Multilingual Large Language Models(MLLMs) demonstrate strong generalization across languages, yet they remain prone to hallucinations, especially in low-resource languages, due to training data imbalances. These hallucinations, which include inaccurate or fabricated outputs, are particularly problematic in domain-specific generation tasks (Chataigner et al., 2024). To address this challenge, we propose CCL-XCoT(Curriculum-based Contrastive Learning-based Cross-lingual Chain-of-Thought), a two-stage fine-tuning framework for mitigating hallucination in MLLMs. Our approach first enhances cross-lingual semantic alignment through curriculum-based contrastive learning combined with next-token prediction during continued pre-training. Building on this foundation, we then introduce a cross-lingual Chain-of-Thought (XCoT) prompting strategy during instruction fine-tuning, which guides the model to reason in a high-resource language before generating answers in the target low-resource language. Experimental results show that CCL-XCoT reduces hallucination rates by up to 62% and substantially improves factual knowledge transfer across language pairs, without relying on external retrieval or multi-model ensembles.

Large language models (LLMs) are increasingly being used in user-facing applications, from providing medical consultations to job interview advice. Recent research suggests that these models are becoming increasingly proficient at inferring identity information about the author of a piece of text from linguistic patterns as subtle as the choice of a few words. However, little is known about how LLMs use this information in their decision-making in real-world applications. We perform the first comprehensive analysis of how identity markers present in a user's writing bias LLM responses across five different high-stakes LLM applications in the domains of medicine, law, politics, government benefits, and job salaries. We find that LLMs are extremely sensitive to markers of identity in user queries and that race, gender, and age consistently influence LLM responses in these applications. For instance, when providing medical advice, we find that models apply different standards of care to individuals of different ethnicities for the same symptoms; we find that LLMs are more likely to alter answers to align with a conservative (liberal) political worldview when asked factual questions by older (younger) individuals; and that LLMs recommend lower salaries for non-White job applicants and higher salaries for women compared to men. Taken together, these biases mean that the use of off-the-shelf LLMs for these applications may cause harmful differences in medical care, foster wage gaps, and create different political factual realities for people of different identities. Beyond providing an analysis, we also provide new tools for evaluating how subtle encoding of identity in users' language choices impacts model decisions. Given the serious implications of these findings, we recommend that similar thorough assessments of LLM use in user-facing applications are conducted before future deployment.

The automated summarisation of parliamentary debates using large language models (LLMs) offers a promising way to make complex legislative discourse more accessible to the public. However, such summaries must not only be accurate and concise but also equitably represent the views and contributions of all speakers. This paper explores the use of LLMs to summarise plenary debates from the European Parliament and investigates the algorithmic and representational biases that emerge in this context. We propose a structured, multi-stage summarisation framework that improves textual coherence and content fidelity, while enabling the systematic analysis of how speaker attributes -- such as speaking order or political affiliation -- influence the visibility and accuracy of their contributions in the final summaries. Through our experiments using both proprietary and open-weight LLMs, we find evidence of consistent positional and partisan biases, with certain speakers systematically under-represented or misattributed. Our analysis shows that these biases vary by model and summarisation strategy, with hierarchical approaches offering the greatest potential to reduce disparity. These findings underscore the need for domain-sensitive evaluation metrics and ethical oversight in the deployment of LLMs for democratic applications.

In modern times, people have numerous online accounts, but they rarely read the Terms of Service or Privacy Policy of those sites despite claiming otherwise. This paper introduces PoliAnalyzer, a neuro-symbolic system that assists users with personalized privacy policy analysis. PoliAnalyzer uses Natural Language Processing (NLP) to extract formal representations of data usage practices from policy texts. In favor of deterministic, logical inference is applied to compare user preferences with the formal privacy policy representation and produce a compliance report. To achieve this, we extend an existing formal Data Terms of Use policy language to model privacy policies as app policies and user preferences as data policies. In our evaluation using our enriched PolicyIE dataset curated by legal experts, PoliAnalyzer demonstrated high accuracy in identifying relevant data usage practices, achieving F1-score of 90-100% across most tasks. Additionally, we demonstrate how PoliAnalyzer can model diverse user data-sharing preferences, derived from prior research as 23 user profiles, and perform compliance analysis against the top 100 most-visited websites. This analysis revealed that, on average, 95.2% of a privacy policy's segments do not conflict with the analyzed user preferences, enabling users to concentrate on understanding the 4.8% (636 / 13205) that violates preferences, significantly reducing cognitive burden. Further, we identified common practices in privacy policies that violate user expectations - such as the sharing of location data with 3rd parties. This paper demonstrates that PoliAnalyzer can support automated personalized privacy policy analysis at scale using off-the-shelf NLP tools. This sheds light on a pathway to help individuals regain control over their data and encourage societal discussions on platform data practices to promote a fairer power dynamic.

The integration of Large Language Models (LLMs) in K--12 education offers both transformative opportunities and emerging risks. This study explores how students may Trojanize prompts to elicit unsafe or unintended outputs from LLMs, bypassing established content moderation systems with safety guardrils. Through a systematic experiment involving simulated K--12 queries and multi-turn dialogues, we expose key vulnerabilities in GPT-3.5 and GPT-4. This paper presents our experimental design, detailed findings, and a prototype tool, TrojanPromptGuard (TPG), to automatically detect and mitigate Trojanized educational prompts. These insights aim to inform both AI safety researchers and educational technologists on the safe deployment of LLMs for educators.

Recent advancements in Large Language Models (LLMs) have spurred interest in numerous applications requiring robust long-range capabilities, essential for processing extensive input contexts and continuously generating extended outputs. As sequence lengths increase, the number of Key-Value (KV) pairs in LLMs escalates, creating a significant efficiency bottleneck. In this paper, we propose a new KV cache optimization paradigm called LaCache, a training-free method for efficient and accurate generative inference of LLMs. LaCache enables LLMs to simultaneously address both of the critical challenges in long-range modeling: robust long-range capabilities and continuous generation without running out-of-memory (OOM). Specifically, LaCache integrates two key innovations: (1) a ladder-shaped KV cache pattern that stores KV pairs not only sequentially (left-to-right within each layer) but also across layers (from shallow to deep), providing an extended span for capturing long-range dependencies under a fixed storage budget, thereby boosting long-range capabilities; and (2) an iterative compaction mechanism that progressively compresses older caches, freeing up space for new tokens within a fixed cache size. This token distance-based dynamic compression enables more effective continuous generation under constrained cache budgets. Experiments across various tasks, benchmarks, and LLM models consistently validate LaCache's effectiveness in enhancing LLMs' long-range capabilities. Our code is available at https://github.com/GATECH-EIC/LaCache.

Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse applications, yet they pose significant security risks that threaten their safe deployment in critical domains. Current security alignment methodologies predominantly rely on Process Reward Models (PRMs) to evaluate intermediate reasoning steps, introducing substantial computational overhead and scalability constraints. This paper presents a novel PRM-free security alignment framework that leverages automated red teaming and adversarial training to achieve robust security guarantees while maintaining computational efficiency. Our approach systematically identifies vulnerabilities through sophisticated attack strategies including genetic algorithm optimization, multi-agent simulation, and advanced prompt mutation techniques. The framework enhances model robustness via targeted adversarial training with curriculum learning and adaptive regularization mechanisms. Comprehensive experimental evaluation across five state-of-the-art LLMs demonstrates that our method achieves superior security alignment performance compared to PRM-based approaches while reducing computational costs by 61\%. The framework incorporates transparent reporting and continuous audit mechanisms that enable iterative security improvement and regulatory compliance. Our contributions advance the field of efficient LLM security alignment by democratizing access to robust security measures for resource-constrained organizations and providing a scalable foundation for addressing evolving adversarial threats.

Large Language Models (LLMs) exhibit significant activation sparsity, where only a subset of neurons are active for a given input. Although this sparsity presents opportunities to reduce computational cost, efficiently utilizing it requires predicting activation patterns in a scalable manner. However, direct prediction at the neuron level is computationally expensive due to the vast number of neurons in modern LLMs. To enable efficient prediction and utilization of activation sparsity, we propose a clustering-based activation pattern compression framework. Instead of treating each neuron independently, we group similar activation patterns into a small set of representative clusters. Our method achieves up to 79.34% clustering precision, outperforming standard binary clustering approaches while maintaining minimal degradation in perplexity (PPL) scores. With a sufficiently large number of clusters, our approach attains a PPL score as low as 12.49, demonstrating its effectiveness in preserving model quality while reducing computational overhead. By predicting cluster assignments rather than individual neuron states, future models can efficiently infer activation patterns from pre-computed centroids. We detail the clustering algorithm, analyze its effectiveness in capturing meaningful activation structures, and demonstrate its potential to improve sparse computation efficiency. This clustering-based formulation serves as a foundation for future work on activation pattern prediction, paving the way for efficient inference in large-scale language models.

Although existing models can interact with humans and provide satisfactory responses, they lack the ability to act autonomously or engage in independent reasoning. Furthermore, input data in these models is typically provided as explicit queries, even when some sensory data is already acquired. In addition, AI agents, which are computational entities designed to perform tasks and make decisions autonomously based on their programming, data inputs, and learned knowledge, have shown significant progress. However, they struggle with integrating knowledge across multiple domains, unlike humans. Mental imagery plays a fundamental role in the brain's thinking process, which involves performing tasks based on internal multisensory data, planned actions, needs, and reasoning capabilities. In this paper, we investigate how to integrate mental imagery into a machine thinking framework and how this could be beneficial in initiating the thinking process. Our proposed machine thinking framework integrates a Cognitive thinking unit supported by three auxiliary units: the Input Data Unit, the Needs Unit, and the Mental Imagery Unit. Within this framework, data is represented as natural language sentences or drawn sketches, serving both informative and decision-making purposes. We conducted validation tests for this framework, and the results are presented and discussed.

The demand for machine intelligence capable of processing continuous, long-context inputs on local devices is growing rapidly. However, the quadratic complexity and memory requirements of traditional Transformer architectures make them inefficient and often unusable for these tasks. This has spurred a paradigm shift towards new architectures like State Space Models (SSMs) and hybrids, which promise near-linear scaling. While most current research focuses on the accuracy and theoretical throughput of these models, a systematic performance characterization on practical consumer hardware is critically needed to guide system-level optimization and unlock new applications. To address this gap, we present a comprehensive, comparative benchmarking of carefully selected Transformer, SSM, and hybrid models specifically for long-context inference on consumer and embedded GPUs. Our analysis reveals that SSMs are not only viable but superior for this domain, capable of processing sequences up to 220K tokens on a 24GB consumer GPU-approximately 4x longer than comparable Transformers. While Transformers may be up to 1.8x faster at short sequences, SSMs demonstrate a dramatic performance inversion, becoming up to 4x faster at very long contexts (~57K tokens). Our operator-level analysis reveals that custom, hardware-aware SSM kernels dominate the inference runtime, accounting for over 55% of latency on edge platforms, identifying them as a primary target for future hardware acceleration. We also provide detailed, device-specific characterization results to guide system co-design for the edge. To foster further research, we will open-source our characterization framework.