AI agents are emerging as a dominant workload in a wide range of applications, promising to be the vehicle that delivers the promised benefits of AI to enterprises and consumers. Unlike conventional software or static inference, agentic workloads are dynamic and structurally complex. Often these agents are directed graphs of compute and IO operations that span multi-modal data input and conversion), data processing and context gathering (e.g vector DB lookups), multiple LLM inferences, tool calls, etc. To scale AI agent usage, we need efficient and scalable deployment and agent-serving infrastructure. To tackle this challenge, in this paper, we present a system design for dynamic orchestration of AI agent workloads on heterogeneous compute infrastructure spanning CPUs and accelerators, both from different vendors and across different performance tiers within a single vendor. The system delivers several building blocks: a framework for planning and optimizing agentic AI execution graphs using cost models that account for compute, memory, and bandwidth constraints of different HW; a MLIR based representation and compilation system that can decompose AI agent execution graphs into granular operators and generate code for different HW options; and a dynamic orchestration system that can place the granular components across a heterogeneous compute infrastructure and stitch them together while meeting an end-to-end SLA. Our design performs a systems level TCO optimization and preliminary results show that leveraging a heterogeneous infrastructure can deliver significant TCO benefits. A preliminary surprising finding is that for some workloads a heterogeneous combination of older generation GPUs with newer accelerators can deliver similar TCO as the latest generation homogenous GPU infrastructure design, potentially extending the life of deployed infrastructure.

This paper presents HITSZ's submission for the IWSLT 2025 Indic track, focusing on speech-to-text translation (ST) for English-to-Indic and Indic-to-English language pairs. To enhance translation quality in this low-resource scenario, we propose an end-to-end system integrating the pre-trained Whisper automated speech recognition (ASR) model with Krutrim, an Indic-specialized large language model (LLM). Experimental results demonstrate that our end-to-end system achieved average BLEU scores of $28.88$ for English-to-Indic directions and $27.86$ for Indic-to-English directions. Furthermore, we investigated the Chain-of-Thought (CoT) method. While this method showed potential for significant translation quality improvements on successfully parsed outputs (e.g. a $13.84$ BLEU increase for Tamil-to-English), we observed challenges in ensuring the model consistently adheres to the required CoT output format.

Deploying Large Language Models (LLMs) on edge devices remains challenging due to their quadratically increasing computations with the sequence length. Existing studies for dynamic attention pruning are designed for hardware with massively parallel computation capabilities, such as GPUs or TPUs, and aim at long context lengths (e.g., 64K), making them unsuitable for edge scenarios. We present DeltaLLM, a training-free framework that exploits temporal sparsity in attention patterns to enable efficient LLM inference across both the prefilling and decoding stages, on resource-constrained edge devices. DeltaLLM introduces an accuracy- and memory-aware delta matrix construction strategy that introduces temporal sparsity, and a context-aware hybrid attention mechanism that combines full attention in a local context window with delta approximation outside it to increase accuracy. We evaluate our framework on the edge-device-friendly BitNet-b1.58-2B-4T model and Llama3.2-1B-Instruct model across diverse language tasks. The results show that on BitNet, our framework increases the attention sparsity from 0% to 60% during the prefilling stage with slight accuracy improvement on the WG task, and 0% to 57% across both the prefilling and decoding stages, with even higher F1 score from 29.63 to 30.97 on SQuAD-v2 task. On the Llama model, it can also achieve up to 60% sparsity during the prefilling stage and around 57% across both stages with negligible accuracy drop. These results demonstrate that DeltaLLM offers a promising solution for efficient edge deployment, requiring no fine-tuning and seamlessly integrating with existing inference pipelines.

Recent advancements in Large Language Models (LLMs) have significantly enhanced their code generation capabilities. However, their robustness against adversarial misuse, particularly through multi-turn malicious coding prompts, remains underexplored. In this work, we introduce code decomposition attacks, where a malicious coding task is broken down into a series of seemingly benign subtasks across multiple conversational turns to evade safety filters. To facilitate systematic evaluation, we introduce \benchmarkname{}, a large-scale benchmark designed to evaluate the robustness of code LLMs against both single-turn and multi-turn malicious prompts. Empirical results across open- and closed-source models reveal persistent vulnerabilities, especially under multi-turn scenarios. Fine-tuning on MOCHA improves rejection rates while preserving coding ability, and importantly, enhances robustness on external adversarial datasets with up to 32.4% increase in rejection rates without any additional supervision.

Large language models (LLMs) possess extensive world knowledge, including geospatial knowledge, which has been successfully applied to various geospatial tasks such as mobility prediction and social indicator prediction. However, LLMs often generate inaccurate geospatial knowledge, leading to geospatial hallucinations (incorrect or inconsistent representations of geospatial information) that compromise their reliability. While the phenomenon of general knowledge hallucination in LLMs has been widely studied, the systematic evaluation and mitigation of geospatial hallucinations remain largely unexplored. To address this gap, we propose a comprehensive evaluation framework for geospatial hallucinations, leveraging structured geospatial knowledge graphs for controlled assessment. Through extensive evaluation across 20 advanced LLMs, we uncover the hallucinations in their geospatial knowledge. Building on these insights, we introduce a dynamic factuality aligning method based on Kahneman-Tversky Optimization (KTO) to mitigate geospatial hallucinations in LLMs, leading to a performance improvement of over 29.6% on the proposed benchmark. Extensive experimental results demonstrate the effectiveness of our benchmark and learning algorithm in enhancing the trustworthiness of LLMs in geospatial knowledge and reasoning tasks.

This paper presents MCP4EDA, the first Model Context Protocol server that enables Large Language Models (LLMs) to control and optimize the complete open-source RTL-to-GDSII design flow through natural language interaction. The system integrates Yosys synthesis, Icarus Verilog simulation, OpenLane place-and-route, GTKWave analysis, and KLayout visualization into a unified LLM-accessible interface, enabling designers to execute complex multi-tool EDA workflows conversationally via AI assistants such as Claude Desktop and Cursor IDE. The principal contribution is a backend-aware synthesis optimization methodology wherein LLMs analyze actual post-layout timing, power, and area metrics from OpenLane results to iteratively refine synthesis TCL scripts, establishing a closed-loop optimization system that bridges the traditional gap between synthesis estimates and physical implementation reality. In contrast to conventional flows that rely on wire-load models, this methodology leverages real backend performance data to guide synthesis parameter tuning, optimization sequence selection, and constraint refinement, with the LLM functioning as an intelligent design space exploration agent. Experimental evaluation on representative digital designs demonstrates 15-30% improvements in timing closure and 10-20% area reduction compared to default synthesis flows, establishing MCP4EDA as the first practical LLM-controlled end-to-end open-source EDA automation system. The code and demo are avaiable at: http://www.agent4eda.com/

The undergraduate thesis (UGTE) plays an indispensable role in assessing a student's cumulative academic development throughout their college years. Although large language models (LLMs) have advanced education intelligence, they typically focus on holistic assessment with only one single evaluation score, but ignore the intricate nuances across multifaceted criteria, limiting their ability to reflect structural criteria, pedagogical objectives, and diverse academic competencies. Meanwhile, pedagogical theories have long informed manual UGTE evaluation through multi-dimensional assessment of cognitive development, disciplinary thinking, and academic performance, yet remain underutilized in automated settings. Motivated by the research gap, we pioneer PEMUTA, a pedagogically-enriched framework that effectively activates domain-specific knowledge from LLMs for multi-granular UGTE assessment. Guided by Vygotsky's theory and Bloom's Taxonomy, PEMUTA incorporates a hierarchical prompting scheme that evaluates UGTEs across six fine-grained dimensions: Structure, Logic, Originality, Writing, Proficiency, and Rigor (SLOWPR), followed by holistic synthesis. Two in-context learning techniques, \ie, few-shot prompting and role-play prompting, are also incorporated to further enhance alignment with expert judgments without fine-tuning. We curate a dataset of authentic UGTEs with expert-provided SLOWPR-aligned annotations to support multi-granular UGTE assessment. Extensive experiments demonstrate that PEMUTA achieves strong alignment with expert evaluations, and exhibits strong potential for fine-grained, pedagogically-informed UGTE evaluations.

The vast majority of Web pages fail to comply with established Web accessibility guidelines, excluding a range of users with diverse abilities from interacting with their content. Making Web pages accessible to all users requires dedicated expertise and additional manual efforts from Web page providers. To lower their efforts and promote inclusiveness, we aim to automatically detect and correct Web accessibility violations in HTML code. While previous work has made progress in detecting certain types of accessibility violations, the problem of automatically detecting and correcting accessibility violations remains an open challenge that we address. We introduce a novel taxonomy classifying Web accessibility violations into three key categories - Syntactic, Semantic, and Layout. This taxonomy provides a structured foundation for developing our detection and correction method and redefining evaluation metrics. We propose a novel method, AccessGuru, which combines existing accessibility testing tools and Large Language Models (LLMs) to detect violations and applies taxonomy-driven prompting strategies to correct all three categories. To evaluate these capabilities, we develop a benchmark of real-world Web accessibility violations. Our benchmark quantifies syntactic and layout compliance and judges semantic accuracy through comparative analysis with human expert corrections. Evaluation against our benchmark shows that AccessGuru achieves up to 84% average violation score decrease, significantly outperforming prior methods that achieve at most 50%.

Large language models (LLMs) are increasingly applied in task-oriented dialogue (TOD) systems but often struggle with long, conditional workflows that involve external tool calls and depend on user-specific information. We present Workflow Adherence via Runtime Parallel Personalization, or WARPP, a training-free, modular framework that combines multi-agent orchestration with runtime personalization to improve workflow adherence in LLM-based systems. By dynamically pruning conditional branches based on user attributes, the framework reduces reasoning overhead and narrows tool selection at runtime. WARPP deploys a parallelized architecture where a dedicated Personalizer agent operates alongside modular, domain-specific agents to dynamically tailor execution paths in real time. The framework is evaluated across five representative user intents of varying complexity within three domains: banking, flights, and healthcare. Our evaluation leverages synthetic datasets and LLM-powered simulated users to test scenarios with conditional dependencies. Our results demonstrate that WARPP outperforms both the non-personalized method and the ReAct baseline, achieving increasingly larger gains in parameter fidelity and tool accuracy as intent complexity grows, while also reducing average token usage, without any additional training.

Text-attributed graphs (TAGs) have emerged as a powerful representation for modeling complex relationships across diverse domains. With the rise of large language models (LLMs), there is growing interest in leveraging their capabilities for graph learning. However, current approaches face significant challenges in embedding structural information into LLM-compatible formats, requiring either computationally expensive alignment mechanisms or manual graph verbalization techniques that often lose critical structural details. Moreover, these methods typically require labeled data from source domains for effective transfer learning, significantly constraining their adaptability. We propose STAG, a novel self-supervised framework that directly quantizes graph structural information into discrete tokens using a frozen codebook. Unlike traditional quantization approaches, our method employs soft assignment and KL divergence guided quantization to address the unique challenges of graph data, which lacks natural tokenization structures. Our framework enables both LLM-based and traditional learning approaches, supporting true zero-shot transfer learning without requiring labeled data even in the source domain. Extensive experiments demonstrate state-of-the-art performance across multiple node classification benchmarks while maintaining compatibility with different LLM architectures, offering an elegant solution to bridging graph learning with LLMs.