Searcharxiv⌕ Search

arXiv subjects

Kaushik Roy

Publications and source records attributed to Kaushik Roy.

At least 37 records · Page 2Linked to original sources
arXiv2026

TRIM: Token-wise Attention-Derived Saliency for Data-Efficient Instruction Tuning

Instruction tuning is essential for aligning large language models (LLMs) to downstream tasks and commonly relies on large, diverse corpora. However, small, high-quality subsets, known as coresets, can deliver comparable or superior results, though curating them remains challenging. Existing methods often rely on coarse, sample-level signals like gradients, an approach that is computationally expensive and overlooks fine-grained features. To address this, we introduce TRIM (Token Relevance via Interpretable Multi-layer Attention), a forward-only, token-centric framework. Instead of using gradients, TRIM operates by matching underlying representational patterns identified via attention-based "fingerprints" from a handful of target samples. Such an approach makes TRIM highly efficient and uniquely sensitive to the structural features that define a task. Coresets selected by our method consistently outperform state-of-the-art baselines by up to 9% on downstream tasks and even surpass the performance of full-data fine-tuning in some settings. By avoiding expensive backward passes, TRIM achieves this at a fraction of the computational cost. These findings establish TRIM as a scalable and efficient alternative for building high-quality instruction-tuning datasets.

cs.CL↗
arXiv2026

HEART: Hyperspherical Embedding Alignment via Kent-Representation Traversal in Diffusion Models

Text-to-image diffusion models can generate visually stunning images, yet, controlling what appears and how it appears, remains surprisingly difficult, especially when operating solely within the constraints of the text-conditioning space. For example, changing a subject or adjusting an attribute often leads to unintended side effects, such as altered backgrounds or distorted details. This is because most existing text-based control methods treat the embedding space as Euclidean and apply simple linear transformations, which do not reflect how semantic concepts are actually organized. In this work, we take a step back and ask: what is the true geometry of these embeddings? We find that text encoder representations lie on a hypersphere, where concepts are not linear directions but structured, anisotropic distributions better captured by Kent distributions. Building on this insight, we propose HEART, a training-free framework that performs Kent-aware geodesic transformations directly on the hypersphere. By respecting the underlying geometry, HEART enables intuitive and precise edits, such as consistent subject replacement and fine-grained attribute control, while preserving the original scene. Importantly, HEART requires no finetuning, inversion, or optimization, and generalizes across diffusion model architectures. Our results show that a simple shift in perspective, from linear to spherical, can unlock fast, and controllable image generation.

cs.CV↗
arXiv2026

REMAP: Regularized Matching and Partial Alignment of Video Embeddings

Real-world instructional videos are long, noisy, and often contain extended background segments, repeated actions, and execution variability that do not correspond to meaningful procedural steps. We propose **REMAP**, an unsupervised framework for procedure learning based on *Regularized Fused Partial Gromov-Wasserstein Optimal Transport*. REMAP relaxes balanced transport constraints, allowing non-informative or redundant frames to remain unmatched through partial transport. The formulation jointly models semantic similarity and temporal structure, while incorporating Laplacian-based smoothness and structural regularization to prevent degenerate alignments and reduce background interference. We evaluate REMAP on large-scale egocentric and third-person benchmarks. The method consistently outperforms state-of-the-art approaches, achieving up to **11.6\% (+4.45pp)** F1 and **19.6\% (+4.73pp)** IoU improvements on EgoProceL, and an average **41\% (+17.15pp)** F1 gain on ProceL and CrossTask. These results highlight the importance of partial alignment in handling real-world procedural variability and demonstrate that REMAP provides a robust and scalable approach for instructional video understanding.

cs.CV↗
arXiv2026

TopoPrune: Robust Data Pruning via Unified Latent Space Topology

Geometric data pruning methods, while practical for leveraging pretrained models, are fundamentally unstable. Their reliance on extrinsic geometry renders them highly sensitive to latent space perturbations, causing performance to degrade during cross-architecture transfer or in the presence of feature noise. We introduce TopoPrune, a framework which resolves this challenge by leveraging topology to capture the stable, intrinsic structure of data. TopoPrune operates at two scales, (1) utilizing a topology-aware manifold approximation to establish a global low-dimensional embedding of the dataset. Subsequently, (2) it employs differentiable persistent homology to perform a local topological optimization on the manifold embeddings, ranking samples by their structural complexity. We demonstrate that our unified dual-scale topological approach ensures high accuracy and precision, particularly at significant dataset pruning rates (e.g., 90%). Furthermore, through the inherent stability properties of topology, TopoPrune is (a) exceptionally robust to noise perturbations of latent feature embeddings and (b) demonstrates superior transferability across diverse network architectures. This study demonstrates a promising avenue towards stable and principled topology-based frameworks for robust data-efficient learning.

cs.LG↗
arXiv2026

CANDI: Contextual Alignment for Niche Domains Question Answering

The deployment of large language models (LLMs) in specialized domains like medical diagnostics and financial advisory necessitates evaluating capabilities beyond general knowledge. Traditional question-answering benchmarks often fail to capture the nuanced contextual grounding, user awareness, and domain understanding these fields require. To address this, we introduce CANDI-QA (Contextual Alignment for Niche Domains Question Answering), a novel dataset evaluating LLMs on delivering accurate, context-sensitive, and user-aligned answers in specialized settings. CANDI-QA features expert-curated question-answer pairs structured into two categories: (1) Information Assistance Questions, which are direct, factual queries requiring precise extraction, and (2) Applied Inference Questions, which are multi-hop reasoning tasks needing situational inference to generate actionable insights. We evaluate over ten diverse language models, from compact open-source to state-of-the-art proprietary systems. As a robust baseline, we present MTSS-Net, a lightweight neuro-symbolic framework combining neural retrieval with rule-based reasoning. Our findings highlight the profound challenges of achieving contextual alignment in niche domains, revealing the limitations of current LLMs without enhanced contextual or symbolic integration. Ultimately, CANDI-QA serves as a critical benchmark for advancing research in context-aware language models, stimulating the development of robust, trustworthy AI for high-stakes domains.

cs.CL↗
arXiv2026

2D-ThermAl: Physics-Informed Framework for Thermal Analysis of Circuits using Generative AI

Thermal analysis is increasingly critical in modern integrated circuits, where non-uniform power dissipation and high transistor densities can cause rapid temperature spikes and reliability concerns. Traditional methods, such as FEM-based simulations offer high accuracy but computationally prohibitive for early-stage design, often requiring multiple iterative redesign cycles to resolve late-stage thermal failures. To address these challenges, we propose 'ThermAl', a physics-informed generative AI framework which effectively identifies heat sources and estimates full-chip transient and steady-state thermal distributions directly from input activity profiles. ThermAl employs a hybrid U-Net architecture enhanced with positional encoding and a Boltzmann regularizer to maintain physical fidelity. Our model is trained on an extensive dataset of heat dissipation maps, ranging from simple logic gates (e.g., inverters, NAND, XOR) to complex designs, generated via COMSOL. Experimental results demonstrate that ThermAl delivers precise temperature mappings for large circuits, with a root mean squared error (RMSE) of only 0.71°C, and outperforms conventional FEM tools by running up to ~200 times faster. We analyze performance across diverse layouts and workloads, and discuss its applicability to large-scale EDA workflows. While thermal reliability assessments often extend beyond 85°C for post-layout signoff, our focus here is on early-stage hotspot detection and thermal pattern learning. To ensure generalization beyond the nominal operating range 25-55°C, we additionally performed cross-validation on an extended dataset spanning 25-95°C maintaining a high accuracy (<2.2% full-scale RMSE) even under elevated temperature conditions representative of peak power and stress scenarios.

cs.LG↗
arXiv2026

OASIS: Online Activation Subspace Learning for Memory-Efficient Training

Training large language models (LLMs) is constrained by memory requirements, with activations accounting for a substantial fraction of the total footprint. Existing approaches reduce memory using low-rank weight parameterizations or low-rank gradient subspaces for optimizer states, while activation memory is addressed through architectural modifications or compression schemes based on periodically updated projections. We propose OASIS, an online activation subspace learning algorithm for memory-efficient training that tracks and continuously updates a low-dimensional activation subspace during training. Intermediate activations are projected onto this evolving subspace, reducing memory without modifying forward-pass computations. The evolving activation subspace induces low-rank gradient representations, enabling both gradients and optimizer states to be maintained directly in this subspace, while a projection-aware optimizer consistently transports optimizer states across subspace updates for stable training. Across various finetuning and pretraining tasks, OASIS achieves up to $2\times$ lower peak memory than full fine-tuning while matching its performance and outperforming prior low-rank methods.

cs.LG↗
arXiv2026

Learning to Localize Leakage of Cryptographic Sensitive Variables

While cryptographic algorithms such as the ubiquitous Advanced Encryption Standard (AES) are secure, *physical implementations* of these algorithms in hardware inevitably 'leak' sensitive data such as cryptographic keys. A particularly insidious form of leakage arises from the fact that hardware consumes power and emits radiation in a manner that is statistically associated with the data it processes and the instructions it executes. Supervised deep learning has emerged as a state-of-the-art tool for carrying out *side-channel attacks*, which exploit this leakage by learning to map power/radiation measurements throughout encryption to the sensitive data operated on during that encryption. In this work we develop a principled deep learning framework for determining the relative leakage due to measurements recorded at different points in time, in order to inform *defense* against such attacks. This information is invaluable to cryptographic hardware designers for understanding *why* their hardware leaks and how they can mitigate it (e.g. by indicating the particular sections of code or electronic components which are responsible). Our framework is based on an adversarial game between a classifier trained to estimate the conditional distributions of sensitive data given subsets of measurements, and a budget-constrained noise distribution which probabilistically erases individual measurements to maximize the loss of this classifier. We demonstrate our method's efficacy and ability to overcome limitations of prior work through extensive experimental comparison on 6 publicly-available power/EM trace datasets from AES, ECC and RSA implementations. Our PyTorch code is available at https://github.com/jimgammell/learning_to_localize_leakage.

cs.LG↗
arXiv2026

LANCE: Low Rank Activation Compression for Efficient On-Device Continual Learning

On-device learning is essential for personalization, privacy, and long-term adaptation in resource-constrained environments. Achieving this requires efficient learning, both fine-tuning existing models and continually acquiring new tasks without catastrophic forgetting. Yet both settings are constrained by high memory cost of storing activations during backpropagation. Existing activation compression methods reduce this cost but rely on repeated low-rank decompositions, introducing computational overhead. Also, such methods have not been explored for continual learning. We propose LANCE (Low-rank Activation Compression), a framework that performs one-shot higher-order Singular Value Decomposition (SVD) to obtain a reusable low-rank subspace for activation projection. This eliminates repeated decompositions, reducing both memory and computation. Moreover, fixed low-rank subspaces further enable on-device continual learning by allocating tasks to orthogonal subspaces without storing large task-specific matrices. Experiments show that LANCE reduces activation storage up to 250$\times$ while maintaining accuracy comparable to full backpropagation on CIFAR-10/100, Oxford-IIIT Pets, Flowers102, and CUB-200 datasets. On continual learning benchmarks (Split CIFAR-100, Split MiniImageNet, 5-Datasets), it performs competitively with orthogonal gradient projection methods at a fraction of the memory cost. These results position LANCE as a practical and scalable solution for efficient fine-tuning and continual learning on edge devices.

cs.LG↗
arXiv2026

SPREAD: Subspace Representation Distillation for Lifelong Imitation Learning

A key challenge in lifelong imitation learning (LIL) is enabling agents to acquire new skills from expert demonstrations while retaining prior knowledge. This requires preserving the low-dimensional manifolds and geometric structures that underlie task representations across sequential learning. Existing distillation methods, which rely on L2-norm feature matching in raw feature space, are sensitive to noise and high-dimensional variability, often failing to preserve intrinsic task manifolds. To address this, we introduce SPREAD, a geometry-preserving framework that employs singular value decomposition (SVD) to align policy representations across tasks within low-rank subspaces. This alignment maintains the underlying geometry of multimodal features, facilitating stable transfer, robustness, and generalization. Additionally, we propose a confidence-guided distillation strategy that applies a Kullback-Leibler divergence loss restricted to the top-M most confident action samples, emphasizing reliable modes and improving optimization stability. Experiments on the LIBERO, lifelong imitation learning benchmark, show that SPREAD substantially improves knowledge transfer, mitigates catastrophic forgetting, and achieves state-of-the-art performance.

cs.LG↗
arXiv2026

WildCross: A Cross-Modal Large Scale Benchmark for Place Recognition and Metric Depth Estimation in Natural Environments

Recent years have seen a significant increase in demand for robotic solutions in unstructured natural environments, alongside growing interest in bridging 2D and 3D scene understanding. However, existing robotics datasets are predominantly captured in structured urban environments, making them inadequate for addressing the challenges posed by complex, unstructured natural settings. To address this gap, we propose WildCross, a cross-modal benchmark for place recognition and metric depth estimation in large-scale natural environments. WildCross comprises over 476K sequential RGB frames with semi-dense depth and surface normal annotations, each aligned with accurate 6DoF poses and synchronized dense lidar submaps. We conduct comprehensive experiments on visual, lidar, and cross-modal place recognition, as well as metric depth estimation, demonstrating the value of WildCross as a challenging benchmark for multi-modal robotic perception tasks. We provide access to the code repository and dataset at https://csiro-robotics.github.io/WildCross.

cs.CV↗
arXiv2026

TraceNAS: Zero-shot LLM Pruning via Gradient Trace Correlation

Structured pruning is essential for efficient deployment of Large Language Models (LLMs). The varying sensitivity of LLM sub-blocks to pruning necessitates the identification of optimal non-uniformly pruned models. Existing methods evaluate the importance of layers, attention heads, or weight channels in isolation. Such localized focus ignores the complex global structural dependencies that exist across the model. Training-aware structured pruning addresses global dependencies, but its computational cost can be just as expensive as post-pruning training. To alleviate the computational burden of training-aware pruning and capture global structural dependencies, we propose TraceNAS, a training-free Neural Architecture Search (NAS) framework that jointly explores structured pruning of LLM depth and width. TraceNAS identifies pruned models that maintain a high degree of loss landscape alignment with the pretrained model using a scale-invariant zero-shot proxy, effectively selecting models that exhibit maximal performance potential during post-pruning training. TraceNAS is highly efficient, enabling high-fidelity discovery of pruned models on a single GPU in 8.5 hours, yielding a 10$\times$ reduction in GPU-hours compared to training-aware methods. Evaluations on the Llama and Qwen families demonstrate that TraceNAS is competitive with training-aware baselines across commonsense and reasoning benchmarks.

cs.LG↗
arXiv2026

SlimDiff: Training-Free, Activation-Guided Hands-free Slimming of Diffusion Models

Diffusion models (DMs), lauded for their generative performance, are computationally prohibitive due to their billion-scale parameters and iterative denoising dynamics. Existing efficiency techniques, such as quantization, timestep reduction, or pruning, offer savings in compute, memory, or runtime but are strictly bottlenecked by reliance on fine-tuning or retraining to recover performance. In this work, we introduce SlimDiff, an automated activation-informed structural compression framework that reduces both attention and feedforward dimensionalities in DMs, while being entirely gradient-free. SlimDiff reframes DM compression as a spectral approximation task, where activation covariances across denoising timesteps define low-rank subspaces that guide dynamic pruning under a fixed compression budget. This activation-aware formulation mitigates error accumulation across timesteps by applying module-wise decompositions over functional weight groups: query--key interactions, value--output couplings, and feedforward projections, rather than isolated matrix factorizations, while adaptively allocating sparsity across modules to respect the non-uniform geometry of diffusion trajectories. SlimDiff achieves up to 35\% acceleration and $\sim$100M parameter reduction over baselines, with generation quality on par with uncompressed models without any backpropagation. Crucially, our approach requires only about 500 calibration samples, over 70$\times$ fewer than prior methods. To our knowledge, this is the first closed-form, activation-guided structural compression of DMs that is entirely training-free, providing both theoretical clarity and practical efficiency.

cs.LG↗
arXiv2026

ELLA: Efficient Lifelong Learning for Adapters in Large Language Models

Large Language Models (LLMs) suffer severe catastrophic forgetting when adapted sequentially to new tasks in a continual learning (CL) setting. Existing approaches are fundamentally limited: replay-based methods are impractical and privacy-violating, while strict orthogonality-based methods collapse under scale: each new task is projected onto an orthogonal complement, progressively reducing the residual degrees of freedom and eliminating forward transfer by forbidding overlap in shared representations. In this work, we introduce ELLA, a training framework built on the principle of selective subspace de-correlation. Rather than forbidding all overlap, ELLA explicitly characterizes the structure of past updates and penalizes alignments along their high-energy, task-specific directions, while preserving freedom in the low-energy residual subspaces to enable transfer. Formally, this is realized via a lightweight regularizer on a single aggregated update matrix. We prove this mechanism corresponds to an anisotropic shrinkage operator that bounds interference, yielding a penalty that is both memory- and compute-constant regardless of task sequence length. ELLA requires no data replay, no architectural expansion, and negligible storage. Empirically, it achieves state-of-the-art CL performance on three popular benchmarks, with relative accuracy gains of up to $9.6\%$ and a $35\times$ smaller memory footprint. Further, ELLA scales robustly across architectures and actively enhances the model's zero-shot generalization performance on unseen tasks, establishing a principled and scalable solution for constructive lifelong LLM adaptation.

cs.LG↗
arXiv2025

LIMO: Low-Power In-Memory-Annealer and Matrix-Multiplication Primitive for Edge Computing

Combinatorial optimization (CO) underpins applications in science and engineering, ranging from logistics to electronic design automation. A classic example is the NP-complete Traveling Salesman Problem (TSP). Finding exact solutions for large-scale TSP instances remains computationally intractable; on von Neumann architectures, such solvers are constrained by the memory wall, incurring compute-memory traffic that grows with instance size. Metaheuristics, such as simulated annealing implemented on compute-in-memory (CiM) architectures, offer a way to mitigate the von Neumann bottleneck. This is accomplished by performing in-memory optimization cycles to rapidly find approximate solutions for TSP instances. Yet this approach suffers from degrading solution quality as instance size increases, owing to inefficient state-space exploration. To address this, we present LIMO, a mixed-signal computational macro that implements an in-memory annealing algorithm with reduced search-space complexity. The annealing process is aided by the stochastic switching of spin-transfer-torque magnetic-tunnel-junctions (STT-MTJs) to escape local minima. For large instances, our macro co-design is complemented by a refinement-based divide-and-conquer algorithm amenable to parallel optimization in a spatial architecture. Consequently, our system comprising several LIMO macros achieves superior solution quality and faster time-to-solution on instances up to 85,900 cities compared to prior hardware annealers. The modularity of our annealing peripherals allows the LIMO macro to be reused for other applications, such as vector-matrix multiplications (VMMs). This enables our architecture to support neural network inference. As an illustration, we show image classification and face detection with software-comparable accuracy, while achieving lower latency and energy consumption than baseline CiM architectures.

cs.ET↗
arXiv2025

AgriRegion: Region-Aware Retrieval for High-Fidelity Agricultural Advice

Large Language Models (LLMs) have demonstrated significant potential in democratizing access to information. However, in the domain of agriculture, general-purpose models frequently suffer from contextual hallucination, which provides non-factual advice or answers are scientifically sound in one region but disastrous in another due to variations in soil, climate, and local regulations. We introduce AgriRegion, a Retrieval-Augmented Generation (RAG) framework designed specifically for high-fidelity, region-aware agricultural advisory. Unlike standard RAG approaches that rely solely on semantic similarity, AgriRegion incorporates a geospatial metadata injection layer and a region-prioritized re-ranking mechanism. By restricting the knowledge base to verified local agricultural extension services and enforcing geo-spatial constraints during retrieval, AgriRegion ensures that the advice regarding planting schedules, pest control, and fertilization is locally accurate. We create a novel benchmark dataset, AgriRegion-Eval, which comprises 160 domain-specific questions across 12 agricultural subfields. Experiments demonstrate that AgriRegion reduces hallucinations by 10-20% compared to state-of-the-art LLMs systems and significantly improves trust scores according to a comprehensive evaluation.

cs.AI↗
arXiv2025

GradientSpace: Unsupervised Data Clustering for Improved Instruction Tuning

Instruction tuning is one of the key steps required for adapting large language models (LLMs) to a broad spectrum of downstream applications. However, this procedure is difficult because real-world datasets are rarely homogeneous; they consist of a mixture of diverse information, causing gradient interference, where conflicting gradients pull the model in opposing directions, degrading performance. A common strategy to mitigate this issue is to group data based on semantic or embedding similarity. However, this fails to capture how data influences model parameters during learning. While recent works have attempted to cluster gradients directly, they randomly project gradients into lower dimensions to manage memory, which leads to accuracy loss. Moreover, these methods rely on expert ensembles which necessitates multiple inference passes and expensive on-the-fly gradient computations during inference. To address these limitations, we propose GradientSpace, a framework that clusters samples directly in full-dimensional gradient space. We introduce an online SVD-based algorithm that operates on LoRA gradients to identify latent skills without the infeasible cost of storing all sample gradients. Each cluster is used to train a specialized LoRA expert along with a lightweight router trained to select the best expert during inference. We show that routing to a single, appropriate expert outperforms expert ensembles used in prior work, while significantly reducing inference latency. Our experiments across mathematical reasoning, code generation, finance, and creative writing tasks demonstrate that GradientSpace leads to coherent expert specialization and consistent accuracy gains over state-of-the-art clustering methods and finetuning techniques.

cs.LG↗
arXiv2025

Experts are all you need: A Composable Framework for Large Language Model Inference

Large Language Models (LLMs) have achieved state-of-the-art accuracies in a variety of natural language processing (NLP) tasks. However, this success comes at the cost of increased model sizes which leads to additional computational burden. Mixture of Experts (MoEs) overcome this bottleneck by decoupling model capacity from computation by only activating a subset of parameters or "experts". However, these models require joint pretraining of these experts along with the router and do not model multi-step reasoning. In contrast, multi-agent frameworks improve reasoning by decomposing complex problems into modular subtasks. However, these frameworks rely on sequential "plan--act--observe" loops, which introduce significant latency. Our work, Comp-LLM, addresses these challenges by introducing a composable inference framework that enables cross-expert collaboration via an explicit sub-query dependency graph. Comp-LLM consists of three components: (1) A Sub-query Generator that decomposes an input query, assigns each sub-query to an appropriate expert using embedding similarity, and constructs a dependency graph; (2) A Query Executor that processes nodes in the graph and identifies opportunities for parallelism based on dependencies and resource constraints; and (3) A Response Aggregator that synthesizes intermediate expert responses into a coherent final answer. Across several benchmarks, Comp-LLM achieves up to 11.01% accuracy improvement over monolithic LLMs of similar size, while offering 1.67x--3.56x reduction in model size with no significant degradation relative to the largest model in its family. Additionally, Comp-LLM provides 1.1x--1.7x latency improvement compared to sequential sub-query processing.

cs.LG↗