Large Language Models (LLMs) remain vulnerable to prompt injection attacks, representing the most significant security threat in production deployments. We present Prompt Fencing, a novel architectural approach that applies cryptographic authentication and data architecture principles to establish explicit security boundaries within LLM prompts. Our approach decorates prompt segments with cryptographically signed metadata including trust ratings and content types, enabling LLMs to distinguish between trusted instructions and untrusted content. While current LLMs lack native fence awareness, we demonstrate that simulated awareness through prompt instructions achieved complete prevention of injection attacks in our experiments, reducing success rates from 86.7% (260/300 successful attacks) to 0% (0/300 successful attacks) across 300 test cases with two leading LLM providers. We implement a proof-of-concept fence generation and verification pipeline with a total overhead of 0.224 seconds (0.130s for fence generation, 0.094s for validation) across 100 samples. Our approach is platform-agnostic and can be incrementally deployed as a security layer above existing LLM infrastructure, with the expectation that future models will be trained with native fence awareness for optimal security. Keywords: Large Language Models, Prompt Injection, Cryptographic Security, Trust Boundaries, LLM Security Note: The experiments described in this paper were conducted in October 2025. This paper was written and submitted in November 2025.

Large language models (LLMs) are increasingly used to generate self-explanations alongside their predictions, a practice that raises concerns about the faithfulness of these explanations, especially in low-resource languages. This study evaluates the faithfulness of LLM-generated explanations in the context of emotion classification in Persian, a low-resource language, by comparing the influential words identified by the model against those identified by human annotators. We assess faithfulness using confidence scores derived from token-level log-probabilities. Two prompting strategies, differing in the order of explanation and prediction (Predict-then-Explain and Explain-then-Predict), are tested for their impact on explanation faithfulness. Our results reveal that while LLMs achieve strong classification performance, their generated explanations often diverge from faithful reasoning, showing greater agreement with each other than with human judgments. These results highlight the limitations of current explanation methods and metrics, emphasizing the need for more robust approaches to ensure LLM reliability in multilingual and low-resource contexts.

Post-training quantization is an effective method for reducing the serving cost of large language models, where the standard approach is to use a round-to-nearest quantization level scheme. However, this often introduces large errors due to outliers in the weights. Proposed mitigation mechanisms include applying adaptive rounding, random rotation transformations or committing to a post-training target using calibration data. Unfortunately, this reliance on calibration data can be severely limiting in some real-world scenarios as such data may be unavailable or subject to privacy regulations. In this paper, we propose algorithms to optimize transformations and adaptive rounding without access to any calibration data. The optimization is achieved by designing a suitable proxy function for the quantization loss without calibration data. To maintain inference efficiency, we perform structured matrix transformations for single matrices. For paired weights that interact directly in the computation graph, we use dual matrix transformations and adaptive rounding methods. We conduct experiments on Gemma 2 models, and observe consistent improvement over the baselines. For Gemma 2 9B quantization, our method improves the average benchmark score from 61.9 to 62.4 for 4-bit quantization and from 52.0 to 60.6 for 3-bit quantization, while adding less than 3% of computation overhead. Furthermore, our method achieves performance comparable to the commonly used GPTQ method, which requires calibration data.

We introduce IndEgo, a multimodal egocentric and exocentric dataset addressing common industrial tasks, including assembly/disassembly, logistics and organisation, inspection and repair, woodworking, and others. The dataset contains 3,460 egocentric recordings (approximately 197 hours), along with 1,092 exocentric recordings (approximately 97 hours). A key focus of the dataset is collaborative work, where two workers jointly perform cognitively and physically intensive tasks. The egocentric recordings include rich multimodal data and added context via eye gaze, narration, sound, motion, and others. We provide detailed annotations (actions, summaries, mistake annotations, narrations), metadata, processed outputs (eye gaze, hand pose, semi-dense point cloud), and benchmarks on procedural and non-procedural task understanding, Mistake Detection, and reasoning-based Question Answering. Baseline evaluations for Mistake Detection, Question Answering and collaborative task understanding show that the dataset presents a challenge for the state-of-the-art multimodal models. Our dataset is available at: https://huggingface.co/datasets/FraunhoferIPK/IndEgo

Financial applications of large language models (LLMs) require factual reliability and computational efficiency, yet current systems often hallucinate details and depend on prohibitively large models. We propose FISCAL (Financial Synthetic Claim-Document Augmented Learning), a modular framework for generating synthetic data tailored to financial fact-checking. Using FISCAL, we generate a dataset called FISCAL-data and use it to train MiniCheck-FISCAL, a lightweight verifier for numerical financial claims. MiniCheck-FISCAL outperforms its baseline, surpasses GPT-3.5 Turbo and other open-source peers of similar size, and approaches the accuracy of much larger systems (20x), such as Mixtral-8x22B and Command R+. On external datasets FinDVer and Fin-Fact, it rivals GPT-4o and Claude-3.5 while outperforming Gemini-1.5 Flash. These results show that domain-specific synthetic data, combined with efficient fine-tuning, enables compact models to achieve state-of-the-art accuracy, robustness, and scalability for practical financial AI. The dataset and scripts are available in the project repository (link provided in the paper).

Conventional AI-driven AMS design automation algorithms remain constrained by their reliance on high-quality datasets to capture underlying circuit behavior, coupled with poor transferability across architectures, and a lack of adaptive mechanisms. This work proposes HeaRT, a foundational reasoning engine for automation loops and a first step toward intelligent, adaptive, human-style design optimization. HeaRT consistently demonstrates reasoning accuracy >97% and Pass@1 performance >98% across our 40-circuit benchmark repository, even as circuit complexity increases, while operating at <0.5x real-time token budget of SOTA baselines. Our experiments show that HeaRT yields >3x faster convergence in both sizing and topology design adaptation tasks across diverse optimization approaches, while preserving prior design intent.

Progress in computer use agents (CUAs) has been constrained by the absence of large and high-quality datasets that capture how humans interact with a computer. While LLMs have thrived on abundant textual data, no comparable corpus exists for CUA trajectories. To address these gaps, we introduce FaraGen, a novel synthetic data generation system for multi-step web tasks. FaraGen can propose diverse tasks from frequently used websites, generate multiple solution attempts, and filter successful trajectories using multiple verifiers. It achieves high throughput, yield, and diversity for multi-step web tasks, producing verified trajectories at approximately $1 each. We use this data to train Fara-7B, a native CUA model that perceives the computer using only screenshots, executes actions via predicted coordinates, and is small enough to run on-device. We find that Fara-7B outperforms other CUA models of comparable size on benchmarks like WebVoyager, Online-Mind2Web, and WebTailBench -- our novel benchmark that better captures under-represented web tasks in pre-existing benchmarks. Furthermore, Fara-7B is competitive with much larger frontier models, illustrating key benefits of scalable data generation systems in advancing small efficient agentic models. We are making Fara-7B open-weight on Microsoft Foundry and HuggingFace, and we are releasing WebTailBench.

Abstract--This study examines whether Low-Rank Adaptation (LoRA) fine-tuned Large Language Models (LLMs) can approximate the performance of fully fine-tuned models in generating human-interpretable decisions and explanations for malware classification. Achieving trustworthy malware detection, particularly when LLMs are involved, remains a significant challenge. We developed an evaluation framework using Bilingual Evaluation Understudy (BLEU), Recall-Oriented Understudy for Gisting Evaluation (ROUGE), and Semantic Similarity Metrics to benchmark explanation quality across five LoRA configurations and a fully fine-tuned baseline. Results indicate that full fine-tuning achieves the highest overall scores, with BLEU and ROUGE improvements of up to 10% over LoRA variants. However, mid-range LoRA models deliver competitive performance--exceeding full fine-tuning on two metrics--while reducing model size by approximately 81% and training time by over 80% on a LoRA model with 15.5% trainable parameters. These findings demonstrate that LoRA offers a practical balance of interpretability and resource efficiency, enabling deployment in resource-constrained environments without sacrificing explanation quality. By providing feature-driven natural language explanations for malware classifications, this approach enhances transparency, analyst confidence, and operational scalability in malware detection systems. Modern AI-based malware detection systems often lack trustworthiness, particularly when LLMs are involved, limiting analysts' ability to validate automated decisions and improve detection strategies.

Abstract--Multi-hop question answering over knowledge graphs remains computationally challenging due to the combinatorial explosion of possible reasoning paths. Recent approaches rely on expensive Large Language Model (LLM) inference for both entity linking and path ranking, limiting their practical deployment. Additionally, LLM-generated answers often lack verifiable grounding in structured knowledge. We present two complementary hybrid algorithms that address both efficiency and verifiability: (1) LLM-Guided Planning that uses a single LLM call to predict relation sequences executed via breadth-first search, achieving near-perfect accuracy (micro-F1 > 0.90) while ensuring all answers are grounded in the knowledge graph, and (2) Embedding-Guided Neural Search that eliminates LLM calls entirely by fusing text and graph embeddings through a lightweight 6.7M-parameter edge scorer, achieving over 100 speedup with competitive accuracy. Through knowledge distillation, we compress planning capability into a 4B-parameter model that matches large-model performance at zero API cost. Evaluation on MetaQA demonstrates that grounded reasoning consistently outperforms ungrounded generation, with structured planning proving more transferable than direct answer generation. Our results show that verifiable multi-hop reasoning does not require massive models at inference time, but rather the right architectural inductive biases combining symbolic structure with learned representations. Knowledge graphs (KGs) have emerged as powerful structures for representing domain-specific, structured information that supports verifiable, multi-hop reasoning. Meanwhile, large language models (LLMs) trained on vast web-scale corpora have achieved impressive fluency and generalization across a wide range of tasks.

Foundation models (FM) have unlocked powerful zero-shot capabilities in vision and language, yet their reliance on internet pretraining data leaves them brittle in unstructured, real-world settings. The messy, real-world data encountered during deployment (e.g. occluded or multilingual text) remains massively underrepresented in existing corpora. Robots, as embodied agents, are uniquely positioned to close this gap: they can act in physical environments to collect large-scale, real-world data that enriches FM training with precisely the examples current models lack. We introduce the Robot-Powered Data Flywheel, a framework that transforms robots from FM consumers into data generators. By deploying robots equipped with FMs in the wild, we enable a virtuous cycle: robots perform useful tasks while collecting real-world data that improves both domain-specific adaptation and domain-adjacent generalization. We instantiate this framework with Scanford, a mobile manipulator deployed in the East Asia Library for 2 weeks. Scanford autonomously scans shelves, identifies books using a vision-language model (VLM), and leverages the library catalog to label images without human annotation. This deployment both aids librarians and produces a dataset to finetune the underlying VLM, improving performance on the domain-specific in-the-wild library setting and on domain-adjacent multilingual OCR benchmarks. Using data collected from 2103 shelves, Scanford improves VLM performance on book identification from 32.0% to 71.8% and boosts domain-adjacent multilingual OCR from 24.8% to 46.6% (English) and 30.8% to 38.0% (Chinese), while saving an ~18.7 hrs of human time. These results highlight how robot-powered data flywheels can both reduce human effort in real deployments and unlock new pathways for continually adapting FMs to the messiness of reality. More details are at: https://scanford-robot.github.io