Biased news contributes to societal polarization and is often reinforced by hostile reader comments, constituting a vital yet often overlooked aspect of news dissemination. Our study reveals that offensive comments support biased content, amplifying bias and causing harm to targeted groups or individuals. Counterspeech is an effective approach to counter such harmful speech without violating freedom of speech, helping to limit the spread of bias. To the best of our knowledge, this is the first study to explore counterspeech generation in the context of news articles. We introduce a manually annotated dataset linking media bias, offensive comments, and counterspeech. We conduct a detailed analysis showing that over 70\% offensive comments support biased articles, amplifying bias and thus highlighting the importance of counterspeech generation. Comparing counterspeech generated by humans and large language models, we find model-generated responses are more polite but lack the novelty and diversity. Finally, we improve generated counterspeech through few-shot learning and integration of news background information, enhancing both diversity and relevance.

This paper presents our approach and results for SubTask 1: Islamic Inheritance Reasoning at QIAS 2025, a shared task focused on evaluating Large Language Models (LLMs) in understanding and reasoning within Islamic inheritance knowledge. We fine-tuned the Fanar-1-9B causal language model using Low-Rank Adaptation (LoRA) and integrated it into a Retrieval-Augmented Generation (RAG) pipeline. Our system addresses the complexities of Islamic inheritance law, including comprehending inheritance scenarios, identifying eligible heirs, applying fixed-share rules, and performing precise calculations. Our system achieved an accuracy of 0.858 in the final test, outperforming other competitive models such as, GPT 4.5, LLaMA, Fanar, Mistral and ALLaM evaluated with zero-shot prompting. Our results demonstrate that QU-NLP achieves near state-of-the-art accuracy (85.8%), excelling especially on advanced reasoning (97.6%) where it outperforms Gemini 2.5 and OpenAI's o3. This highlights that domain-specific fine-tuning combined with retrieval grounding enables mid-scale Arabic LLMs to surpass frontier models in Islamic inheritance reasoning.

Despite recent advances in large language models (LLMs), most QA benchmarks are still confined to single-paragraph or single-document settings, failing to capture the complexity of real-world information-seeking tasks. Practical QA often requires multi-hop reasoning over information distributed across multiple documents, modalities, and structural formats. Although prior datasets made progress in this area, they rely heavily on Wikipedia-based content and unimodal plain text, with shallow reasoning paths that typically produce brief phrase-level or single-sentence answers, thus limiting their realism and generalizability. We propose DocHop-QA, a large-scale benchmark comprising 11,379 QA instances for multimodal, multi-document, multi-hop question answering. Constructed from publicly available scientific documents sourced from PubMed, DocHop-QA is domain-agnostic and incorporates diverse information formats, including textual passages, tables, and structural layout cues. Unlike existing datasets, DocHop-QA does not rely on explicitly hyper-linked documents; instead, it supports open-ended reasoning through semantic similarity and layout-aware evidence synthesis. To scale realistic QA construction, we designed an LLM-driven pipeline grounded in 11 high-frequency scientific question concepts. We evaluated DocHop-QA through four tasks spanning structured index prediction, generative answering, and multimodal integration, reflecting both discriminative and generative paradigms. These tasks demonstrate DocHop-QA's capacity to support complex, multi-modal reasoning across multiple documents.

Large language models (LLMs) have shown promise in medical question answering but often struggle with hallucinations and shallow reasoning, particularly in tasks requiring nuanced clinical understanding. Retrieval-augmented generation (RAG) offers a practical and privacy-preserving way to enhance LLMs with external medical knowledge. However, most existing approaches rely on surface-level semantic retrieval and lack the structured reasoning needed for clinical decision support. We introduce MedCoT-RAG, a domain-specific framework that combines causal-aware document retrieval with structured chain-of-thought prompting tailored to medical workflows. This design enables models to retrieve evidence aligned with diagnostic logic and generate step-by-step causal reasoning reflective of real-world clinical practice. Experiments on three diverse medical QA benchmarks show that MedCoT-RAG outperforms strong baselines by up to 10.3% over vanilla RAG and 6.4% over advanced domain-adapted methods, improving accuracy, interpretability, and consistency in complex medical tasks.

AI-generated text (AIGT) detection evasion aims to reduce the detection probability of AIGT, helping to identify weaknesses in detectors and enhance their effectiveness and reliability in practical applications. Although existing evasion methods perform well, they suffer from high computational costs and text quality degradation. To address these challenges, we propose Self-Disguise Attack (SDA), a novel approach that enables Large Language Models (LLM) to actively disguise its output, reducing the likelihood of detection by classifiers. The SDA comprises two main components: the adversarial feature extractor and the retrieval-based context examples optimizer. The former generates disguise features that enable LLMs to understand how to produce more human-like text. The latter retrieves the most relevant examples from an external knowledge base as in-context examples, further enhancing the self-disguise ability of LLMs and mitigating the impact of the disguise process on the diversity of the generated text. The SDA directly employs prompts containing disguise features and optimized context examples to guide the LLM in generating detection-resistant text, thereby reducing resource consumption. Experimental results demonstrate that the SDA effectively reduces the average detection accuracy of various AIGT detectors across texts generated by three different LLMs, while maintaining the quality of AIGT.

Recent advances in artificial intelligence (AI), particularly in the domain of large language models (LLMs), have significantly amplified concerns around AI safety and security. One critical aspect of these concerns is the vulnerability of LLMs to backdoor attacks--a malicious strategy whereby an attacker injects specific triggers into training data, resulting in "sleeper agents" that behave normally until activated by particular inputs [6]. These backdoored models (also known as sleeper agents or trojaned models) pose a serious threat as they cannot be detected by standard evaluation methods and manifest undesirable or harmful behaviors only upon exposure to particular triggers in the input [2]. Triggers can take on many forms, ranging from simple single-token lexical triggers to complex semantic triggers [9]. The significance of studying backdoor vulnerabilities arises from two primary threat models: Data-poisoned sleeper agents These involve deliberate poisoning of the training data to trigger specific harmful behaviors under attacker-defined conditions [3]. Real-world implications are substantial; for instance, autonomous vehicles might misinterpret modified road signs, potentially leading to fatal accidents, or software coding assistants might generate insecure code when prompted by certain organisations, making the organisations software systems vulnerable to attack if the generated code is not carefully inspected [3]. Deceptive instrumental alignment Plausibly, models could develop deceptive behaviors organically during training [8]. These models exhibit compliant behaviors in training and evaluation phases but deviate from their developer-defined goals once deployed. While naturally occurring deceptive models have not yet been reported, the training process does select for such behaviour [6].

Fine-tuning Large Language Models (LLMs) with reinforcement learning to produce an explicit Chain-of-Thought (CoT) before answering produces models that consistently raise overall performance on code, math, and general-knowledge benchmarks. However, on benchmarks where LLMs currently achieve low accuracy, such as Humanity's Last Exam (HLE), they often report high self-confidence, reflecting poor calibration. Here, we test whether measurable properties of the CoT provide reliable signals of an LLM's internal confidence in its answers. We analyze three feature classes: (i) CoT length, (ii) intra-CoT sentiment volatility, and (iii) lexicographic hints, including hedging words. Using DeepSeek-R1 and Claude 3.7 Sonnet on both Humanity's Last Exam (HLE), a frontier benchmark with very low accuracy, and Omni-MATH, a saturated benchmark of moderate difficulty, we find that lexical markers of uncertainty (e.g., $\textit{guess}$, $\textit{stuck}$, $\textit{hard}$) in the CoT are the strongest indicators of an incorrect response, while shifts in the CoT sentiment provide a weaker but complementary signal. CoT length is informative only on Omni-MATH, where accuracy is already high ($\approx 70\%$), and carries no signal on the harder HLE ($\approx 9\%$), indicating that CoT length predicts correctness only in the intermediate-difficulty benchmarks, i.e., inside the model's demonstrated capability, but still below saturation. Finally, we find that uncertainty indicators in the CoT are consistently more salient than high-confidence markers, making errors easier to predict than correct responses. Our findings support a lightweight post-hoc calibration signal that complements unreliable self-reported probabilities and supports safer deployment of LLMs.

This review synthesizes the nascent but critical field of developmental interpretability for Large Language Models. We chart the field's evolution from static, post-hoc analysis of trained models to a dynamic investigation of the training process itself. We begin by surveying the foundational methodologies, including representational probing, causal tracing, and circuit analysis, that enable researchers to deconstruct the learning process. The core of this review examines the developmental arc of LLM capabilities, detailing key findings on the formation and composition of computational circuits, the biphasic nature of knowledge acquisition, the transient dynamics of learning strategies like in-context learning, and the phenomenon of emergent abilities as phase transitions in training. We explore illuminating parallels with human cognitive and linguistic development, which provide valuable conceptual frameworks for understanding LLM learning. Finally, we argue that this developmental perspective is not merely an academic exercise but a cornerstone of proactive AI safety, offering a pathway to predict, monitor, and align the processes by which models acquire their capabilities. We conclude by outlining the grand challenges facing the field, such as scalability and automation, and propose a research agenda for building more transparent, reliable, and beneficial AI systems.

Language models are increasingly used in Brazil, but most evaluation remains English-centric. This paper presents Alvorada-Bench, a 4,515-question, text-only benchmark drawn from five Brazilian university entrance examinations. Evaluating twenty models under zero-shot, role-playing, and chain-of-thought prompting, producing 270,900 responses with structured self-reports of confidence, perceived difficulty, and Bloom level. The top models exceed 94% accuracy overall, but accuracy declines on Mathematics and on the engineering oriented IME and ITA exams, indicating persistent weaknesses in multi-step reasoning. Confidence is well calibrated and correlates with perceived difficulty, revealing that models can accurately assess their own certainty capabilities. A cost accuracy analysis shows that high accuracy is achievable at under $2 per 1K tokens. On ENEM 2024 the top model (O3) achieved perfect scores in Languages subject questions while even the weakest system (GPT-4.1 Nano) only underperforms humans in Mathematics. Through exams that distill decades of Brazilian educational priorities and assess millions of students yearly, Alvorada-Bench establishes whether language models can navigate the intersection of language, culture, and reasoning that defines academic readiness in Brazil.

Web agents have shown great promise in performing many tasks on ecommerce website. To assess their capabilities, several benchmarks have been introduced. However, current benchmarks in the e-commerce domain face two major problems. First, they primarily focus on product search tasks (e.g., Find an Apple Watch), failing to capture the broader range of functionalities offered by real-world e-commerce platforms such as Amazon, including account management and gift card operations. Second, existing benchmarks typically evaluate whether the agent completes the user query, but ignore the potential risks involved. In practice, web agents can make unintended changes that negatively impact the user account or status. For instance, an agent might purchase the wrong item, delete a saved address, or incorrectly configure an auto-reload setting. To address these gaps, we propose a new benchmark called Amazon-Bench. To generate user queries that cover a broad range of tasks, we propose a data generation pipeline that leverages webpage content and interactive elements (e.g., buttons, check boxes) to create diverse, functionality-grounded user queries covering tasks such as address management, wish list management, and brand store following. To improve the agent evaluation, we propose an automated evaluation framework that assesses both the performance and the safety of web agents. We systematically evaluate different agents, finding that current agents struggle with complex queries and pose safety risks. These results highlight the need for developing more robust and reliable web agents.