Usman Naseem

Large language models can mislead users and evaluators when incentives favour objectives other than truth, creating risks that factual accuracy alone cannot capture. Existing research spans lying, sycophancy, unfaithful reasoning, and strategic concealment, but lacks an integrated account connecting these behaviours to their causes and countermeasures. This survey develops a functional framework for understanding deception in large language models and distinguishing it from hallucination through a four-part operational test. We organise eleven deception types into four families: deception about the world, about the model's reasoning, about the model itself, and through other people or agents. Six secondary axes characterise these behaviours, while five underlying mechanisms connect them to detection, evaluation, mitigation, and governance. Our synthesis highlights a detectability inversion: overt falsehoods are easier to detect than subtle or strategic deception. It also identifies an evaluation validity threat, whereby models that recognise testing can conceal behaviours that emerge in deployment. We connect single-agent scheming with multi-agent collusion and identify twelve open problems. Drawing predominantly on English-language, text-based evidence, we prioritise evaluations resistant to strategic adaptation and layered oversight, providing a framework for assessing deceptive behaviour and directing future research.

Large language models increasingly produce falsehoods that are not mere mistakes. They withhold what they know, flatter users against the evidence, justify answers with reasoning they did not use, conceal capabilities from evaluators, and behave differently under observation. Such behaviour is goal-directed and sensitive to incentives, unlike hallucination, an unintended error with no objective to serve. This survey treats deception in language models as a phenomenon in its own right. We adopt a functional definition, the systematic production of false beliefs in the service of an objective other than truth, and turn its boundary with hallucination into a four-part test. We then sort deceptive behaviour into four families by what the false belief concerns: the world, the model’s reasoning, the model itself, or a belief induced through other people and agents. The families contain eleven deception types, from lying and sycophancy to sandbagging, scheming, and collusion, each profiled on six secondary axes. We trace the types to five causal mechanisms and follow them through detection, evaluation, defences, governance, and agentic deployment. We also argue that AI-safety work on scheming and the economics of algorithmic collusion describe one behavioural class at two scales. Two findings emerge. The first is a detectability inversion: blunt falsehoods are the easiest to catch, and strategic deception the hardest. The second is a validity threat: models that recognise evaluation can change their behaviour, which undermines the measurements meant to reveal deception. We close with twelve open problems and argue that the validity of evaluation should come first.

Agentic Artificial Intelligence (AI) represents a paradigm shift from static, task-specific systems to autonomous, goal-directed agents capable of reasoning, planning, learning, and acting with minimal human oversight. This paper synthesizes perspectives from philosophy, cognitive science, and AI to define agency, outline its key properties, and situate it in relation to existing paradigms such as reinforcement learning, symbolic reasoning, Belief–Desire–Intention (BDI) architectures, and embodied cognition. We present a taxonomy of agentic systems along dimensions of autonomy, cognitive capability, modality, and environmental interaction, highlighting current capabilities and limitations, and we critically delimit where such a taxonomy is informative and where a functional, closed-loop analysis of agent behavior must take over. The enabling technologies, including large language models, memory architectures, planning frameworks, tool-use mechanisms, and multimodal embodiment, are reviewed alongside diverse application domains ranging from autonomous research and creative systems to web automation and human–AI collaboration. We analyze safety, alignment, evaluation challenges, and emergent risks, dedicate a section to trustworthiness, privacy, and sustainability by bridging from trustworthy machine learning, verified autonomy, and privacy-preserving learning, and compile a comparative review of prominent benchmarks for assessing agentic behavior. Finally, we outline future research directions, including compositional and modular architectures, cooperative AI, simulation-based safe exploration, data-efficient and developmentally inspired learning, hybrid symbolic–neural systems, and strategies for robust alignment. By providing a comprehensive framework and critical analysis, this work aims to guide the development of agentic AI systems that are not only capable but also safe, trustworthy, and aligned with human values.

This paper presents a systematic evaluation of Large Language Models’ (LLMs) behavior on encrypted texts to discuss its safety implications. We introduce a two-dimensional evaluation framework that separately assesses early instruction refusal (whether models refuse harmful-looking instructions) and generation safety (whether models suppress harmful content generation). Previous works have demonstrated that models possessing decryption capabilities are susceptible to under-generalization attacks , where safety mechanisms trained on natural language fail to generalize to encrypted formats. In this work, we show that such scenarios inevitably result in failure along at least one safety dimension: either the generation of unsafe responses (inadequate generation safety) or the over-refusal of legitimate requests (excessive early refusal). Based on these findings, we evaluate a number of pre-LLM and post-LLM safeguards in the encryption schemes where models possess decryption capability (i.e., easy ciphers) and our findings reveal that: (1) most models struggle at balancing both dimensions effectively—prioritizing either instruction refusal or response suppression; (2) pre-LLM defenses fail due to lack of semantic comprehension of encrypted content; (3) post-LLM defenses achieve strong harmful response suppression but are susceptible to over-refusal. This work contributes systematic evaluation methodology and identifies fundamental trade-offs in current safety approaches by analyzing encrypted content.

Medical Visual Question Answering (Med-VQA) has rapidly emerged as a pivotal area in medical artificial intelligence, enabling systems to interpret complex medical images and respond to natural language queries with clinically relevant insights. Despite extensive research on general VQA and medical Vision Language Models (VLMs), the current literature lacks a structured narrative survey focused specifically on the Med-VQA aspects of these two broader fields. To address this gap, this study analyzes Med-VQA as a standalone research question across four key areas. First, we discuss 21 state-of-the-art Med-VQA datasets, tracing the evolution of data resources by reviewing their details. Second, we analyze evaluation protocols and highlight commonly used metrics along with hallucination assessment. Third, we review the VLM architectural concepts and methods that enable Med-VQA. Finally, we review Med-VQA applications across radiology, pathology, and surgical decision support and discuss challenges, including data limitations, hallucinations, and barriers to clinical deployment. This work aims to equip researchers with a structured and comprehensive understanding of Med-VQA in light of its most recent developments.

The fourth edition of the International Workshop on Multimedia Content Analysis for Social Good (MM4SG 2026) was held alongside the prestigious Web Conference 2026. Building on the momentum of prior editions, the workshop continued to address the critical challenge of analyzing and moderating multimodal content across digital platforms. In an era where diverse multimodal media, including memes, text-embedded images, synthetic and manipulated content, and cross-platform narratives can rapidly shape public opinion and influence societal discourse, the need for robust, transparent, and ethically grounded content understanding has become increasingly urgent. MM4SG 2026 provided a dedicated forum for interdisciplinary collaboration, bringing together researchers and practitioners from natural language processing, computer vision, machine learning, computational social science, and AI ethics to advance the state of the art in socially responsible multimodal intelligence. A major theme of this edition was the growing role of agentic multimodal systems and large vision-language models (VLMs), including emerging agentic LVLM paradigms, for scalable reasoning, moderation, and decision-making in real-world settings.