arXiv subjects

Zafar Ayyub Qazi

Publications and source records attributed to Zafar Ayyub Qazi.

12 recordsLinked to original sources

arXiv2026

Affix Cache for Diffusion Large Language Models

Diffusion Large Language Models (DLLMs) enable non-autoregressive decoding, but efficient inference support remains immature: unlike autoregressive models, whose requests reuse a shared prefix key-value (KV) cache, DLLMs use bidirectional attention, so a shared context's KV states depend on the tokens still being decoded, leaving directly reused caches stale and full recomputation necessary. We present ACache, a cross-request cache reuse mechanism for shared spans, or affixes, at any position: prefix, infix, or suffix. ACache measures the influence of affix tokens on the masked generation region to identify a small request-specific subset as Anchor Tokens, and recomputes only their KV states while reusing the remaining affix cache. Built on state-of-the-art intra-request caching mechanisms, ACache recovers most of the accuracy lost to direct affix-cache reuse on average when recomputing around 20% of affix tokens, and at that budget preserves more accuracy than selection criteria adapted from prior cross-request cache-reuse systems. We co-design ACache with a modern inference engine, whose attention reads each request's recomputed Anchor KV states alongside one affix cache shared across concurrent requests. Against the same system with only intra-request caching, ACache cuts recompute latency by up to 56.7%, translating to as much as 1.71$\times$ end-to-end throughput, while reducing peak KV cache memory by up to 45.8%.

cs.CL↗

arXiv2026

COMED: The Missing Middle Between Routing and Collaboration in Multi-LLM Inference

No single Large Language Model (LLM) is uniformly reliable across queries, motivating multi-model inference systems that either route among models or combine their outputs. However, routing stops after selecting an initial model, while dense collaboration invokes peers on every query. We show that collaboration is non-monotonic: peers can recover failures that no model solves alone, but can also corrupt initially correct answers. We introduce COMED (Controlled Model Escalation for Multi-LLM Deliberation), a post-anchor controller for selective cross-model collaboration. COMED uses anchor self-consistency, router margin, and a lightweight peer probe to accept confident answers, verify ambiguous cases, and escalate only when collaboration is likely beneficial. We formalize this trade-off with a rescue-harm decomposition showing that selective collaboration improves when rescued errors outweigh collaboration-induced harms. Across medical, scientific, and general reasoning benchmarks, COMED improves fixed and routed anchors in all 16 open-weight settings, with gains up to +10.7 percentage points on MedQA while invoking fewer models and using fewer decoded tokens than dense collaboration. On HLE with frontier models, COMED improves GPT-5.5 from 23.1% to 28.1%, outperforming dense collaboration and achieving the best results.

cs.CL↗

arXiv2026

Separating Clicks from Baits: Using Large Language Models to Detect Misleading YouTube Thumbnails

Misleading video thumbnails on platforms like YouTube are a pervasive problem, undermining user trust and platform integrity. This paper proposes a novel multi-modal detection pipeline that uses Large Language Models (LLMs) to flag misleading thumbnails. We first construct a comprehensive dataset of 2,843 videos from eight countries, including 1,359 misleading thumbnail videos that collectively amassed over 7.6 billion views, providing a unique cross-cultural perspective on this global issue. Our detection pipeline integrates video-to-text descriptions, thumbnail images, and subtitle transcripts to holistically analyze content and flag misleading thumbnails. Through extensive experimentation and prompt engineering, we evaluate the performance of four frontier-level LLMs, including GPT-4o, GPT-4o Mini, Claude 3.5 Sonnet, and Gemini-1.5 Flash. We further evaluate open-weight vision-language models, LLaVA-v1.5 and Qwen2.5-VL-7B-Instruct, to assess the generalizability of our approach beyond proprietary systems. Our findings show the effectiveness of LLMs in identifying misleading thumbnails, with Claude 3.5 Sonnet consistently showing strong performance, achieving an accuracy of 93.8%, precision over 92%, and recall exceeding 94% in certain scenarios. Beyond evaluating detection performance, we conducted a careful failure analysis to understand when LLMs fail in identifying misleading thumbnails. We discuss the implications of our findings for content moderation, user experience, and the ethical considerations of deploying such systems at scale. Our findings pave the way for more transparent, trustworthy video platforms and stronger content integrity for audiences worldwide.

cs.SI↗

arXiv2026

Efficient and Adaptable Detection of Malicious LLM Prompts via Bootstrap Aggregation

Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language understanding, reasoning, and generation. However, these systems remain susceptible to malicious prompts that induce unsafe or policy-violating behavior through harmful requests, jailbreak techniques, and prompt injection attacks. Existing defenses face fundamental limitations: black-box moderation APIs offer limited transparency and adapt poorly to evolving threats, while white-box approaches using large LLM judges impose prohibitive computational costs and require expensive retraining for new attacks. Current systems force designers to choose between performance, efficiency, and adaptability. To address these challenges, we present BAGEL (Bootstrap AGgregated Ensemble Layer), a modular, lightweight, and incrementally updatable framework for malicious prompt detection. BAGEL employs a bootstrap aggregation and mixture of expert inspired ensemble of fine-tuned models, each specialized on a different attack dataset. At inference, BAGEL uses a random forest router to identify the most suitable ensemble member, then applies stochastic selection to sample additional members for prediction aggregation. When new attacks emerge, BAGEL updates incrementally by fine-tuning a small prompt-safety classifier (86M parameters) and adding the resulting model to the ensemble. BAGEL achieves an F1 score of 0.92 by selecting just 5 ensemble members (430M parameters), outperforming OpenAI Moderation API and ShieldGemma which require billions of parameters. Performance remains robust after nine incremental updates, and BAGEL provides interpretability through its router's structural features. Our results show ensembles of small finetuned classifiers can match or exceed billion-parameter guardrails while offering the adaptability and efficiency required for production systems.

cs.LG↗

arXiv2025

TweakLLM: A Routing Architecture for Dynamic Tailoring of Cached Responses

Large Language Models (LLMs) process millions of queries daily, making efficient response caching a compelling optimization for reducing cost and latency. However, preserving relevance to user queries using this approach proves difficult due to the personalized nature of chatbot interactions and the limited accuracy of semantic similarity search. To address this, we present TweakLLM, a novel routing architecture that employs a lightweight LLM to dynamically adapt cached responses to incoming prompts. Through comprehensive evaluation, including user studies with side-by-side comparisons, satisfaction voting, as well as multi-agent LLM debates, we demonstrate that TweakLLM maintains response quality comparable to frontier models while significantly improving cache effectiveness. Our results across real-world datasets highlight TweakLLM as a scalable, resource-efficient caching solution for high-volume LLM deployments without compromising user experience.

cs.LG↗

arXiv2025

ThumbnailTruth: A Multi-Modal LLM Approach for Detecting Misleading YouTube Thumbnails Across Diverse Cultural Settings

Misleading video thumbnails on platforms like YouTube are a pervasive problem, undermining user trust and platform integrity. This paper proposes a novel multi-modal detection pipeline that uses Large Language Models (LLMs) to flag misleading thumbnails. We first construct a comprehensive dataset of 2,843 videos from eight countries, including 1,359 misleading thumbnail videos that collectively amassed over 7.6 billion views -- providing a unique cross-cultural perspective on this global issue. Our detection pipeline integrates video-to-text descriptions, thumbnail images, and subtitle transcripts to holistically analyze content and flag misleading thumbnails. Through extensive experimentation and prompt engineering, we evaluate the performance of state-of-the-art LLMs, including GPT-4o, GPT-4o Mini, Claude 3.5 Sonnet, and Gemini-1.5 Flash. Our findings show the effectiveness of LLMs in identifying misleading thumbnails, with Claude 3.5 Sonnet consistently showing strong performance, achieving an accuracy of 93.8\%, precision over 92\%, and recall exceeding 94\% in certain scenarios. We discuss the implications of our findings for content moderation, user experience, and the ethical considerations of deploying such systems at scale. Our findings pave the way for more transparent, trustworthy video platforms and stronger content integrity for audiences worldwide.

cs.SI↗

arXiv2025

Semantic Caching for Improving Web Affordability

The rapid growth of web content has led to increasingly large webpages, posing significant challenges for Internet affordability, especially in developing countries where data costs remain prohibitively high. We propose semantic caching using Large Language Models (LLMs) to improve web affordability by enabling reuse of semantically similar images within webpages. Analyzing 50 leading news and media websites, encompassing 4,264 images and over 40,000 image pairs, we demonstrate potential for significant data transfer reduction, with some website categories showing up to 37% of images as replaceable. Our proof-of-concept architecture shows users can achieve approximately 10% greater byte savings compared to exact caching. We evaluate both commercial and open-source multi-modal LLMs for assessing semantic replaceability. GPT-4o performs best with a low Normalized Root Mean Square Error of 0.1735 and a weighted F1 score of 0.8374, while the open-source LLaMA 3.1 model shows comparable performance, highlighting its viability for large-scale applications. This approach offers benefits for both users and website operators, substantially reducing data transmission. We discuss ethical concerns and practical challenges, including semantic preservation, user-driven cache configuration, privacy concerns, and potential resistance from website operators

cs.NI↗

arXiv2024

Warping the Edge: Where Instant Mobility in 5G Meets Stateful Applications

Edge computing is considered a key paradigm for supporting real-time applications over 5G networks, as hosting applications at the network edge can substantially reduce delays. A significant fraction of real-time applications over 5G are expected to be highly mobile applications. However, one challenge with hosting mobile applications on the network edge is ensuring that users continue to get low latency as they move across different locations. This requires the support to handover clients to different edge sites with negligible application delays. However, many edge applications are stateful and can experience significant downtime during state migration over 5G. This paper addresses the problem of enabling stateful mobile edge applications in 5G networks. We first identify the key architectural issues and then propose a new system design, EdgeWarp, that mitigates delays during mobility through proactive application state migration. To enable this, we extend the existing edge data stores with the design of a novel two-step application state synchronization protocol, that leverages the early prediction of the target edge host. Additionally, EdgeWarp prioritizes the handover of latency-sensitive edge applications by communicating their latency requirements to the 5G control plane at the beginning of a data session. Our evaluation with real edge applications shows up to a 15.4x reduction in application downtime under mobility. We have made our anonymized code publicly accessible here.

cs.NI↗

arXiv2024

Rethinking Image Compression on the Web with Generative AI

The rapid growth of the Internet, driven by social media, web browsing, and video streaming, has made images central to the Web experience, resulting in significant data transfer and increased webpage sizes. Traditional image compression methods, while reducing bandwidth, often degrade image quality. This paper explores a novel approach using generative AI to reconstruct images at the edge or client-side. We develop a framework that leverages text prompts and provides additional conditioning inputs like Canny edges and color palettes to a text-to-image model, achieving up to 99.8% bandwidth savings in the best cases and 92.6% on average, while maintaining high perceptual similarity. Empirical analysis and a user study show that our method preserves image meaning and structure more effectively than traditional compression methods, offering a promising solution for reducing bandwidth usage and improving Internet affordability with minimal degradation in image quality.

cs.NI↗

arXiv2021

FedPrune: Towards Inclusive Federated Learning

Federated learning (FL) is a distributed learning technique that trains a shared model over distributed data in a privacy-preserving manner. Unfortunately, FL's performance degrades when there is (i) variability in client characteristics in terms of computational and memory resources (system heterogeneity) and (ii) non-IID data distribution across clients (statistical heterogeneity). For example, slow clients get dropped in FL schemes, such as Federated Averaging (FedAvg), which not only limits overall learning but also biases results towards fast clients. We propose FedPrune; a system that tackles this challenge by pruning the global model for slow clients based on their device characteristics. By doing so, slow clients can train a small model quickly and participate in FL which increases test accuracy as well as fairness. By using insights from Central Limit Theorem, FedPrune incorporates a new aggregation technique that achieves robust performance over non-IID data. Experimental evaluation shows that Fed- Prune provides robust convergence and better fairness compared to Federated Averaging.

cs.LG↗

arXiv2016

MAC Rate Adaptation and Cross Layer Behavior for Vehicular WiFi Access: An Experimental Study

Vehicular WiFi is different from conventional WiFi access. Firstly, as the connections arise opportunistically, they are short lived and intermittent. Secondly, at vehicular speeds channel conditions change rapidly. Under these conditions, the MAC bit rate adaptation plays a critical role to ensure that devices can maximize throughput performance. The goal of this work is to gain a deeper understanding of current MAC bit rate adaptation algorithms and their interaction with higher layer protocols under vehicular settings. Towards this end, we evaluate the performance of four MAC bit rate adaptation algorithms; SampleRate, AMRR, Minstrel, and RRAA, and their interactions with transport layer protocols such as TCP in real world vehicular settings. Our experimental study reveals that in general these algorithm perform poorly in vehicular settings, and can have an extremely adverse impact on TCP performance.

cs.NI↗

arXiv2014

A Framework to Quantify the Benefits of Network Functions Virtualization in Cellular Networks

Network functions virtualization (NFV) is an appealing vision that promises to dramatically reduce capital and operating expenses for cellular providers. However, existing efforts in this space leave open broad issues about how NFV deployments should be instantiated or how they should be provisioned. In this paper, we present an initial attempt at a framework that will help network operators systematically evaluate the potential benefits that different points in the NFV design space can offer.

cs.NI↗