Manoj Ghuhan Arivazhagan

Vinayshekhar Bannihatti Kumar, Manoj Ghuhan Arivazhagan, Disha Makhija et al.

Prior KV cache compression schemes empirically demonstrate that the prompt cache is partially redundant during decoding, dropping or summarising entries with little accuracy loss. We ask when and what kind of redundancy: at which layers, after how many decoding steps, and in what form can the prompt span KV cache be replaced without breaking the task. A controlled splice intervention swept over layer cutoff and decoding steps shows this redundancy is about form (chat template scaffolding) rather than content. Replacing the upper layer prompt span KV cache with KV cache from a chat template scaffold whose user content is a neutral filler recovers near clean accuracy, while zeroing the same slots collapses accuracy. The dissociation replicates across the Qwen3, Gemma 3, and Llama 3 families on multiple datasets.

Vinayshekhar Bannihatti Kumar, Disha Makhija, Manoj Ghuhan Arivazhagan et al.

Large language models (LLMs) achieve high pass rates on code generation benchmarks, yet whether they can transfer this ability to languages absent from pretraining remains poorly understood. We introduce PyLang, a minimal imperative language absent from all pretraining corpora, and evaluate frontier models zero-shot and fine-tuned Qwen3 (4B, 8B, 32B) on 352 problems. We find that fine-tuning quickly teaches syntax but fails to transfer semantic competence: Python outperforms PyLang by up to 19% across all configurations, and no intervention (multi-task learning, preference tuning, code infilling, or latent-space objectives) closes the gap. An LLM judge reveals that frontier models select an identical algorithm to Python 80% of the time, yet cannot translate it into a working PyLang implementation., and CKA analysis confirms that fine-tuned models converge to nearly identical internal representations across languages (CKA > 0.97) while diverging at the output stage. We term this the implementation fidelity gap: models possess language-agnostic algorithmic understanding but cannot express it in an unfamiliar language. Our findings highlight the need for training methods that decouple reasoning from language-specific realization.

Nishanth Sridhar Nakshatri, Shamik Roy, Manoj Ghuhan Arivazhagan et al.

LLMs often fail to handle temporal knowledge conflicts--contradictions arising when facts evolve over time within their training data. Existing studies evaluate this phenomenon through benchmarks built on structured knowledge bases like Wikidata, but they focus on widely-covered, easily-memorized popular entities and lack the dynamic structure needed to fairly evaluate LLMs with different knowledge cut-off dates. We introduce evolveQA, a benchmark specifically designed to evaluate LLMs on temporally evolving knowledge, constructed from 3 real-world, time-stamped corpora: AWS updates, Azure changes, and WHO disease outbreak reports. Our framework identifies naturally occurring knowledge evolution and generates questions with gold answers tailored to different LLM knowledge cut-off dates. Through extensive evaluation of 12 open and closed-source LLMs across 3 knowledge probing formats, we demonstrate significant performance drops of up to 31% on evolveQA compared to static knowledge questions.

Disha Makhija, Manoj Ghuhan Arivazhagan, Vinayshekhar Bannihatti Kumar et al.

Membership inference attacks (MIAs) reveal whether specific data was used to train machine learning models, serving as important tools for privacy auditing and compliance assessment. Recent studies have reported that MIAs perform only marginally better than random guessing against large language models, suggesting that modern pre-training approaches with massive datasets may be free from privacy leakage risks. Our work offers a complementary perspective to these findings by exploring how examining LLMs' internal representations, rather than just their outputs, may provide additional insights into potential membership inference signals. Our framework, \emph{memTrace}, follows what we call \enquote{neural breadcrumbs} extracting informative signals from transformer hidden states and attention patterns as they process candidate sequences. By analyzing layer-wise representation dynamics, attention distribution characteristics, and cross-layer transition patterns, we detect potential memorization fingerprints that traditional loss-based approaches may not capture. This approach yields strong membership detection across several model families achieving average AUC scores of 0.85 on popular MIA benchmarks. Our findings suggest that internal model behaviors can reveal aspects of training data exposure even when output-based signals appear protected, highlighting the need for further research into membership privacy and the development of more robust privacy-preserving training techniques for large language models.

Manoj Ghuhan Arivazhagan, Vinay Aggarwal, Aaditya Kumar Singh et al.

The emerging paradigm of federated learning strives to enable collaborative training of machine learning models on the network edge without centrally aggregating raw data and hence, improving data privacy. This sharply deviates from traditional machine learning and necessitates the design of algorithms robust to various sources of heterogeneity. Specifically, statistical heterogeneity of data across user devices can severely degrade the performance of standard federated averaging for traditional machine learning applications like personalization with deep learning. This paper pro-posesFedPer, a base + personalization layer approach for federated training of deep feedforward neural networks, which can combat the ill-effects of statistical heterogeneity. We demonstrate effectiveness ofFedPerfor non-identical data partitions ofCIFARdatasetsand on a personalized image aesthetics dataset from Flickr.

Ramit Pahwa, Manoj Ghuhan Arivazhagan, Ankur Garg et al.

Deploying trained convolutional neural networks (CNNs) to mobile devices is a challenging task because of the simultaneous requirements of the deployed model to be fast, lightweight and accurate. Designing and training a CNN architecture that does well on all three metrics is highly non-trivial and can be very time-consuming if done by hand. One way to solve this problem is to compress the trained CNN models before deploying to mobile devices. This work asks and answers three questions on compressing CNN models automatically: a) How to control the trade-off between speed, memory and accuracy during model compression? b) In practice, a deployed model may not see all classes and/or may not need to produce all class labels. Can this fact be used to improve the trade-off? c) How to scale the compression algorithm to execute within a reasonable amount of time for many deployments? The paper demonstrates that a model compression algorithm utilizing reinforcement learning with architecture search and knowledge distillation can answer these questions in the affirmative. Experimental results are provided for current state-of-the-art CNN model families for image feature extraction like VGG and ResNet with CIFAR datasets.