for dexterous manipulation, using a hardware-agnostic contact representation and action-conditioned proximity dynamics. This allows for improved success rates and robustness in manipulation tasks under unreliable visual feedback.

A new decoding framework, Early-Bird (EB), is proposed to accelerate diffusion large language models (dLLMs) by adaptively grouping tokens with similar uncertainty into variable-length blocks and learning to unmask tokens in parallel. This approach achieves 3.53-18.76x higher throughput than the vanilla decoding method and is deployable as a plug-in with negligible training and inference overhead.

Researchers have introduced MASA, a novel approach to wild test-time adaptation that uses a frozen multimodal large language model to provide structured semantic descriptions for object families and nuisance factors. This helps to break the self-referential loop in model adaptation and improves accuracy in limited-batch, mixed-domain, and imbalanced-label-shift settings.