In emergency situations, every second counts. The deployment of Large Language Models (LLMs) in time-sensitive, low or zero-connectivity environments remains limited. Current models are computationally intensive and unsuitable for low-tier devices often used by first responders or civilians. A major barrier to developing lightweight, domain-specific solutions is the lack of high-quality datasets tailored to first aid and emergency response. To address this gap, we introduce FirstAidQA, a synthetic dataset containing 5,500 high-quality question answer pairs that encompass a wide range of first aid and emergency response scenarios. The dataset was generated using a Large Language Model, ChatGPT-4o-mini, with prompt-based in-context learning, using texts from the Vital First Aid Book (2019). We applied preprocessing steps such as text cleaning, contextual chunking, and filtering, followed by human validation to ensure accuracy, safety, and practical relevance of the QA pairs. FirstAidQA is designed to support instruction-tuning and fine-tuning of LLMs and Small Language Models (SLMs), enabling faster, more reliable, and offline-capable systems for emergency settings. We publicly release the dataset to advance research on safety-critical and resource-constrained AI applications in first aid and emergency response. The dataset is available on Hugging Face at https://huggingface.co/datasets/i-am-mushfiq/FirstAidQA.

With the rapid advancement of AI models, their deployment across diverse tasks has become increasingly widespread. A notable emerging application is leveraging AI models to assist in reviewing scientific papers. However, recent reports have revealed that some papers contain hidden, injected prompts designed to manipulate AI reviewers into providing overly favorable evaluations. In this work, we present an early systematic investigation into this emerging threat. We propose two classes of attacks: (1) static attack, which employs a fixed injection prompt, and (2) iterative attack, which optimizes the injection prompt against a simulated reviewer model to maximize its effectiveness. Both attacks achieve striking performance, frequently inducing full evaluation scores when targeting frontier AI reviewers. Furthermore, we show that these attacks are robust across various settings. To counter this threat, we explore a simple detection-based defense. While it substantially reduces the attack success rate, we demonstrate that an adaptive attacker can partially circumvent this defense. Our findings underscore the need for greater attention and rigorous safeguards against prompt-injection threats in AI-assisted peer review.

Foundation models (FMs) have emerged as a transformative paradigm in medical image analysis, offering the potential to provide generalizable, task-agnostic solutions across a wide range of clinical tasks and imaging modalities. Their capacity to learn transferable representations from large-scale data has the potential to address the limitations of conventional task-specific models. However, adaptation of FMs to real-world clinical practice remains constrained by key challenges, including domain shifts, limited availability of high-quality annotated data, substantial computational demands, and strict privacy requirements. This review presents a comprehensive assessment of strategies for adapting FMs to the specific demands of medical imaging. We examine approaches such as supervised fine-tuning, domain-specific pretraining, parameter-efficient fine-tuning, self-supervised learning, hybrid methods, and multimodal or cross-modal frameworks. For each, we evaluate reported performance gains, clinical applicability, and limitations, while identifying trade-offs and unresolved challenges that prior reviews have often overlooked. Beyond these established techniques, we also highlight emerging directions aimed at addressing current gaps. These include continual learning to enable dynamic deployment, federated and privacy-preserving approaches to safeguard sensitive data, hybrid self-supervised learning to enhance data efficiency, data-centric pipelines that combine synthetic generation with human-in-the-loop validation, and systematic benchmarking to assess robust generalization under real-world clinical variability. By outlining these strategies and associated research gaps, this review provides a roadmap for developing adaptive, trustworthy, and clinically integrated FMs capable of meeting the demands of real-world medical imaging.

Speculative decoding (SD) has emerged as an effective technique to accelerate large language model (LLM) inference without compromising output quality. However, the achievable speedup largely depends on the effectiveness of the drafting model. While model-based methods like EAGLE-2 are accurate but costly, retrieval-enhanced methods like SAM-Decoding rely on heuristic switching strategies that often trigger unnecessary retrievals. To address this, we propose ReSpec (\textbf{Re}trieval-enhanced \textbf{Spe}culative Decoding), a novel framework that transforms heuristic drafter switching into adaptive decision-making. ReSpec features three core innovations: 1) An \textbf{entropy-guided adaptive trigger} quantifies contextual predictability to initiate retrieval only when uncertainty is low, avoiding costly low-quality speculations. 2) A \textbf{feedback-driven candidate selection} leverages historical feedback to organize multiple high-quality candidates for parallel verification, maximizing retrieval utility. 3) A source-aware \textbf{relaxed verification strategy} applies strict checks to model-generated drafts while using a relaxed verification for retrieved drafts, achieving a better balance between accuracy and efficiency. Extensive experiments on Spec-Bench demonstrate that ReSpec achieves state-of-the-art acceleration,outperforming EAGLE-2 and SAM-Decoding by over $33\%$ and $25\%$, respectively, while maintaining output quality.

Large language models (LLMs) are reshaping numerous facets of our daily lives, leading widespread adoption as web-based services. Despite their versatility, LLMs face notable challenges, such as generating hallucinated content and lacking access to up-to-date information. Lately, to address such limitations, Retrieval-Augmented Generation (RAG) has emerged as a promising direction by generating responses grounded in external knowledge sources. A typical RAG system consists of i) a retriever that probes a group of relevant passages from a knowledge base and ii) a generator that formulates a response based on the retrieved content. However, as with other AI systems, recent studies demonstrate the vulnerability of RAG, such as knowledge corruption attacks by injecting misleading information. In response, several defense strategies have been proposed, including having LLMs inspect the retrieved passages individually or fine-tuning robust retrievers. While effective, such approaches often come with substantial computational costs. In this work, we introduce RAGDefender, a resource-efficient defense mechanism against knowledge corruption (i.e., by data poisoning) attacks in practical RAG deployments. RAGDefender operates during the post-retrieval phase, leveraging lightweight machine learning techniques to detect and filter out adversarial content without requiring additional model training or inference. Our empirical evaluations show that RAGDefender consistently outperforms existing state-of-the-art defenses across multiple models and adversarial scenarios: e.g., RAGDefender reduces the attack success rate (ASR) against the Gemini model from 0.89 to as low as 0.02, compared to 0.69 for RobustRAG and 0.24 for Discern-and-Answer when adversarial passages outnumber legitimate ones by a factor of four (4x).

Role-play has become a key testbed for generative models, expanding from text-only dialogue to multimodal interaction. Extending role-play to speech captures prosody, emotion, and delivery, but also poses new evaluation challenges. Current pipelines often use audio large language models (ALLMs) as zero-shot judges, which miss paralinguistic cues, collapse multiple aspects into coarse scores, and rely on synthetic speech references that fail to reflect real-world roles. We present Speech-DRAME, a unified framework that contributes at three levels: (i) Speech-DRAME-EvalBench, an evaluation benchmark with bilingual human-annotated data and protocols for training and testing speech evaluation models (SEMs), (ii) DRAME-Eval, a fine-tuned evaluation model, which substantially outperforms zero-shot and few-shot ALLMs, and (iii) Speech-DRAME-RoleBench, a speech role-play benchmark that leverages DRAME-Eval as an automatic judge to compare speech foundation models (SFMs). Speech-DRAME distinguishes between two complementary evaluation strategies: Archetype Evaluation, a top-down approach measuring adherence to broad role archetypes, and Realism Evaluation, a bottom-up approach grounded in real human speech that emphasizes nuanced role quality. Compared to zero-shot ALLM judges, DRAME-Eval achieves stronger agreement with human ratings (Pearson correlation from 0.480 to 0.629 in archetypes, and 0.390 to 0.625 in realism). By integrating transparent benchmark resources, modeling approaches, and system-level evaluation, Speech-DRAME provides the first comprehensive, reproducible foundation for assessing spoken role-play.

Vision-language-action (VLA) models have significantly advanced robotic learning, enabling training on large-scale, cross-embodiment data and fine-tuning for specific robots. However, state-of-the-art autoregressive VLAs struggle with multi-robot collaboration. We introduce embodiment transfer learning, denoted as ET-VLA, a novel framework for efficient and effective transfer of pre-trained VLAs to multi-robot. ET-VLA's core is Synthetic Continued Pretraining (SCP), which uses synthetically generated data to warm up the model for the new embodiment, bypassing the need for real human demonstrations and reducing data collection costs. SCP enables the model to learn correct actions and precise action token numbers. Following SCP, the model is fine-tuned on target embodiment data. To further enhance the model performance on multi-embodiment, we present the Embodied Graph-of-Thought technique, a novel approach that formulates each sub-task as a node, that allows the VLA model to distinguish the functionalities and roles of each embodiment during task execution. Our work considers bimanual robots, a simple version of multi-robot to verify our approaches. We validate the effectiveness of our method on both simulation benchmarks and real robots covering three different bimanual embodiments. In particular, our proposed ET-VLA \space can outperform OpenVLA on six real-world tasks over 53.2%. We will open-source all codes to support the community in advancing VLA models for robot learning.

Abstract--The immense diversity in the culture and culinary of Indian cuisines calls attention to the major shortcoming of the existing Visual Question Answering(VQA) systems which are inclined towards the foods from western regionRecent attempt towards building a VQA dataset for Indian food is a step towards addressing this challenge. However, their approach towards VQA follows a two-step process in which the answer is generated first, followed by the explanation of the expected answer . In this work, we claim that food VQA requires to follow a multi-step reasoning process to arrive at an accurate answer, especially in the context of India food, which involves understanding complex culinary context and identifying relationships between various food items. With this hypothesis we create reasoning chains upon the QA with minimal human intervention. With augmentation of reasoning chains, we observed accuracy improvement of an average 10 percentage points on the baseline. We provide detailed analysis in terms the effect of addition of reasoning chains for the Indian Food VQA task. One of the most important part of culture and social aspects in everyday life is food. In a country like India, food highlights immense diversity based on geography, religion, and traditions of different regions. A single mealcontain items which differ in preparation, presentation and flavor. This richness in the culinary and the culture, poses unique set of challenges for AI systems that target the understanding of content related to Indian food. A powerful framework that has emerged to connect visual and language reasoning is Visual Question Answering(VQA) [6].

The performance of Large Language Models (LLMs) and the associated dollar costs of API calls can fluctuate over time, potentially invalidating conclusions drawn in prior research. To address this, we propose a Fair Evaluation protocol for Test-Time Compute (FEval-TTC), designed to ensure consistent assessment of test-time compute (TTC) methods, regardless of such fluctuations. FEval-TTC focuses on the evaluation of TTC methods that utilize underlying Chains-of-Thought (CoT). It supports evaluations across multiple LLMs on a diverse set of mathematical and commonsense reasoning datasets. The few-shot prompting and answer extraction processes are standardized across datasets, reducing both time and monetary overhead for researchers. Furthermore, we provide a cost modelling procedure that estimates both the token and dollar cost per query, facilitating equitable comparisons of prevalent TTC methods. We open-source FEval-TTC for public use at https://github.com/networkslab/feval_ttc .

Detecting machine-generated text (MGT) has emerged as a critical challenge, driven by the rapid advancement of large language models (LLMs) capable of producing highly realistic, human-like content. However, the performance of current approaches often degrades significantly under domain shift. To address this challenge, we propose a novel framework designed to capture both domain-specific and domain-general MGT patterns through a two-stage Disentangled mixturE-of-ExpeRts (DEER) architecture. First, we introduce a disentangled mixture-of-experts module, in which domain-specific experts learn fine-grained, domain-local distinctions between human and machine-generated text, while shared experts extract transferable, cross-domain features. Second, to mitigate the practical limitation of unavailable domain labels during inference, we design a reinforcement learning-based routing mechanism that dynamically selects the appropriate experts for each input instance, effectively bridging the train-inference gap caused by domain uncertainty. Extensive experiments on five in-domain and five out-of-domain benchmark datasets demonstrate that DEER consistently outperforms state-of-the-art methods, achieving average F1-score improvements of 1.39% and 5.32% on in-domain and out-of-domain datasets respectively, along with accuracy gains of 1.35% and 3.61% respectively. Ablation studies confirm the critical contributions of both disentangled expert specialization and adaptive routing to model performance.