Searcharxiv⌕ Search

arXiv subjects

Shubham Kumar

Publications and source records attributed to Shubham Kumar.

At least 19 recordsLinked to original sources
arXiv2026

Towards a Belief-Based World Model for LLM Agents

Large language models (LLMs) are being used as policies for autonomous decision-making and planning in many domains. Despite their strong reasoning capabilities, LLMs struggle with long-horizon tasks, especially under partial observability. World models are a promising way to enhance policy performance, both during training and inference. During inference, agents currently use world models to simulate the consequences of candidate actions before committing to an action, which can improve decision-making. However, we argue that simulation alone is an incomplete interface for decision-making under partial observability: simulation doesn't adequately capture uncertainty about the current state, which agents may need for accurate decision-making. We address this limitation with Belief-Based World Models (BB-WMs), which model and maintain a belief that LLMs can query to access information on what is known and uncertain about the current state. Before developing methods to learn accurate BB-WMs, we first ask a more fundamental question: does exposing a world model's belief directly to an LLM policy improve decision-making? Our results show that giving LLM agents access to world model beliefs improves task performance under partial observability, while remaining complementary to existing simulation-based world models. Code is released at https://github.com/skumar-ml/belief-world-models.

cs.AI↗
arXiv2026

Why We Created Yet Another Memory Framework: Understanding MGA's Role in Next-Gen Database Systems

Despite the presence of multiple memory regions in modern database systems, supporting an efficient form of memory remains a challenge under production constraints. In enterprise-grade data systems, existing abstractions impose a trade-off between coarse-grained global sharing and strict process isolation, resulting in data copying, memory fragmentation, and limited support for controlled sharing. These challenges become more pronounced as workloads grow more diverse, and systems must tolerate process failures while maintaining predictable performance. This paper introduces the Managed Global Area (MGA), a scoped shared-memory abstraction in Oracle AI Database that addresses these limitations. MGA allows components to explicitly define allocation source, membership, and coordination semantics across selected processes while integrating with a production database engine. Unlike fully shared memory regions in Oracle, such as the System Global Area (SGA), MGA supports dynamic process membership and modular memory usage without imposing system-wide visibility. We evaluate MGA on analytical and AI workloads that stress shared-memory execution, including TPC-H hash joins and ONNX Runtime inference. Under concurrent execution, MGA reduces latency for join-intensive TPC-H queries by up to 35%. For ONNX- based inference, MGA-enabled model sharing reduces memory footprint by up to 90% and lowers large-model inference latency by up to 37%. These results demonstrate that dynamically scoped shared memory can improve both efficiency and predictability in production database systems.

cs.DB↗
arXiv2026

Unveiling Structural Bottlenecks of Dynamic Disorder in a Density-Tunable Glass Former: From Strong to Fragile Regimes

Fragility characterizes how rapidly a glass-forming liquid slows down upon supercooling, but whether strong and fragile behaviors arise from the same microscopic relaxation mechanism remains unclear. Here, we address this question using a density-tunable soft-repulsive binary mixture spanning distinct fragility regimes and analyze particle jump dynamics within the framework of dynamic disorder. Across these regimes, we show that increasing fragility leads to progressively broader cage-lifetime distributions and increasingly non-exponential survival probabilities, revealing non-Poisson cage-to-jump statistics governed by fluctuating jump rates and slowly evolving structural variables. To characterize their structural origin, we first identify the neighbor ranks most strongly coupled to jump motion using Kullback-Leibler divergence and Pearson correlation analyses. We then introduce a structural slowness parameter that combines these neighbor-distance fluctuations into a reduced slow coordinate for constructing the slow-fluctuation survival probability. A comparison with the actual survival probability shows that localized neighbor-distance fluctuations control the jump rate in the strong regime, whereas extended neighbor rearrangements become relevant in the intermediate and fragile regimes, increasing the effective dimensionality of the slow-variable space. In the fragile regime, distance-based descriptors alone become insufficient at the lowest temperature, where the Voronoi free volume captures additional cage-volume fluctuations in the rate-controlling slow variable. Point-to-set correlations grow with fragility, but the spatial extent of the slow variables exceeds the point-to-set length. These results show that fragility changes the structural bottleneck for microscopic rate fluctuations, linking dynamic disorder and multidimensional slow variables.

cond-mat.soft↗
arXiv2026

A Fractional-Memory Physics-Informed Neural Network with Fast History Compression for Tempered Fractional Coupled Phase-Field Systems

Tempered time-fractional coupled phase-field (tTFCP) systems are used to model interfacial phenomena involving memory-dependent transport and relaxation mechanisms. Numerical solutions to these systems are challenging due to the simultaneous presence of nonlocal temporal operators, weak initial singularities, moving diffuse interfaces, and strongly coupled multiphysics dynamics. In this work, we introduce FM-tfPINN (fractional-memory physics-informed neural network), which is used for forward simulation and inverse parameter identification in tempered fractional coupled phase-field systems. Unlike conventional fractional PINNs, which enforce memory effects solely through residual constraints, our framework incorporates tempered fractional memory directly into the neural representation via latent memory-source functions and a tempered fractional integral operator. We develop a fast shifted residual formulation based on graded temporal meshes and sum-of-exponentials (SOE) history compression to efficiently evaluate the tempered fractional operators. This framework combines interface-aware and residual-adaptive collocation strategies, improving resolution near evolving diffuse interfaces. A unified, physics-informed loss formulation allows for the forward prediction and inverse recovery of unknown physical parameters from sparse observations. We assess the proposed method on a class of tempered fractional corrosion phase-field models, including one-dimensional corrosion-front propagation, activation- and diffusion-controlled regimes, two-dimensional pitting corrosion, and inverse mobility identification problems. The numerical results demonstrate the accurate recovery of coupled phase and concentration fields, the robust prediction of physically relevant interface diagnostics, and the reliable estimation of parameters from limited data.

math.NA↗
arXiv2026

Super-Arrhenius Dynamic Slowdown Revealed by Slow Variable Modulation in the Fragile Supercooled Liquid

The super-Arrhenius dynamic slowdown in fragile supercooled liquids remains one of the central unresolved questions in condensed matter physics. In this study, we analyze particle jump dynamics in a prototypical fragile glass-forming liquid, the Kob-Andersen Lennard-Jones (KALJ) model. Using the displacement of jumping particles as the reaction coordinate, we demonstrate the emergence of non-Poissonian dynamics as the temperature decreases. In the mildly supercooled regime, the outer region of the first coordination shell of a jumping particle exhibits a significant distribution shift during the jump motion. By comparing the survival probability with its slow-fluctuation limit using this distribution as a slow variable, we confirm that particles in this region modulate the jump dynamics, enhance the jump rate fluctuations, and thereby induce the dynamic slowdown as supercooling proceeds. As the temperature decreases, this behavior extends to the outer regions of the second coordination shell and beyond, intensifying the dynamic slowdown. This spatial growth of the slow variables responsible for dynamic disorder exhibits close correspondence with an increase in the static correlation length. These results provide a microscopic mechanism for the super-Arrhenius dynamic slowdown in the KALJ model.

cond-mat.soft↗
arXiv2026

Minimal, Local, Causal Explanations for Jailbreak Success in Large Language Models

Safety trained large language models (LLMs) can often be induced to answer harmful requests through jailbreak prompts. Because we lack a robust understanding of why LLMs are susceptible to jailbreaks, future frontier models operating more autonomously in higher-stakes settings may similarly be vulnerable to such attacks. Prior work has studied jailbreak success by examining the model's intermediate representations, identifying directions in this space that causally encode concepts like harmfulness and refusal. Then, they globally explain all jailbreak attacks as attempting to reduce or strengthen these concepts (e.g., reduce harmfulness). However, different jailbreak strategies may succeed by strengthening or suppressing different intermediate concepts, and the same jailbreak strategy may not work for different harmful request categories (e.g., violence vs. cyberattack); thus, we seek to give a local explanation -- i.e., why did this specific jailbreak succeed? To address this gap, we introduce LOCA, a method that gives Local, CAusal explanations of jailbreak success by identifying a minimal set of interpretable, intermediate representation changes that causally induce model refusal on an otherwise successful jailbreak request. We evaluate LOCA on harmful original-jailbreak pairs from a large jailbreak benchmark across Gemma, Llama, and Qwen chat models, comparing against prior methods adapted to this setting. LOCA can successfully induce refusal by making, on average, six interpretable changes; prior work routinely fails to achieve refusal even after 20 changes. LOCA is a step toward mechanistic, local explanations of jailbreak success in LLMs. Code publicly available at https://github.com/skumar-ml/loca-jailbreaks

cs.AI↗
arXiv2026

DHFP-PE: Dual-Precision Hybrid Floating Point Processing Element for AI Acceleration

The rapid adoption of low-precision arithmetic in artificial intelligence and edge computing has created a strong demand for energy-efficient and flexible floating-point multiply-accumulate (MAC) units. This paper presents a dual-precision floating-point MAC processing element supporting FP8 (E4M3, E5M2) and FP4 (2 x E2M1, 2 x E1M2) formats, specifically optimized for low-power and high-throughput AI workloads. The proposed architecture employs a novel bit-partitioning technique that enables a single 4-bit unit multiplier to operate either as a standard 4 x 4 multiplier for FP8 or as two parallel 2 x 2 multipliers for 2-bit operands, achieving maximum hardware utilization without duplicating logic. Implemented in 28 nm technology, the proposed PE achieves an operating frequency of 1.94 GHz with an area of 0.00396 mm^2 and power consumption of 2.13 mW, resulting in up to 60.4% area reduction and 86.6% power savings compared to state-of-the-art designs, making it well suited for energy-constrained AI inference and mixed-precision computing applications when deployed within larger accelerator architectures.

cs.AR↗
arXiv2026

Constant Depth Digital-Analog Counterdiabatic Quantum Computing

We introduce a digital-analog quantum computing framework that enables counterdiabatic protocols to be implemented at constant circuit depth, allowing fast and resource-efficient quantum state preparation on current quantum hardware. Counterdiabatic protocols suppress diabatic excitations in finite-time adiabatic evolution, but their practical application is limited by the non-local structure of the required Hamiltonians and the resource overhead of fully digital implementations. Counterdiabatic terms can be expressed as truncated expansions of nested commutators of the adiabatic Hamiltonian and its parametric derivative. Here, we show how this algebraic structure can be efficiently realized in a digital-analog setting using commutator product formulas. Using native multi-qubit analog interactions augmented by local single-qubit rotations, this approach enables higher-order counterdiabatic protocols whose implementation requires a constant number of analog blocks for any fixed truncation order, independent of system size. We demonstrate the method for two-dimensional spin models and analyze the associated approximation errors. These results show that digital-analog quantum computing enables a qualitatively new resource scaling for counterdiabatic protocols and related quantum control primitives, with direct implications for quantum simulation, optimization, and algorithmic state preparation on current quantum devices.

quant-ph↗
arXiv2025

Dynamic Slowdown and Spatial Correlations in Viscous Silica Melt: Perspectives from Dynamic Disorder

The dynamic slowdown in glass-forming liquids remains a central topic in condensed matter science. Here, we report a theoretical investigation of the microscopic origin of the slowdown in amorphous silica, a prototypical strong glass former with a tetrahedral network structure. Using molecular dynamics simulations, we analyze atomic jump dynamics, the elementary structural change processes underlying relaxation. We find that the jump statistics deviate from Poisson behavior with decreasing temperature, reflecting the emergence of dynamic disorder in which slowly evolving variables modulate the jump motion. The slowdown is species-dependent: for silicon, the primary constraint arises from the fourth-nearest oxygen neighbor, while at lower temperatures, the fourth-nearest silicon also becomes relevant; for oxygen, the dominant influence comes from the second-nearest silicon neighbors. As the system is cooled, the jump dynamics become increasingly slow and intermittent, proceeding in a higher-dimensional space of multiple slow variables that reflect cooperative rearrangements of the network. Species-resolved point-to-set correlations further reveal that the spatial extent of cooperative relaxation grows differently for silicon and oxygen, directly linking their relaxation asymmetry to the extent of collective motion. Together, these results provide a microscopic framework linking dynamic disorder, species-dependent constraints, and cooperative correlations, offering deeper insight into the slowdown of strong glass-forming networks.

cond-mat.soft↗
arXiv2025

Zero-Shot Grammar Competency Estimation Using Large Language Model Generated Pseudo Labels

Grammar competency estimation is essential for assessing linguistic proficiency in both written and spoken language; however, the spoken modality presents additional challenges due to its spontaneous, unstructured, and disfluent nature. Developing accurate grammar scoring models further requires extensive expert annotation, making large-scale data creation impractical. To address these limitations, we propose a zero-shot grammar competency estimation framework that leverages unlabeled data and Large Language Models (LLMs) without relying on manual labels. During training, we employ LLM-generated predictions on unlabeled data by using grammar competency rubric-based prompts. These predictions, treated as pseudo labels, are utilized to train a transformer-based model through a novel training framework designed to handle label noise effectively. We show that the choice of LLM for pseudo-label generation critically affects model performance and that the ratio of clean-to-noisy samples during training strongly influences stability and accuracy. Finally, a qualitative analysis of error intensity and score prediction confirms the robustness and interpretability of our approach. Experimental results demonstrate the efficacy of our approach in estimating grammar competency scores with high accuracy, paving the way for scalable, low-resource grammar assessment systems.

cs.CL↗
arXiv2025

Digitized Counterdiabatic Quantum Sampling

We propose digitized counterdiabatic quantum sampling (DCQS), a hybrid quantum-classical algorithm for efficient sampling from energy-based models, such as low-temperature Boltzmann distributions. The method utilizes counterdiabatic protocols, which suppress non-adiabatic transitions, with an iterative bias-field procedure that progressively steers the sampling toward low-energy regions. We observe that the samples obtained at each iteration correspond to approximate Boltzmann distributions at effective temperatures. By aggregating these samples and applying classical reweighting, the method reconstructs the Boltzmann distribution at a desired temperature. We define a scalable performance metric, based on the Kullback-Leibler divergence and the total variation distance, to quantify convergence toward the exact Boltzmann distribution. DCQS is validated on one-dimensional Ising models with random couplings up to 124 qubits, where exact results are available through transfer-matrix methods. We then apply it to a higher-order spin-glass Hamiltonian with 156 qubits executed on IBM quantum processors. We show that classical sampling algorithms, including Metropolis-Hastings and the state-of-the-art low-temperature technique parallel tempering, require up to three orders of magnitude more samples to match the quality of DCQS, corresponding to an approximately 2x runtime advantage. Boltzmann sampling underlies applications ranging from statistical physics to machine learning, yet classical algorithms exhibit exponentially slow convergence at low temperatures. Our results thus demonstrate a robust route toward scalable and efficient Boltzmann sampling on current quantum processors.

quant-ph↗
arXiv2025

Quantum Combinatorial Reasoning for Large Language Models

We design and implement a quantum combinatorial reasoning framework for large language models (QCR-LLM), integrating a real quantum computer in the hybrid workflow. QCR-LLM reformulates reasoning aggregation as a higher-order unconstrained binary optimization (HUBO) problem. In this sense, reasoning fragments are represented as binary variables and their interactions encode statistical relevance, logical coherence, and semantic redundancy. We tackle the resulting high-order optimization problem both classically, via simulated annealing, and quantumly through the bias-field digitized counterdiabatic quantum optimizer (BF-DCQO) executed on IBM's superconducting digital quantum processors. Experiments on BIG-Bench Extra Hard (BBEH) benchmarks demonstrate that our QCR-LLM consistently improves reasoning accuracy across multiple LLM backbones, surpassing reasoning-native systems such as o3-high and DeepSeek R1 by up to $+9\,$pp. Despite requiring multiple reasoning samples per query, our QCR-LLM remains approximately five times more energy-efficient than o3-high, owing to the low per-token energy footprint of its GPT-4o backbone. These results constitute the first experimental evidence of quantum-assisted reasoning, showing that hybrid quantum-classical optimization can efficiently enhance reasoning coherence, interpretability, and sustainability in large-scale language models. We have opened the doors to the emergence of quantum intelligence, where harder prompts require quantum optimizers at quantum-advantage level.

quant-ph↗
arXiv2025

A Geomechanically-Informed Framework for Wellbore Trajectory Prediction: Integrating First-Principles Kinematics with a Rigorous Derivation of Gated Recurrent Networks

Accurate wellbore trajectory prediction is a paramount challenge in subsurface engineering, governed by complex interactions between the drilling assembly and heterogeneous geological formations. This research establishes a comprehensive, mathematically rigorous framework for trajectory prediction that moves beyond empirical modeling to a geomechanically-informed, data-driven surrogate approach.The study leverages Log ASCII Standard (LAS) and wellbore deviation (DEV) data from 14 wells in the Gulfaks oil field, treating petrophysical logs not merely as input features, but as proxies for the mechanical properties of the rock that fundamentally govern drilling dynamics. A key contribution of this work is the formal derivation of wellbore kinematic models, including the Average Angle method and Dogleg Severity, from the first principles of vector calculus and differential geometry, contextualizing them as robust numerical integration schemes. The core of the predictive model is a Gated Recurrent Unit (GRU) network, for which we provide a complete, step-by-step derivation of the forward propagation dynamics and the Backpropagation Through Time (BPTT) training algorithm. This detailed theoretical exposition, often omitted in applied studies, clarifies the mechanisms by which the network learns temporal dependencies. The methodology encompasses a theoretically justified data preprocessing pipeline, including feature normalization, uniform depth resampling, and sequence generation. Trajectory post-processing and error analysis are conducted using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the Coefficient of Determination (R2).

physics.geo-ph↗
arXiv2025

Dendritic Computing with Multi-Gate Ferroelectric Field-Effect Transistors

Although inspired by neuronal systems in the brain, artificial neural networks generally employ point-neurons, which offer far less computational complexity than their biological counterparts. Neurons have dendritic arbors that connect to different sets of synapses and offer local non-linear accumulation - playing a pivotal role in processing and learning. Inspired by this, we propose a novel neuron design based on a multi-gate ferroelectric field-effect transistor that mimics dendrites. It leverages ferroelectric nonlinearity for local computations within dendritic branches, while utilizing the transistor action to generate the final neuronal output. The branched architecture paves the way for utilizing smaller crossbar arrays in hardware integration, leading to greater efficiency. Using an experimentally calibrated device-circuit-algorithm co-simulation framework, we demonstrate that networks incorporating our dendritic neurons achieve superior performance in comparison to much larger networks without dendrites ($\sim$17$\times$ fewer trainable weight parameters). These findings suggest that dendritic hardware can significantly improve computational efficiency, and learning capacity of neuromorphic systems optimized for edge applications.

cs.ET↗
arXiv2025

Measuring the (Un)Faithfulness of Concept-Based Explanations

Deep vision models perform input-output computations that are hard to interpret. Concept-based explanation methods (CBEMs) increase interpretability by re-expressing parts of the model with human-understandable semantic units, or concepts. Checking if the derived explanations are faithful -- that is, they represent the model's internal computation -- requires a surrogate that combines concepts to compute the output. Simplifications made for interpretability inevitably reduce faithfulness, resulting in a tradeoff between the two. State-of-the-art unsupervised CBEMs (U-CBEMs) are seemingly more interpretable, while also being more faithful to the model. However, we observe that the reported improvement in faithfulness artificially results from either (1) using overly complex surrogates, which introduces an unmeasured cost to the explanation's interpretability, or (2) relying on deletion-based approaches that, as we demonstrate, do not properly measure faithfulness. We propose Surrogate Faithfulness (SURF), which (1) replaces prior complex surrogates with a simple, linear surrogate that measures faithfulness without changing the explanation's interpretability and (2) introduces well-motivated metrics that assess loss across all output classes, not just the predicted class. We validate SURF with a measure-over-measure study by proposing a simple sanity check -- explanations with random concepts should be less faithful -- which prior surrogates fail. SURF enables the first reliable faithfulness benchmark of U-CBEMs, revealing that many visually compelling U-CBEMs are not faithful. Code is released at https://github.com/skumar-ml/surf-eval .

cs.LG↗
arXiv2025

Global Convergence and Rate Analysis of the Steepest Descent Method for Uncertain Multiobjective Optimization via a Robust Optimization Approach

In this article, we extend our previous work (Applicable Analysis, 2024, pp. 1-25) on the steepest descent method for uncertain multiobjective optimization problems. While that study established local convergence, it did not address global convergence and the rate of convergence of the steepest descent algorithm. To bridge this gap, we provide rigorous proofs for both global convergence and the linear convergence rate of the steepest descent algorithm. Global convergence analysis strengthens the theoretical foundation of the steepest descent method for uncertain multiobjective optimization problems, offering deeper insights into its efficiency and robustness across a broader class of optimization problems. These findings enhance the method's practical applicability and contribute to the advancement of robust optimization techniques.

math.OC↗
arXiv2025

Robust Optimization Approach for Solving Uncertain Multiobjective Optimization Problems Using the Projected Gradient Method

Numerous real-world applications of uncertain multiobjective optimization problems (UMOPs) can be found in science, engineering, business, and management. To handle the solution of uncertain optimization problems, robust optimization is a relatively new field. An extended version of the projected gradient method (PGM) for a deterministic smooth multiobjective optimization problem (MOP) is presented in the current study as a PGM for UMOP. An objective-wise worst-case cost (OWWC) type robust counterpart is considered, and the PGM is used to solve a UMOP by using OWWC. A projected gradient descent algorithm is created using theoretical findings. It is demonstrated that the projected gradient descent algorithm's generated sequence converges to the robust counterpart's weak Pareto optimal solution, which will be the robust weak Pareto optimal solution for UMOP. Under a few reasonable presumptions, the projected gradient descent algorithm's full convergent behavior is also justified. Finally, numerical tests are presented to validate the proposed method.

math.OC↗
arXiv2025

Solution of Uncertain Multiobjective Optimization Problems by Using Nonlinear Conjugate Gradient Method

This paper introduces a nonlinear conjugate gradient method (NCGM) for addressing the robust counterpart of uncertain multiobjective optimization problems (UMOPs). Here, the robust counterpart is defined as the minimum across objective-wise worst-case scenarios. There are some drawbacks to using scalarization techniques to solve the robust counterparts of UMOPs, such as the pre-specification and restrictions of weights, and function importance that is unknown beforehand. NCGM is free from any kind of priori chosen scalars or ordering information of objective functions as accepted in scalarization methods. With the help of NCGM, we determine the critical point for the robust counterpart of UMOP, which is the robust critical point for UMOP. To tackle this robust counterpart using the NCGM, the approach involves constructing and solving a subproblem to determine a descent direction. Subsequently, a new direction is derived based on parameter selection methods such as Fletcher-Reeves, conjugate descent, Dai-Yuan, Polak-Ribi$\grave{e}$re-Polyak, and Hestenes-Stiefel. An Armijo-type inexact line search is employed to identify an appropriate step length. Utilizing descent direction and step length, a sequence is generated, and convergence of the proposed method is established. The effectiveness of the proposed method is verified and compared against an existing method using a set of test problems.

math.OC↗