The proliferation of Model Context Protocol (MCP) servers in enterprise infrastructure management has revolutionized AI-driven automation while introducing critical multi-layered security vulnerabilities that traditional cybersecurity frameworks cannot adequately address. This paper presents a comprehensive intelligent guardrail system that addresses the unique security challenges of MCP-driven infrastructure management through a novel four-layer defense architecture. Our solution employs a dedicated guardian LLM that interprets natural language policies and applies contextual reasoning to complex infrastructure scenarios, providing dynamic policy enforcement that adapts to user roles, operational timing, and system context. Unlike existing rule-based security systems, our approach implements guardrails at four distinct control points: input message filtering, tool selection validation, execution-time verification, and post-action auditing. The system addresses critical gaps in existing security solutions by providing infrastructure-specific threat modeling, real-time policy adaptation, and comprehensive audit trails with explainable decision-making through confidence scores and detailed reasoning. Our evaluation demonstrates the system's effectiveness in preventing command injection, privilege escalation, and tool poisoning attacks across various enterprise infrastructure scenarios while maintaining operational agility essential for modern data center management.

Hallucinated object mentions remain a persistent failure mode of vision-language models (VLMs) across generation tasks such as image captioning and visual question answering: outputs may be fluent, yet include entities not supported by visual evidence.Existing mitigation approaches often reduce hallucinations at the cost of degraded generation quality or require expensive retraining and taskspecific supervision.We introduce CEBC, a lightweight, training-free framework for lowhallucination vision-language generation based on conformal evidence-bounded minimal editing.CEBC first produces a strong base output (via greedy decoding or best-of-K sampling), then applies an evidence-bounded editing step that minimally revises or suppresses unsupported object mentions using constraints derived from an external visual detector.Crucially, the evidence threshold is conformally calibrated on a small held-out set via quantiles of detector confidence scores, enabling explicit and controllable hallucination risk at test time.To balance factuality and informativeness, we further introduce a risk-first, qualityaware selection rule that prioritizes evidenceconsistent generations while regularizing unnecessary length or lexical drift.Extensive experiments on MS-COCO and GQA for image captioning, and POPE for VQA evaluation across multiple VLMs demonstrate that CEBC consistently reduces hallucination rates (CHAIR S , CHAIR I , POPE) while maintaining or improving standard generation quality metrics (CIDEr, BLEU, CLIPScore).CEBC establishes a stronger factuality-quality Pareto frontier without any additional model training or access to paired supervision beyond an offthe-shelf detector.

Gayathri Saranathan, Cong Xu, Mahammad Parwez Alam, Tarun Kumar, Martin Foltin, Soon Yee Wong, Suparna Bhattacharya. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Training vision-based Autonomous driving models is a challenging problem with enormous practical implications. One of the main challenges is the requirement of storage and processing of vast volumes of (possibly redundant) driving video data. In this paper, we study the problem of data-efficient training of autonomous driving systems. We argue that in the context of an edge-device deployment, multi-criteria online video frame subset selection is an appropriate technique for developing such frameworks. We study existing convex optimization based solutions and show that they are unable to provide solution with high weightage to loss of selected video frames. We design a novel multi-criteria online subset selection algorithm, TMCOSS, which uses a thresholded concave function of selection variables. Extensive experiments using driving simulator CARLA show that we are able to drop 80% of the frames, while succeeding to complete 100% of the episodes. We also show that TMCOSS improves performance on the crucial affordance "Relative Angle" during turns, on inclusion of bucket-specific relative angle loss (BL), leading to selection of more frames in those parts. TMCOSS also achieves an 80% reduction in number of training video frames, on real-world videos from the standard BDD and Cityscapes datasets, for the tasks of drivable area segmentation, and semantic segmentation.

A wide range of multimodal sensors such as sensors, video cameras, wearables worn by users in an IoT powered smarthome provide important albeit huge amount of data through which applications derive meaningful Activities of Daily Living (ADLs). Storing this massive amount of data is a significant challenge for efficient execution of the corresponding applications to meet real-time demands; there is scope for improving it as it is found that a substantial amount of the data produced may be unimportant. In this paper, we propose an end-to-end system GradeSense, which implements a grading mechanism based on multimodal data fusion by categorizing ‘important’ data. GradeSense is made complete by an applicationindependent storage module that leverages our grading scheme (as opposed to traditional usage-based models) for efficient storing. Activity prediction algorithms perform well (up to 17% improvement) with this now-fused and important data which is a mere fraction of the entire data, achieving 87% data reduction on average in faster storage tier. The throughput of GradeSense, measured through runtime, gives improvement up to 46% due to our enhanced data-categorization and storing mechanism.

In a recent pioneering approach LDA was used to discover cross cutting concerns(CCC) automatically from software codebases. LDA though successful in detecting prominent concerns, fails to detect many useful CCCs including ones that may be heavily executed but elude discovery because they do not have a strong prevalence in source-code. We pose this problem as that of discovering topics that rarely occur in individual documents, which we will refer to as subtle topics. Recently an interesting approach, namely focused topic models(FTM) was proposed for detecting rare topics. FTM, though successful in detecting topics which occur prominently in very few documents, is unable to detect subtle topics. Discovering subtle topics thus remains an important open problem. To address this issue we propose subtle topic models(STM). STM uses a generalized stick breaking process(GSBP) as a prior for defining multiple distributions over topics. This hierarchical structure on topics allows STM to discover rare topics beyond the capabilities of FTM. The associated inference is non-standard and is solved by exploiting the relationship between GSBP and generalized Dirichlet distribution. Empirical results show that STM is able to discover subtle CCC in two benchmark code-bases, a feat which is beyond the scope of existing topic models, thus demonstrating the potential of the model in automated concern discovery, a known difficult problem in Software Engineering. Furthermore it is observed that even in general text corpora STM outperforms the state of art in discovering subtle topics.

Framework based software tends to get bloated by accumulating optional features (or concerns) just-in-case they are needed. The good news is that such feature bloat need not always cause runtime execution bloat. The bad news is that often enough, only a few statements from an optional concern may cause execution bloat that may result in as much as 50% runtime overhead.

The presence of software bloat in large flexible software systems can hurt energy efficiency. However, identifying and mitigating bloat is fairly effort intensive. To enable such efforts to be directed where there is a substantial potential for energy savings, we investigate the impact of bloat on power consumption under different situations. We conduct the first systematic experimental study of the joint power-performance implications of bloat across a range of hardware and software configurations on modern server platforms. The study employs controlled experiments to expose different effects of a common type of Java runtime bloat, excess temporary objects, in the context of the SPECPower_ssj2008 workload. We introduce the notion of equi-performance power reduction to characterize the impact, in addition to peak power comparisons. The results show a wide variation in energy savings from bloat reduction across these configurations. Energy efficiency benefits at peak performance tend to be most pronounced when bloat affects a performance bottleneck and non-bloated resources have low energy-proportionality. Equi-performance power savings are highest when bloated resources have a high degree of energy proportionality.

Managing a combined store consisting of database data and file data in a robust and consistent manner is a challenge for database systems and content management systems. In such a hybrid system, images, videos, engineering drawings, etc. are stored as files on a file server while meta-data referencing/indexing such files is created and stored in a relational database to take advantage of efficient search. In this paper we describe solutions for two potentially problematic aspects of such a data management system: backup/recovery and data consistency. We present algorithms for performing backup and recovery of the DBMS data in a coordinated fashion with the files on the file servers. Our algorithms for coordinated backup and recovery have been implemented in the IBM DB2/DataLinks product [1]. We also propose an efficient solution to the problem of maintaining consistency between the content of a file and the associated meta-data stored in the DBMS from a reader's point of view without holding long duration locks on meta-data tables. In the model, an object is directly accessed and edited in-place through normal file system APIs using a reference obtained via an SQL Query on the database. To relate file modifications to meta-data updates, the user issues an update through the DBMS, and commits both file and meta-data updates together.