Breaking the Safety-Capability Tradeoff: Reinforcement Learning with Verifiable Rewards Maintains Safety Guardrails in LLMs

Restora-Flow has been introduced as a training-free method for image restoration that utilizes flow matching sampling guided by a degradation mask. This innovative approach aims to enhance the quality of image restoration tasks such as inpainting, super-resolution, and denoising while addressing the long processing times and over-smoothing issues faced by existing methods.

RobustMerge has been introduced as a parameter-efficient model merging method designed for multi-task learning in machine learning language models (MLLMs), emphasizing direction robustness during the merging process. This approach addresses the challenges of merging expert models without data leakage, which has become increasingly important as model sizes and data complexity grow.

The recent introduction of EmoFeedback$^2$ aims to enhance continuous emotional image generation (C-EICG) by utilizing a large vision-language model (LVLM) to provide reward and textual feedback, addressing the limitations of existing methods that struggle with emotional continuity and fidelity. This paradigm allows for better alignment of generated images with user emotional descriptions.

BengaliFig has been introduced as a new challenge set aimed at evaluating figurative and culturally grounded reasoning in Bengali, a language that is considered low-resource. The dataset comprises 435 unique riddles from Bengali traditions, annotated across five dimensions to assess reasoning types and cultural depth, and is designed for use with large language models (LLMs).

The introduction of DesignPref marks a significant advancement in the field of visual design generation, presenting a dataset of 12,000 pairwise comparisons of UI designs rated by 20 professional designers. This dataset highlights the subjective nature of design preferences, revealing substantial disagreement among trained designers, as indicated by a Krippendorff's alpha of 0.25 for binary preferences.

Gram2Vec is introduced as a grammatical style embedding system that transforms documents into a higher dimensional space by analyzing the normalized relative frequencies of grammatical features in the text. This method offers inherent interpretability compared to traditional neural approaches, with applications demonstrated in authorship verification and AI detection.

A systematic analysis of test-time thinking in large vision-language models (LVLMs) has been conducted, revealing that generating explicit intermediate reasoning chains can enhance performance, but excessive thinking may lead to incorrect outcomes. The study evaluated ten variants from the InternVL3.5 and Qwen3-VL families on the MMMU-val dataset, highlighting the importance of short lookback phrases that refer back to the image for successful visual reasoning.

A new framework has been proposed to quantify modality contributions in multimodal models by utilizing Partial Information Decomposition (PID). This approach addresses the limitations of existing methods that conflate contribution with performance metrics, particularly in cross-attention architectures where modalities interact. The algorithm developed enables scalable, inference-only analysis of predictive information in internal embeddings.