Most evaluations of large language models focus on standard tasks such as factual question answering or short summarization. This research expands that scope in two directions: first, by comparing two retrieval strategies, Graph RAG, structured knowledge-graph based, and Advanced RAG, hybrid keyword-semantic search, for QA; and second, by evaluating whether LLMs can generate high quality non-traditional academic outputs, specifically slide decks and podcast scripts. We implemented a prototype combining Meta's LLaMA 3 70B open weight and OpenAI's GPT 4o mini API based. QA performance was evaluated using both human ratings across eleven quality dimensions and large language model judges for scalable cross validation. GPT 4o mini with Advanced RAG produced the most accurate responses. Graph RAG offered limited improvements and led to more hallucinations, partly due to its structural complexity and manual setup. Slide and podcast generation was tested with document grounded retrieval. GPT 4o mini again performed best, though LLaMA 3 showed promise in narrative coherence. Human reviewers were crucial for detecting layout and stylistic flaws, highlighting the need for combined human LLM evaluation in assessing emerging academic outputs.

The rapid advancement of humanoid robotics has intensified the need for robust and adaptable controllers to enable stable and efficient locomotion across diverse platforms. However, developing such controllers remains a significant challenge because existing solutions are tailored to specific robot designs, requiring extensive tuning of reward functions, physical parameters, and training hyperparameters for each embodiment. To address this challenge, we introduce H-Zero, a cross-humanoid locomotion pretraining pipeline that learns a generalizable humanoid base policy. We show that pretraining on a limited set of embodiments enables zero-shot and few-shot transfer to novel humanoid robots with minimal fine-tuning. Evaluations show that the pretrained policy maintains up to 81% of the full episode duration on unseen robots in simulation while enabling few-shot transfer to unseen humanoids and upright quadrupeds within 30 minutes of fine-tuning.

Indirect prompt injection attacks (IPIAs), where large language models (LLMs) follow malicious instructions hidden in input data, pose a critical threat to LLMpowered agents. In this paper, we present IntentGuard, a general defense framework based on instruction-following intent analysis. The key insight of Intent-Guard is that the decisive factor in IPIAs is not the presence of malicious text, but whether the LLM intends to follow instructions from untrusted data. Building on this insight, IntentGuard leverages an instruction-following intent analyzer (IIA) to identify which parts of the input prompt the model recognizes as actionable instructions, and then flag or neutralize any overlaps with untrusted data segments. To instantiate the framework, we develop an IIA that uses three "thinking intervention" strategies to elicit a structured list of intended instructions from reasoning-enabled LLMs. These techniques include start-of-thinking prefilling, end-of-thinking refinement, and adversarial in-context demonstration. We evaluate IntentGuard on two agentic benchmarks (AgentDojo and Mind2Web) using two reasoning-enabled LLMs (Qwen-3-32B and gpt-oss-20B). Results demonstrate that IntentGuard achieves (1) no utility degradation in all but one setting and (2) strong robustness against adaptive prompt injection attacks (e.g., reducing attack success rates from 100% to 8.5% in a Mind2Web scenario). Indirect prompt injection attacks (IPIAs) (Greshake et al., 2023), where large language models (LLMs) follow malicious instructions hidden in the input data, have emerged as a top security concern for LLM-powered agents. Although many defenses have been proposed, each faces fundamental limitations. Finetuning-based defenses (Chen et al., 2024; 2025b) are costly and lack interpretability; auxiliary classifiers for IPIA detection Shi et al. (2025); Hung et al. (2024) often fail to generalize and are vulnerable to adaptive attacks; system-level rule enforcement Debenedetti et al. (2025) can impact agent utility while offering little robustness against attacks that do not alter control and data flows (e.g., injecting misinformation or phishing links into an email summary). In this paper, we approach the prompt injection problem from a new perspective: instruction-following intent analysis. For an LLM to effectively follow instructions, it must have an internal mechanism to decide which parts of a prompt it recognizes as actionable instructions.

Organisms constantly pivot between tasks such as evading predators, foraging, traversing rugged terrain, and socializing, often within milliseconds. Remarkably, they preserve knowledge of once-learned environments sans catastrophic forgetting, a phenomenon neuroscientists hypothesize, is due to a singular neural circuitry dynamically overlayed by neuromodulatory agents such as dopamine and acetylcholine. In parallel, deep learning research addresses analogous challenges via domain generalization (DG) and continual learning (CL), yet these methods remain siloed, despite the brains ability to perform them seamlessly. In particular, prior work has not explored architectures involving associative memories (AMs), which are an integral part of biological systems, to jointly address these tasks. We propose Memory-Integrated Reconfigurable Adapters (MIRA), a unified framework that integrates Hopfield-style associative memory modules atop a shared backbone. Associative memory keys are learned post-hoc to index and retrieve an affine combination of stored adapter updates for any given task or domain on a per-sample basis. By varying only the task-specific objectives, we demonstrate that MIRA seamlessly accommodates domain shifts and sequential task exposures under one roof. Empirical evaluations on standard benchmarks confirm that our AM-augmented architecture significantly enhances adaptability and retention: in DG, MIRA achieves SoTA out-of-distribution accuracy, and in incremental learning settings, it outperforms architectures explicitly designed to handle catastrophic forgetting using generic CL algorithms. By unifying adapter-based modulation with biologically inspired associative memory, MIRA delivers rapid task switching and enduring knowledge retention in a single extensible architecture, charting a path toward more versatile and memory-augmented AI systems.

Transformer models have significantly advanced Natural Language Processing (NLP), demonstrating strong performance in English. However, their effectiveness in Arabic, particularly for Named Entity Recognition (NER), remains limited, even with larger pre-trained models. This performance gap stems from multiple factors, including tokenisation, dataset quality, and annotation inconsistencies. Existing studies often analyze these issues in isolation, failing to capture their joint effect on system behaviour and performance. We introduce DeformAr (Debugging and Evaluation Framework for Transformer-based NER Systems), a novel framework designed to investigate and explain the performance discrepancy between Arabic and English NER systems. DeformAr integrates a data extraction library and an interactive dashboard, supporting two modes of evaluation: cross-component analysis and behavioural analysis. The framework divides each language into dataset and model components to examine their interactions. The analysis proceeds in two stages. First, cross-component analysis provides systematic diagnostic measures across data and model subcomponents, addressing the "what," "how," and "why" behind observed discrepancies. The second stage applies behavioural analysis by combining interpretability techniques with token-level metrics, interactive visualisations, and representation space analysis. These stages enable a component-aware diagnostic process that detects model behaviours and explains them by linking them to underlying representational patterns and data factors. DeformAr is the first Arabic-specific, component-based interpretability tool, offering a crucial resource for advancing model analysis in under-resourced languages.

Large Vision-Language Models (LVLMs) have shown impressive multimodal understanding capabilities, yet their robustness is poorly understood. In this paper, we investigate the structural vulnerabilities of LVLMs to identify any critical neurons whose removal triggers catastrophic collapse. In this context, we propose CAN, a method to detect Consistently Activated Neurons and to locate critical neurons by progressive masking. Experiments on LLaVA-1.5-7b-hf and InstructBLIP-Vicuna-7b reveal that masking only a tiny portion of the language model's feed-forward networks (just as few as four neurons in extreme cases) suffices to trigger catastrophic collapse. Notably, critical neurons are predominantly localized in the language model rather than in the vision components, and the down-projection layer is a particularly vulnerable structure. We also observe a consistent two-stage collapse pattern: initial expressive degradation followed by sudden, complete collapse. Our findings provide important insights for safety research in LVLMs.

Large Language Models (LLMs) have shown remarkable abilities in a wide variety of reasoning and factual knowledge tasks (Achiam et al., 2023; Bubeck et al., 2023; Katz et al., 2024), but their large size makes inference expensive. With the advent of agentic systems that interact with the external world, LLMs are poised to become even more ubiquitous in science, technology, and society, but the tremendous inference cost presents a challenge for realizing the full potential of these agents. One way to quell the computational expense associated with LLM inference is to use small language models (SLMs). With orders of magnitude fewer parameters, SLMs are faster, cheaper, and easier to finetune for specialised skills like tool use, making them natural specialists using proprietary data or within agentic systems (Belcak et al., 2025). Training language models typically involves three stages: pre-training on large general-purpose corpora, supervised finetuning (SFT), and reinforcement learning from human feedback (RLHF) or from verifiable rewards (RLVR) (Ouyang et al., 2022).

Existing foundation models (FMs) in the medical domain often require extensive fine-tuning or rely on training resource-intensive decoders, while many existing encoders are pretrained with objectives biased toward specific tasks. This illustrates a need for a strong, task-agnostic foundation model that requires minimal fine-tuning beyond feature extraction. In this work, we introduce a suite of task-agnostic pretraining of CT foundation models (TAP-CT): a simple yet effective adaptation of Vision Transformers (ViTs) and DINOv2 for volumetric data, enabling scalable self-supervised pretraining directly on 3D CT volumes. Our approach incorporates targeted modifications to patch embeddings, positional encodings, and volumetric augmentations, making the architecture depth-aware while preserving the simplicity of the underlying architectures. We show that large-scale 3D pretraining on an extensive in-house CT dataset (105K volumes) yields stable, robust frozen representations that generalize strongly across downstream tasks. To promote transparency and reproducibility, and to establish a powerful, low-resource baseline for future research in medical imaging, we will release all pretrained models, experimental configurations, and downstream benchmark code at https://huggingface.co/fomofo/tap-ct-b-3d.

Probabilistic forecasting is essential for modern risk management, allowing decision-makers to quantify uncertainty in critical systems. This paper tackles this challenge using the volatile REFIT household dataset, which is complicated by a large structural data gap. We first address this by conducting a rigorous comparative experiment to select a Seasonal Imputation method, demonstrating its superiority over linear interpolation in preserving the data's underlying distribution. We then systematically evaluate a hierarchy of models, progressing from classical baselines (SARIMA, Prophet) to machine learning (XGBoost) and advanced deep learning architectures (LSTM). Our findings reveal that classical models fail to capture the data's non-linear, regime-switching behavior. While the LSTM provided the most well-calibrated probabilistic forecast, the Temporal Fusion Transformer (TFT) emerged as the superior all-round model, achieving the best point forecast accuracy (RMSE 481.94) and producing safer, more cautious prediction intervals that effectively capture extreme volatility.

Embedding spaces are fundamental to modern AI, translating raw data into high-dimensional vectors that encode rich semantic relationships. Y et, their internal structures remain opaque, with existing approaches often sacrificing semantic coherence for structural regularity or incurring high computational overhead to improve interpretability. To address these challenges, we introduce the Semantic Field Subspace (SFS), a geometry-preserving, context-aware representation that captures local semantic neighborhoods within the embedding space. We also propose SAF ARI (SemAntic Field subspAce deteRmInation), an unsupervised, modality-agnostic algorithm that uncovers hierarchical semantic structures using a novel metric called Semantic Shift, which quantifies how semantics evolve as SFSes evolve. To ensure scalability, we develop an efficient approximation of Semantic Shift that replaces costly SVD computations, achieving a 15 30 speedup with average errors below 0.01. Extensive evaluations across six real-world text and image datasets show that SFSes outperform standard classifiers not only in classification but also in nuanced tasks such as political bias detection, while SAF ARI consistently reveals interpretable and generalizable semantic hierarchies. This work presents a unified framework for structuring, analyzing, and scaling semantic understanding in embedding spaces.