Reward-based alignment methods for large language models (LLMs) face two key limitations: vulnerability to reward hacking, where models exploit flaws in the reward signal; and reliance on brittle, labor-intensive prompt engineering when LLMs are used as reward models. We introduce Meta Policy Optimization (MPO), a framework that addresses these challenges by integrating a meta-reward model that dynamically refines the reward model's prompt throughout training. In MPO, the meta-reward model monitors the evolving training context and continuously adjusts the reward model's prompt to maintain high alignment, providing an adaptive reward signal that resists exploitation by the policy. This meta-learning approach promotes a more stable policy optimization, and greatly reduces the need for manual reward prompt design. It yields performance on par with or better than models guided by extensively hand-crafted reward prompts. Furthermore, we show that MPO maintains its effectiveness across diverse tasks, from essay writing to mathematical reasoning, without requiring specialized reward designs. Beyond standard RLAIF, MPO's meta-learning formulation is readily extensible to higher-level alignment frameworks. Overall, this method addresses theoretical and practical challenges in reward-based RL alignment for LLMs, paving the way for more robust and adaptable alignment strategies. The code and data can be accessed at: https://github.com/minnesotanlp/mpo
Large-scale pre-trained Vision-Language Models (VLMs) have significantly advanced transfer learning across diverse tasks. However, adapting these models with limited few-shot data often leads to overfitting, undermining their ability to generalize to new tasks. To address this, we propose Multi-Modal Representation Learning (MMRL), which introduces a shared, learnable, modality-agnostic representation space. MMRL generates space tokens projected into both text and image encoders as representation tokens, enabling more effective cross-modal interactions. Unlike prior methods that mainly optimize class token features, MMRL inserts representation tokens into higher encoder layers--where task-specific features are more prominent--while preserving general knowledge in the lower layers. During training, both class and representation features are jointly optimized: a trainable projection layer is applied to representation tokens for task adaptation, while the projection layer for class token remains frozen to retain pre-trained knowledge. To further promote generalization, we introduce a regularization term aligning class and text features with the frozen VLM's zero-shot features. At inference, a decoupling strategy uses both class and representation features for base tasks, but only class features for novel tasks due to their stronger generalization. Building upon this, we propose MMRL++, a parameter-efficient and interaction-aware extension that significantly reduces trainable parameters and enhances intra-modal interactions--particularly across the layers of representation tokens--allowing gradient sharing and instance-specific information to propagate more effectively through the network. Extensive experiments on 15 datasets demonstrate that MMRL and MMRL++ consistently outperform state-of-the-art methods, achieving a strong balance between task-specific adaptation and generalization.
This study critically distinguishes between AI Agents and Agentic AI, offering a structured conceptual taxonomy, application mapping, and challenge analysis to clarify their divergent design philosophies and capabilities. We begin by outlining the search strategy and foundational definitions, characterizing AI Agents as modular systems driven by Large Language Models (LLMs) and Large Image Models (LIMs) for narrow, task-specific automation. Generative AI is positioned as a precursor, with AI Agents advancing through tool integration, prompt engineering, and reasoning enhancements. In contrast, Agentic AI systems represent a paradigmatic shift marked by multi-agent collaboration, dynamic task decomposition, persistent memory, and orchestrated autonomy. Through a sequential evaluation of architectural evolution, operational mechanisms, interaction styles, and autonomy levels, we present a comparative analysis across both paradigms. Application domains such as customer support, scheduling, and data summarization are contrasted with Agentic AI deployments in research automation, robotic coordination, and medical decision support. We further examine unique challenges in each paradigm including hallucination, brittleness, emergent behavior, and coordination failure and propose targeted solutions such as ReAct loops, RAG, orchestration layers, and causal modeling. This work aims to provide a definitive roadmap for developing robust, scalable, and explainable AI agent and Agentic AI-driven systems. >AI Agents, Agent-driven, Vision-Language-Models, Agentic AI Decision Support System, Agentic-AI Applications
Large Language Models (LLMs) are increasingly embedded in autonomous systems and public-facing environments, yet they remain susceptible to jailbreak vulnerabilities that may undermine their security and trustworthiness. Adversarial suffixes are considered to be the current state-of-the-art jailbreak, consistently outperforming simpler methods and frequently succeeding even in black-box settings. Existing defenses rely on access to the internal architecture of models limiting diverse deployment, increase memory and computation footprints dramatically, or can be bypassed with simple prompt engineering methods. We introduce $\textbf{Adversarial Suffix Filtering}$ (ASF), a lightweight novel model-agnostic defensive pipeline designed to protect LLMs against adversarial suffix attacks. ASF functions as an input preprocessor and sanitizer that detects and filters adversarially crafted suffixes in prompts, effectively neutralizing malicious injections. We demonstrate that ASF provides comprehensive defense capabilities across both black-box and white-box attack settings, reducing the attack efficacy of state-of-the-art adversarial suffix generation methods to below 4%, while only minimally affecting the target model's capabilities in non-adversarial scenarios.
Our goal is to understand how post-training methods, such as fine-tuning, alignment, and unlearning, modify language model behavior and representations. We are particularly interested in the brittle nature of these modifications that makes them easy to bypass through prompt engineering or relearning. Recent results suggest that post-training induces shallow context-dependent ``circuits'' that suppress specific response patterns. This could be one explanation for the brittleness of post-training. To test this hypothesis, we design an unlearning algorithm, Layered Unlearning (LU), that creates distinct inhibitory mechanisms for a growing subset of the data. By unlearning the first $i$ folds while retaining the remaining $k - i$ at the $i$th of $k$ stages, LU limits the ability of relearning on a subset of data to recover the full dataset. We evaluate LU through a combination of synthetic and large language model (LLM) experiments. We find that LU improves robustness to adversarial relearning for several different unlearning methods. Our results contribute to the state-of-the-art of machine unlearning and provide insight into the effect of post-training updates.
Vision-language models (VLMs) offer flexible object detection through natural language prompts but suffer from performance variability depending on prompt phrasing. In this paper, we introduce a method for automated prompt refinement using a novel metric called the Contrastive Class Alignment Score (CCAS), which ranks prompts based on their semantic alignment with a target object class while penalizing similarity to confounding classes. Our method generates diverse prompt candidates via a large language model and filters them through CCAS, computed using prompt embeddings from a sentence transformer. We evaluate our approach on challenging object categories, demonstrating that our automatic selection of high-precision prompts improves object detection accuracy without the need for additional model training or labeled data. This scalable and model-agnostic pipeline offers a principled alternative to manual prompt engineering for VLM-based detection systems.
We introduce a method of meta-prompting that jointly produces fluent text for complex tasks while optimizing the similarity of neural states between a human's mental expectation and a Large Language Model's (LLM) neural processing. A technique of agentic reinforcement learning is applied, in which an LLM as a Judge (LLMaaJ) teaches another LLM, through in-context learning, how to produce content by interpreting the intended and unintended generated text traits. To measure human mental beliefs around content production, users modify long form AI-generated text articles before publication at the US Open 2024 tennis Grand Slam. Now, an LLMaaJ can solve the Theory of Mind (ToM) alignment problem by anticipating and including human edits within the creation of text from an LLM. Throughout experimentation and by interpreting the results of a live production system, the expectations of human content reviewers had 100% of alignment with AI 53.8% of the time with an average iteration count of 4.38. The geometric interpretation of content traits such as factualness, novelty, repetitiveness, and relevancy over a Hilbert vector space combines spatial volume (all trait importance) with vertices alignment (individual trait relevance) enabled the LLMaaJ to optimize on Human ToM. This resulted in an increase in content quality by extending the coverage of tennis action. Our work that was deployed at the US Open 2024 has been used across other live events within sports and entertainment.
In the creative practice of text-to-image generation (TTI), images are generated from text prompts. However, TTI models are trained to always yield an output, even if the prompt contains unknown terms. In this case, the model may generate what we call "default images": images that closely resemble each other across many unrelated prompts. We argue studying default images is valuable for designing better solutions for TTI and prompt engineering. In this paper, we provide the first investigation into default images on Midjourney, a popular image generator. We describe our systematic approach to create input prompts triggering default images, and present the results of our initial experiments and several small-scale ablation studies. We also report on a survey study investigating how default images affect user satisfaction. Our work lays the foundation for understanding default images in TTI and highlights challenges and future research directions.
Large language models (LLMs) are inherently vulnerable to unintended privacy breaches. Consequently, systematic red-teaming research is essential for developing robust defense mechanisms. However, current data extraction methods suffer from several limitations: (1) rely on dataset duplicates (addressable via deduplication), (2) depend on prompt engineering (now countered by detection and defense), and (3) rely on random-search adversarial generation. To address these challenges, we propose DMRL, a Data- and Model-aware Reward Learning approach for data extraction. This technique leverages inverse reinforcement learning to extract sensitive data from LLMs. Our method consists of two main components: (1) constructing an introspective reasoning dataset that captures leakage mindsets to guide model behavior, and (2) training reward models with Group Relative Policy Optimization (GRPO), dynamically tuning optimization based on task difficulty at both the data and model levels. Comprehensive experiments across various LLMs demonstrate that DMRL outperforms all baseline methods in data extraction performance.
We present a novel approach to symbolic regression using vision-capable large language models (LLMs) and the ideas behind Google DeepMind's Funsearch. The LLM is given a plot of a univariate function and tasked with proposing an ansatz for that function. The free parameters of the ansatz are fitted using standard numerical optimisers, and a collection of such ans\"atze make up the population of a genetic algorithm. Unlike other symbolic regression techniques, our method does not require the specification of a set of functions to be used in regression, but with appropriate prompt engineering, we can arbitrarily condition the generative step. By using Kolmogorov Arnold Networks (KANs), we demonstrate that ``univariate is all you need'' for symbolic regression, and extend this method to multivariate functions by learning the univariate function on each edge of a trained KAN. The combined expression is then simplified by further processing with a language model.
Over the past three years, the rapid advancement of Large Language Models (LLMs) has had a profound impact on multiple areas of Artificial Intelligence (AI), particularly in Natural Language Processing (NLP) across diverse languages, including Arabic. Although Arabic is considered one of the most widely spoken languages across 27 countries in the Arabic world and used as a second language in some other non-Arabic countries as well, there is still a scarcity of Arabic resources, datasets, and tools. Arabic NLP tasks face various challenges due to the complexities of the Arabic language, including its rich morphology, intricate structure, and diverse writing standards, among other factors. Researchers have been actively addressing these challenges, demonstrating that pre-trained Large Language Models (LLMs) trained on multilingual corpora achieve significant success in various Arabic NLP tasks. This study provides an overview of using large language models (LLMs) for the Arabic language, highlighting early pre-trained Arabic Language models across various NLP applications and their ability to handle diverse Arabic content tasks and dialects. It also provides an overview of how techniques like finetuning and prompt engineering can enhance the performance of these models. Additionally, the study summarizes common Arabic benchmarks and datasets while presenting our observations on the persistent upward trend in the adoption of LLMs.
Graphical User Interface (GUI) agents, driven by Multi-modal Large Language Models (MLLMs), have emerged as a promising paradigm for enabling intelligent interaction with digital systems. This paper provides a structured survey of recent advances in GUI agents, focusing on architectures enhanced by Reinforcement Learning (RL). We first formalize GUI agent tasks as Markov Decision Processes and discuss typical execution environments and evaluation metrics. We then review the modular architecture of (M)LLM-based GUI agents, covering Perception, Planning, and Acting modules, and trace their evolution through representative works. Furthermore, we categorize GUI agent training methodologies into Prompt-based, Supervised Fine-Tuning (SFT)-based, and RL-based approaches, highlighting the progression from simple prompt engineering to dynamic policy learning via RL. Our summary illustrates how recent innovations in multimodal perception, decision reasoning, and adaptive action generation have significantly improved the generalization and robustness of GUI agents in complex real-world environments. We conclude by identifying key challenges and future directions for building more capable and reliable GUI agents.
Electronic Health Records (EHRs) are digital records of patient information, often containing unstructured clinical text. Named Entity Recognition (NER) is essential in EHRs for extracting key medical entities like problems, tests, and treatments to support downstream clinical applications. This paper explores prompt-based medical entity recognition using large language models (LLMs), specifically GPT-4o and DeepSeek-R1, guided by various prompt engineering techniques, including zero-shot, few-shot, and an ensemble approach. Among all strategies, GPT-4o with prompt ensemble achieved the highest classification performance with an F1-score of 0.95 and recall of 0.98, outperforming DeepSeek-R1 on the task. The ensemble method improved reliability by aggregating outputs through embedding-based similarity and majority voting.
Prompt engineering has shown remarkable success with large language models, yet its systematic exploration in computer vision remains limited. In semantic segmentation, both textual and visual prompts offer distinct advantages: textual prompts through open-vocabulary methods allow segmentation of arbitrary categories, while visual reference prompts provide intuitive reference examples. However, existing benchmarks evaluate these modalities in isolation, without direct comparison under identical conditions. We present Show or Tell (SoT), a novel benchmark specifically designed to evaluate both visual and textual prompts for semantic segmentation across 14 datasets spanning 7 diverse domains (common scenes, urban, food, waste, parts, tools, and land-cover). We evaluate 5 open-vocabulary methods and 4 visual reference prompt approaches, adapting the latter to handle multi-class segmentation through a confidence-based mask merging strategy. Our extensive experiments reveal that open-vocabulary methods excel with common concepts easily described by text but struggle with complex domains like tools, while visual reference prompt methods achieve good average results but exhibit high variability depending on the input prompt. Through comprehensive quantitative and qualitative analysis, we identify the strengths and weaknesses of both prompting modalities, providing valuable insights to guide future research in vision foundation models for segmentation tasks.
Integrating structured knowledge from Knowledge Graphs (KGs) into Large Language Models (LLMs) remains a key challenge for symbolic reasoning. Existing methods mainly rely on prompt engineering or fine-tuning, which lose structural fidelity or incur high computational costs. Building on recent encoding techniques which integrate graph embeddings within the LLM input as tokens, we extend this paradigm to the KG domain by leveraging Knowledge Graph Embedding (KGE) models, thus enabling graph-aware reasoning. Our approach is model-agnostic, resource-efficient, and compatible with any LLMs. Extensive experimentation on synthetic and real-world datasets shows that our method improves reasoning performance over established baselines, further achieving the best trade-off in terms of accuracy and efficiency against state-of-the-art LLMs.
Generating high-quality images without prompt engineering expertise remains a challenge for text-to-image (T2I) models, which often misinterpret poorly structured prompts, leading to distortions and misalignments. While humans easily recognize these flaws, metrics like CLIP fail to capture structural inconsistencies, exposing a key limitation in current evaluation methods. To address this, we introduce PromptIQ, an automated framework that refines prompts and assesses image quality using our novel Component-Aware Similarity (CAS) metric, which detects and penalizes structural errors. Unlike conventional methods, PromptIQ iteratively generates and evaluates images until the user is satisfied, eliminating trial-and-error prompt tuning. Our results show that PromptIQ significantly improves generation quality and evaluation accuracy, making T2I models more accessible for users with little to no prompt engineering expertise.
In this work, we enable gamers to share their gaming experience on social media by automatically generating eye-catching highlight reels from their gameplay session Our automation will save time for gamers while increasing audience engagement. We approach the highlight generation problem by first identifying intervals in the video where interesting events occur and then concatenate them. We developed an in-house gameplay event detection dataset containing interesting events annotated by humans using VIA video annotator. Traditional techniques for highlight detection such as game engine integration requires expensive collaboration with game developers. OCR techniques which detect patches of specific images or texts require expensive per game engineering and may not generalize across game UI and different language. We finetuned a multimodal general purpose video understanding model such as X-CLIP using our dataset which generalizes across multiple games in a genre without per game engineering. Prompt engineering was performed to improve the classification performance of this multimodal model. Our evaluation showed that such a finetuned model can detect interesting events in first person shooting games from unseen gameplay footage with more than 90% accuracy. Moreover, our model performed significantly better on low resource games (small dataset) when trained along with high resource games, showing signs of transfer learning. To make the model production ready, we used ONNX libraries to enable cross platform inference. These libraries also provide post training quantization tools to reduce model size and inference time for deployment. ONNX runtime libraries with DirectML backend were used to perform efficient inference on Windows OS. We show that natural language supervision in the X-CLIP model leads to data efficient and highly performant video recognition models.
With the ever-growing adoption of artificial intelligence (AI), AI-based software and its negative impact on the environment are no longer negligible, and studying and mitigating this impact has become a critical area of research. However, it is currently unclear which role environmental sustainability plays during AI adoption in industry and how AI regulations influence Green AI practices and decision-making in industry. We therefore aim to investigate the Green AI perception and management of industry practitioners. To this end, we conducted a total of 11 interviews with participants from 10 different organizations that adopted AI-based software. The interviews explored three main themes: AI adoption, current efforts in mitigating the negative environmental impact of AI, and the influence of the EU AI Act and the Corporate Sustainability Reporting Directive (CSRD). Our findings indicate that 9 of 11 participants prioritized business efficiency during AI adoption, with minimal consideration of environmental sustainability. Monitoring and mitigation of AI's environmental impact were very limited. Only one participant monitored negative environmental effects. Regarding applied mitigation practices, six participants reported no actions, with the others sporadically mentioning techniques like prompt engineering, relying on smaller models, or not overusing AI. Awareness and compliance with the EU AI Act are low, with only one participant reporting on its influence, while the CSRD drove sustainability reporting efforts primarily in larger companies. All in all, our findings reflect a lack of urgency and priority for sustainable AI among these companies. We suggest that current regulations are not very effective, which has implications for policymakers. Additionally, there is a need to raise industry awareness, but also to provide user-friendly techniques and tools for Green AI practices.
As large language models (LLMs) become widely deployed, concerns about their safety and alignment grow. An approach to steer LLM behavior, such as mitigating biases or defending against jailbreaks, is to identify which parts of a prompt influence specific aspects of the model's output. Token-level attribution methods offer a promising solution, but still struggle in text generation, explaining the presence of each token in the output separately, rather than the underlying semantics of the entire LLM response. We introduce ConceptX, a model-agnostic, concept-level explainability method that identifies the concepts, i.e., semantically rich tokens in the prompt, and assigns them importance based on the outputs' semantic similarity. Unlike current token-level methods, ConceptX also offers to preserve context integrity through in-place token replacements and supports flexible explanation goals, e.g., gender bias. ConceptX enables both auditing, by uncovering sources of bias, and steering, by modifying prompts to shift the sentiment or reduce the harmfulness of LLM responses, without requiring retraining. Across three LLMs, ConceptX outperforms token-level methods like TokenSHAP in both faithfulness and human alignment. Steering tasks boost sentiment shift by 0.252 versus 0.131 for random edits and lower attack success rates from 0.463 to 0.242, outperforming attribution and paraphrasing baselines. While prompt engineering and self-explaining methods sometimes yield safer responses, ConceptX offers a transparent and faithful alternative for improving LLM safety and alignment, demonstrating the practical value of attribution-based explainability in guiding LLM behavior.
Vision-language-action (VLA) models represent an important step toward general-purpose robotic systems by integrating visual perception, language understanding, and action execution. However, systematic evaluation of these models, particularly their zero-shot generalization capabilities in out-of-distribution (OOD) environments, remains limited. In this paper, we introduce MultiNet v0.2, a comprehensive benchmark designed to evaluate and analyze the generalization performance of state-of-the-art VLM and VLA models-including GPT-4o, GPT-4.1, OpenVLA,Pi0 Base, and Pi0 FAST-on diverse procedural tasks from the Procgen benchmark. Our analysis reveals several critical insights: (1) all evaluated models exhibit significant limitations in zero-shot generalization to OOD tasks, with performance heavily influenced by factors such as action representation and task complexit; (2) VLAs generally outperform other models due to their robust architectural design; and (3) VLM variants demonstrate substantial improvements when constrained appropriately, highlighting the sensitivity of model performance to precise prompt engineering.