In vision-language modeling, critic models are typically trained to evaluate outputs -- assigning scalar scores or pairwise preferences -- rather than to generate responses. This separation from policy models, which produce the responses, is so entrenched that critics are rarely considered for direct policy use. In this work, we challenge this convention. We propose to reorganize preference-labeled critic datasets into verifiable training signals and perform reinforcement learning directly on a base generative model, producing LLaVA-Critic-R1, a multimodal critic trained to optimize preference judgments while retaining full generation ability. Surprisingly, LLaVA-Critic-R1 emerges not only as a top-performing critic but also as a competitive policy model -- matching or surpassing specialized reasoning VLMs trained with in-domain data across 26 visual reasoning and understanding benchmarks, with an average gain of +5.7% over its base model (Qwen-2.5-VL-7B). Extending this approach to existing strong reasoning VLMs yields LLaVA-Critic-R1+, which further advances policy performance without sacrificing critic quality, achieving a SoTA performance of 71.9 on MMMU at the 7B scale. Finally, we show that the enhanced critic ability benefits inference: applying self-critique at test time yields an average +13.8% improvement on five representative reasoning tasks without additional training. Our results reveal that RL training on critic data can produce a unified model excelling at both evaluation and generation, offering a simple path toward scalable, self-improving multimodal systems.

We investigate the efficacy of Large Language Models (LLMs) in detecting implicit and explicit hate speech, examining whether models with minimal safety alignment (uncensored) might provide more objective classification capabilities compared to their heavily-aligned (censored) counterparts. While uncensored models theoretically offer a less constrained perspective free from moral guardrails that could bias classification decisions, our results reveal a surprising trade-off: censored models significantly outperform their uncensored counterparts in both accuracy and robustness, achieving 78.7% versus 64.1% strict accuracy. However, this enhanced performance comes with its own limitation -- the safety alignment acts as a strong ideological anchor, making censored models resistant to persona-based influence, while uncensored models prove highly malleable to ideological framing. Furthermore, we identify critical failures across all models in understanding nuanced language such as irony. We also find alarming fairness disparities in performance across different targeted groups and systemic overconfidence that renders self-reported certainty unreliable. These findings challenge the notion of LLMs as objective arbiters and highlight the need for more sophisticated auditing frameworks that account for fairness, calibration, and ideological consistency.

Multimodal Large Language Models (MLLMs) have made significant progress in bridging visual perception with high-level textual reasoning. However, they face a fundamental contradiction: while excelling at complex semantic understanding, these models often fail at basic visual tasks that require precise detail perception. This deficiency primarily stems from the prevalent architectural reliance on a single vision encoder optimized for high-level semantic alignment, which inherently sacrifices the ability to capture fine-grained visual information. To address this issue, we introduce Fusion to Enhance (FtZ), a novel vision tower framework. FtZ moves beyond the single-encoder design by innovatively composing a semantically powerful anchor encoder with a perception-rich augmenting encoder via a lightweight Multi-Head Cross-Attention mechanism. Experimental results demonstrate that on several challenging benchmarks demanding fine-grained visual understanding, such as TextVQA, POPE, MMMU, MME and MM-Vet, our FtZ model significantly outperforms baselines that use only a single encoder or existing feature fusion methods. This work proves that composing heterogeneous expert encoders is an efficient and effective path to overcoming the visual perception bottleneck in current MLLMs, offering a new design paradigm for building next-generation AI systems with stronger perceptual capabilities.

-- The human-robot interaction (HRI) field has traditionally used Wizard-of-Oz (WoZ) controlled robots to explore navigation, conversational dynamics, human-in-the-loop interactions, and more to explore appropriate robot behaviors in everyday settings. However, existing WoZ tools are often limited to one context, making them less adaptable across different settings, users, and robotic platforms. T o mitigate these issues, we introduce a Context-Adaptable Robot Interface System (CARIS) that combines advanced robotic capabilities such teleoperation, human perception, human-robot dialogue, and multimodal data recording. Through pilot studies, we demonstrate the potential of CARIS to WoZ control a robot in two contexts: 1) mental health companion and as a 2) tour guide. Furthermore, we identified areas of improvement for CARIS, including smoother integration between movement and communication, clearer functionality separation, recommended prompts, and one-click communication options to enhance the usability wizard control of CARIS. This project offers a publicly available, context-adaptable tool for the HRI community, enabling researchers to streamline data-driven approaches to intelligent robot behavior . The human-robot interaction (HRI) field has long explored intelligent systems that complement human skills in both social and task-oriented contexts [1]. Social robots must seamlessly coordinate perception, conversation, navigation, and other high-level functions to engage naturally with humans. This is a demanding task that requires specialized frameworks, diverse toolsets, and depends on robust communication protocols to enable efficient, real-time exchanges between the robot control system and people.

Text-to-music models capture broad attributes such as instrumentation or mood, but fine-grained stylistic control remains an open challenge. Existing stylization methods typically require retraining or specialized conditioning, which complicates reproducibility and limits policy compliance when artist names are restricted. We study whether lightweight, human-readable modifiers sampled from a large language model can provide a policy-robust alternative for stylistic control. Using MusicGen-small, we evaluate two artists: Billie Eilish (vocal pop) and Ludovico Einaudi (instrumental piano). For each artist, we use fifteen reference excerpts and evaluate matched seeds under three conditions: baseline prompts, artist-name prompts, and five descriptor sets. All prompts are generated using a large language model. Evaluation uses both VGGish and CLAP embeddings with distributional and per-clip similarity measures, including a new min-distance attribution metric. Results show that artist names are the strongest control signal across both artists, while name-free descriptors recover much of this effect. This highlights that existing safeguards such as the restriction of artist names in music generation prompts may not fully prevent style imitation. Cross-artist transfers reduce alignment, showing that descriptors encode targeted stylistic cues. We also present a descriptor table across ten contemporary artists to illustrate the breadth of the tokens. Together these findings define the name-free gap, the controllability difference between artist-name prompts and policy-compliant descriptors, shown through a reproducible evaluation protocol for prompt-level controllability.

The rapid growth of hardware vulnerabilities has created an urgent need for systematic and scalable analysis methods. Unlike software flaws, which are often patchable post-deployment, hardware weaknesses remain embedded across product lifecycles, posing persistent risks to processors, embedded devices, and IoT platforms. Existing efforts such as the MITRE CWE Hardware List (2021) relied on expert-driven Delphi surveys, which lack statistical rigor and introduce subjective bias, while large-scale data-driven foundations for hardware weaknesses have been largely absent. In this work, we propose LLM-HyPZ, an LLM-assisted hybrid framework for zero-shot knowledge extraction and refinement from vulnerability corpora. Our approach integrates zero-shot LLM classification, contextualized embeddings, unsupervised clustering, and prompt-driven summarization to mine hardware-related CVEs at scale. Applying LLM-HyPZ to the 2021-2024 CVE corpus (114,836 entries), we identified 1,742 hardware-related vulnerabilities. We distilled them into five recurring themes, including privilege escalation via firmware and BIOS, memory corruption in mobile and IoT systems, and physical access exploits. Benchmarking across seven LLMs shows that LLaMA 3.3 70B achieves near-perfect classification accuracy (99.5%) on a curated validation set. Beyond methodological contributions, our framework directly supported the MITRE CWE Most Important Hardware Weaknesses (MIHW) 2025 update by narrowing the candidate search space. Specifically, our pipeline surfaced 411 of the 1,026 CVEs used for downstream MIHW analysis, thereby reducing expert workload and accelerating evidence gathering. These results establish LLM-HyPZ as the first data-driven, scalable approach for systematically discovering hardware vulnerabilities, thereby bridging the gap between expert knowledge and real-world vulnerability evidence.

The rapid digital transformation of Fourth Industrial Revolution (4IR) systems is reshaping workforce needs, widening skill gaps, especially for older workers. With growing emphasis on STEM skills such as robotics, automation, artificial intelligence (AI), and security, large-scale re-skilling and up-skilling are required. Training programs must address diverse backgrounds, learning styles, and motivations to improve persistence and success, while ensuring rapid, cost-effective workforce development through experiential learning. To meet these challenges, we present an adaptive tutoring framework that combines generative AI with Retrieval-Augmented Generation (RAG) to deliver personalized training. The framework leverages document hit rate and Mean Reciprocal Rank (MRR) to optimize content for each learner, and is benchmarked against human-generated training for alignment and relevance. We demonstrate the framework in 4IR cybersecurity learning by creating a synthetic QA dataset emulating trainee behavior, while RAG is tuned on curated cybersecurity materials. Evaluation compares its generated training with manually curated queries representing realistic student interactions. Responses are produced using large language models (LLMs) including GPT-3.5 and GPT-4, assessed for faithfulness and content alignment. GPT-4 achieves the best performance with 87% relevancy and 100% alignment. Results show this dual-mode approach enables the adaptive tutor to act as both a personalized topic recommender and content generator, offering a scalable solution for rapid, tailored learning in 4IR education and workforce development.

Generative models have achieved impressive results in text to image tasks, significantly advancing visual content creation. However, this progress comes at a cost, as such models rely heavily on large-scale training data and may unintentionally replicate copyrighted elements, creating serious legal and ethical challenges for real-world deployment. To address these concerns, researchers have proposed various strategies to mitigate copyright risks, most of which are prompt based methods that filter or rewrite user inputs to prevent explicit infringement. While effective in handling obvious cases, these approaches often fall short in more subtle situations, where seemingly benign prompts can still lead to infringing outputs. To address these limitations, this paper introduces Assessing and Mitigating Copyright Risks (AMCR), a comprehensive framework which i) builds upon prompt-based strategies by systematically restructuring risky prompts into safe and non-sensitive forms, ii) detects partial infringements through attention-based similarity analysis, and iii) adaptively mitigates risks during generation to reduce copyright violations without compromising image quality. Extensive experiments validate the effectiveness of AMCR in revealing and mitigating latent copyright risks, offering practical insights and benchmarks for the safer deployment of generative models.

Among the hardest tasks for humans are those found in competitive programming where problems require sophisticated algorithmic thinking, puzzle solving, and the creation of effective code. As a domain to assess language models (LMs), it has not received enough attention, though. This study presents the ICPC benchmark, which consists of 254 international collegiate programming contest (ICPC) tasks. Each problem includes official analysis, reference code, and sample, high-quality unit, and hidden tests. We are able to develop and evaluate a variety of LM inference techniques for competitive programming with these resources. With zero-shot chain-of-thought prompting, we find that o1 only achieves a 19.1\% pass@1 solve rate. With our best inference technique, which combines multi-turn self-judge with reflection and retrieval over episodic information, raises this to 42.2\%. Furthermore, we conduct a new human-in-the-loop investigation to gain a deeper understanding of the remaining difficulties. Surprisingly, we discover that o1 can solve 17 out of 18 problems that were previously unsolvable by any model or technique with just a few specific instructions. A footstep toward LMs with grounded, imaginative, and algorithmic thinking is provided by our quantitative findings and qualitative research. We open-source our code and data at https://github.com/kraritt/zolve.

This paper presents presents three distinct systems developed for the M-DAIGT shared task on detecting AI generated content in news articles and academic abstracts. The systems includes: (1) A fine-tuned RoBERTa-base classifier, (2) A classical TF-IDF + Support Vector Machine (SVM) classifier , and (3) An Innovative ensemble model named Candace, leveraging probabilistic features extracted from multiple Llama-3.2 models processed by a customTransformer encoder.The RoBERTa-based system emerged as the most performant, achieving near-perfect results on both development and test sets.