arXiv subjects

Rahul Mazumder

Publications and source records attributed to Rahul Mazumder.

At least 19 recordsLinked to original sources

arXiv2026

Convex Mixed-Integer Programming for Invariant Predictions from Multi-environment Data

Heterogeneous data provide an opportunity to identify causal predictors by revealing which predictive relationships remain invariant. Under a linear model, we introduce a least-squares coefficient-and-risk (LSCR) invariance principle, requiring both ordinary least squares coefficients and prediction risks to remain constant across environments. Our model neither specifies an intervention mechanism nor assumes a structural causal model for the entire system. We develop an invariance score that quantifies departures from LSCR invariance and use its optimization to construct two complementary procedures for causal predictor identification. First, we propose an estimator that jointly minimizes prediction risk and invariance penalty, and establish non-asymptotic guarantees for estimation and recovery of the causal parents under suitable identifiability conditions. Second, without requiring full identifiability, we develop a testing procedure that minimizes the invariance score over all predictor sets excluding a given variable, enabling valid causal discoveries with asymptotic error control via a multiplier bootstrap. Both procedures admit convex mixed-integer optimization formulations that avoid exhaustive subset enumeration while providing certificates of optimality. Simulations and real data experiments demonstrate substantial statistical and computational improvements over existing methods.

stat.ME↗

arXiv2026

Reconstruction-Optimized Expert Pruning for Mixture-of-Experts Language Models

Mixture-of-Experts (MoE) models reduce computation through sparse expert activation, but deployment still requires storing a large number of expert parameters. We study post-training expert pruning using task-domain calibration data as a subset-selection problem: given a pretrained MoE layer, select a fixed number of experts that best reconstruct its original output. Exact subset search is combinatorial and becomes intractable for modern MoE models with many experts per layer. We develop OptEP, a scalable continuous relaxation that directly optimizes differentiable expert-selection variables, followed by a lightweight reconstruction-based update of the retained expert weights. Across multiple MoE architectures and tasks spanning mathematical reasoning and code generation, our approach improves over the strongest existing expert-pruning baselines, with absolute accuracy gains of up to 15.05% on Qwen3-30B-A3B, 14.55% on Qwen3-235B-A22B, 11.06% on GLM-4.5-Air, and 16.20% on Mixtral-8x7B. The improvements are particularly pronounced under aggressive pruning, demonstrating that reconstruction-driven expert selection and weight adaptation can preserve substantially more of the original model's performance while reducing the number of retained experts.

cs.LG↗

arXiv2026

Two For The Price Of One: Curvature-Aware Pruning Preserves Influence Functions Too

Post-training pruning enables efficient deployment, but conventional methods do not explicitly preserve internal-gradient alignment between sparse and dense models. We study novel connections between downstream performance of compressed models and the ability to retain gradient information via influence-function preservation. Central to this connection is a curvature-aware objective that accounts for both layer activations and backpropagated gradients. Despite its promise, this objective remains underexplored in pruning due to a lack of efficient solvers. We introduce $\texttt{CACTUS}$, a novel ADMM-based optimizer supporting both pruning and sparse-plus-low-rank decompositions of pre-trained weights. Compared to LLM-Surgeon which targets the same objective, $\texttt{CACTUS}$ can achieve over an order-of-magnitude speedup. At $50\%$ pruning of Qwen3-4B, $\texttt{CACTUS}$ reduces the C4 perplexity gap to the dense model by $45\%$, improves compressed-to-dense per-sample gradient cosine similarity on WikiText-2 from $0.670$ to $0.755$, and runs over $28\times$ faster than LLM-Surgeon on substantially weaker hardware. By preserving influence functions, $\texttt{CACTUS}$ supports effective data selection. On Llama-3.2-3B, fine-tuning a fresh dense model on the $5\%$ of data selected by $\texttt{CACTUS}$ achieves $49.44\%$ accuracy on BIG-Bench Hard, compared with $47.22\%$ using ALPS and $49.54\%$ using dense-model selection. Finally, integrating $\texttt{CACTUS}$ with LoGRA, a recent gradient-projection strategy, on Qwen3-4B with MLP layers pruned to 2:4 sparsity and a rank-64 correction retains a $0.711$ cosine alignment with dense LoGRA-projected C4 gradients.

cs.LG↗

arXiv2026

3BASiL: An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs

Sparse plus Low-Rank $(\mathbf{S} + \mathbf{LR})$ decomposition of Large Language Models (LLMs) has emerged as a promising direction in model compression, aiming to decompose pre-trained model weights into a sum of sparse and low-rank matrices $(\mathbf{W} \approx \mathbf{S} + \mathbf{LR})$. Despite recent progress, existing methods often suffer from substantial performance degradation compared to dense models. In this work, we introduce 3BASiL-TM, an efficient one-shot post-training method for $(\mathbf{S} + \mathbf{LR})$ decomposition of LLMs that addresses this gap. Our approach first introduces a novel 3-Block Alternating Direction Method of Multipliers (ADMM) method, termed 3BASiL, to minimize the layer-wise reconstruction error with convergence guarantees. We then design an efficient transformer-matching (TM) refinement step that jointly optimizes the sparse and low-rank components across transformer layers. This step minimizes a novel memory-efficient loss that aligns outputs at the transformer level. Notably, the TM procedure is universal as it can enhance any $(\mathbf{S} + \mathbf{LR})$ decomposition, including pure sparsity. Our numerical experiments show that 3BASiL-TM reduces the WikiText2 perplexity gap relative to dense LLaMA-8B model by over 30% under a (2:4 Sparse + 64 LR) configuration, compared to prior methods. Moreover, our method achieves over 2.5x faster compression runtime on an A100 GPU compared to SOTA $(\mathbf{S} + \mathbf{LR})$ method. Our code is available at https://github.com/mazumder-lab/3BASiL.

cs.LG↗

arXiv2026

ThinQuant: Scalable Rotation Learning for Weight and Activation Quantization of LLMs

Learned rotations play an important role in enabling low-bit weight and activation quantization of large language models by smoothing outliers in the activation distribution. State-of-the-art approaches include gradient-based procedures such as SpinQuant and computationally friendlier gradient-free approaches such as DartQuant, but both remain hard to scale to the largest architectures. To address the computational bottlenecks in gradient-free rotation learning, we introduce two ideas for efficiency, (i) a data selection procedure which reduces the required number of calibration data points, and (ii) an exact reduction of the associated optimization on this reduced calibration set. Our data selection procedure exploits the geometric structure of the convex hull of the activations. Using this idea, we show that a carefully selected calibration set with several orders of magnitude fewer activations than state-of-the-art rotation-based methods can match their performance in low-bit quantization settings. Under this extreme data efficiency, the selected activations span an $r$-dimensional subspace with $r<d$, making optimization over a $d\times d$ rotation equivalent to optimizing a $d\times r$ matrix on the Stiefel manifold. We solve this reduced problem using an efficient ADMM algorithm that iteratively employs thin matrix updates at every step, hence the name ThinQuant. For Llama-3-70B with W4A4KV4 quantization, ThinQuant completes the entire rotation calibration in under 12 minutes and achieves a WikiText-2 perplexity of 5.63, compared with 7.55 for DartQuant, which requires 111 minutes. Unlike SpinQuant and DartQuant, ThinQuant also scales to Llama-3.1-405B on a single H200 GPU, completing rotation calibration in just over 2 hours and achieving WikiText-2 perplexity of 2.97 at W4A4, compared with 3.48 for GPTAQ+QuaRoT.

cs.LG↗

arXiv2026

A GPU-accelerated Nonlinear Branch-and-Bound Framework for Sparse Linear Models

We study exact sparse linear regression with an $\ell_0-\ell_2$ penalty and develop a branch-and-bound (BnB) algorithm explicitly designed for GPU execution. Starting from a perspective reformulation, we derive an interval relaxation that can be solved by ADMM with closed-form, coordinate-wise updates. We structure these updates so that the main work at each BnB node reduces to batched matrix-vector operations with a shared data matrix, enabling fine-grained parallelism across coordinates and coarse-grained parallelism across many BnB nodes on a single GPU. Feasible solutions (upper bounds) are generated by a projected gradient method on the active support, implemented in a batched fashion so that many candidate supports are updated in parallel on the GPU. We discuss practical design choices such as memory layout, batching strategies, and load balancing across nodes that are crucial for obtaining good utilization on modern GPUs. On synthetic and real high-dimensional datasets, our GPU-based approach achieves clear runtime improvements over a CPU implementation of our method, an existing specialized BnB method, and commercial MIP solvers.

math.OC↗

arXiv2026

MOONSHOT : A Framework for Multi-Objective Pruning of Vision and Large Language Models

Weight pruning is a common technique for compressing large neural networks. We focus on the challenging post-training one-shot setting, where a pre-trained model is compressed without any retraining. Existing one-shot pruning methods typically optimize a single objective, such as a layer-wise reconstruction loss or a second-order Taylor approximation of the training loss. We highlight that neither objective alone is consistently the most effective across architectures and sparsity levels. Motivated by this insight, we propose MOONSHOT, a general and flexible framework that extends any single-objective pruning method into a multi-objective formulation by jointly optimizing both the layer-wise reconstruction error and second-order Taylor approximation of the training loss. MOONSHOT acts as a wrapper around existing pruning algorithms. To enable this integration while maintaining scalability to billion-parameter models, we propose modeling decisions and introduce an efficient procedure for computing the inverse Hessian, preserving the efficiency of state-of-the-art one-shot pruners. When combined with state-of-the-art pruning methods on Llama-3.2 and Llama-2 models, MOONSHOT reduces C4 perplexity by up to 32.6% at 2:4 sparsity and improves zero-shot mean accuracy across seven classification benchmarks by up to 4.9 points. On Vision Transformers, it improves accuracy on ImageNet-1k by over 5 points at 70% sparsity, and on ResNet-50, it yields a 4-point gain at 90% sparsity.

cs.LG↗

arXiv2026

ADMM-Q: An Improved Hessian-based Weight Quantizer for Post-Training Quantization of Large Language Models

Quantization is an effective strategy to reduce the storage and computation footprint of large language models (LLMs). Post-training quantization (PTQ) is a leading approach for compressing LLMs. Popular weight quantization procedures, including GPTQ and RTN, suffer in model utility, especially at aggressive quantization levels (sub-4-bit). We propose ADMM-Q, a novel weight quantization algorithm that considers the layer-wise quantization problem. Our algorithm is based on a combinatorial variant of the Alternating Direction Method of Multipliers (ADMM). Our operator-splitting procedure updates weights continuously to minimize the layer-wise reconstruction error, while gradually enforcing the quantization constraints with convergence guarantees. We propose additional algorithmic enhancements (e.g., penalty scheduling, preconditioning, and a local search post-processing step) to make ADMM-Q efficient at LLM scale. ADMM-Q is modular and can be used as a drop-in replacement for any weight quantizer within existing quantization pipelines: ADMM-Q is fully composable with existing techniques including range clipping, learned or random rotations, and activation scaling. Using ADMM-Q in place of GPTQ on Qwen3-8B, we decrease WikiText-2 perplexity in: (i) the W3A16 weight-only setting (12.85 $\rightarrow$ 10.06); (ii) the W4A8 SmoothQuant procedure (9.29 $\rightarrow$ 8.68); and (iii) the W2A4KV4 SpinQuant procedure (66.11 $\rightarrow$ 19.42).

cs.LG↗

arXiv2026

Large-Scale Regularized Matching on GPU Clusters

Production decision systems such as ad allocation or content matching involve millions of users and thousands of items, reducing to large-scale linear programs with sparse block-diagonal structure across users. These LPs are solved repeatedly on recurring cadences over slowly evolving inputs. Three system gaps stand out. Scale: production instances routinely exceed the memory capacity of GPU solvers such as cuPDLP and D-PDLP under fixed hardware budgets. Temporal instability: solution variability across runs induces downstream churn and complicates SLAs, yet existing solvers provide no explicit control. Extensibility: CPU-based solvers such as DuaLip-Scala converge slowly and couple problem formulation to fixed schemas, making new constraint families difficult to express. We present a distributed multi-GPU LP solver built natively in PyTorch with systems-algorithm co-design for this structure. It adopts column-sharded parallelism with fused Triton kernels and batched operations to reduce per-iteration overhead. As users grow, only local computation increases, while communication is limited to a reduction of item-level dual variables, yielding near-linear scaling with GPU count at fixed item size. We also adopt ridge-regularized LPs to improve stability, a control absent from existing GPU solvers. A continuation schedule over the regularization parameter balances convergence speed and solution fidelity. Finally, we introduce an operator-centric programming model that replaces DuaLip-Scala's schema-bound interface with composable primitives, enabling new formulations without modifying the solve loop or distributed infrastructure. On synthetic workloads, our system achieves order-of-magnitude wall-clock speedup over DuaLip-Scala, near-linear multi-GPU scaling (3.86x on 4 GPUs), and scales beyond the reach of existing GPU solvers.

cs.DC↗

arXiv2026

Reasoning Models Can be Accurately Pruned Via Chain-of-Thought Reconstruction

Reasoning language models such as DeepSeek-R1 produce long chain-of-thought traces during inference time which make them costly to deploy at scale. We show that using compression techniques such as neural network pruning produces greater performance loss than in typical language modeling tasks, and in some cases can make the model slower since they cause the model to produce more thinking tokens but with worse performance. We show that this is partly due to the fact that standard LLM pruning methods often focus on input reconstruction, whereas reasoning is a decode-dominated task. We introduce a simple, drop-in fix: during pruning we jointly reconstruct activations from the input and the model's on-policy chain-of-thought traces. This "Reasoning-Aware Compression" (RAC) integrates seamlessly into existing pruning workflows such as SparseGPT, and boosts their performance significantly. Code reproducing the results in the paper can be found at: https://github.com/RyanLucas3/RAC

cs.AI↗

arXiv2026

Computation of Least Trimmed Squares: A Branch-and-Bound framework with Hyperplane Arrangement Enhancements

We study computational aspects of a key problem in robust statistics -- the penalized least trimmed squares (LTS) regression problem, a robust estimator that mitigates the influence of outliers in data by capping residuals with large magnitudes. Although statistically attractive, penalized LTS is NP-hard, and existing mixed-integer optimization (MIO) formulations scale poorly due to weak relaxations and exponential worst-case complexity in the number of observations. We propose a new MIO formulation that embeds hyperplane arrangement logic into a perspective reformulation, explicitly enforcing structural properties of optimal solutions. We show that, if the number of features is fixed, the resulting branch-and-bound tree is of polynomial size in the sample size. Moreover, we develop a tailored branch-and-bound algorithm that uses first-order methods with dual bounds to solve node relaxations efficiently. Computational experiments on synthetic and real datasets demonstrate substantial improvements over existing MIO approaches: on synthetic instances with 5000 samples and 20 features, our tailored solver reaches a 1% gap in 1 minute while competing approaches fail to do so within one hour. These gains enable exact robust regression at significantly larger sample sizes in low-dimensional settings.

math.OC↗

arXiv2026

Sparse Gaussian Graphical Models with Discrete Optimization: Computational and Statistical Perspectives

We consider the problem of learning a sparse graph underlying an undirected Gaussian graphical model, a key problem in statistical machine learning. Given $n$ samples from a multivariate Gaussian distribution with $p$ variables, the goal is to estimate the $p \times p$ inverse covariance matrix (aka precision matrix), assuming it is sparse (i.e., has a few nonzero entries). We propose GraphL0BnB, a new estimator based on an $\ell_0$-penalized version of the pseudo-likelihood function, while most earlier approaches are based on the $\ell_1$-relaxation. Our estimator can be formulated as a convex mixed integer program (MIP) which can be difficult to compute beyond $p\approx 100$ using off-the-shelf commercial solvers. To solve the MIP, we propose a custom nonlinear branch-and-bound (BnB) framework that solves node relaxations with tailored first-order methods. As a key component of our BnB framework, we propose large-scale solvers for obtaining good primal solutions that are of independent interest. We derive novel statistical guarantees (estimation and variable selection) for our estimator and discuss how our approach improves upon existing estimators. Our numerical experiments on real and synthetic datasets suggest that our BnB framework offers significant advantages over off-the-shelf commercial solvers, and our approach has favorable performance (both in terms of runtime and statistical performance) compared to the state-of-the-art approaches for learning sparse graphical models.

cs.LG↗

arXiv2026

Sparse PCA: A New Scalable Estimator Based On Integer Programming

We consider the Sparse Principal Component Analysis (SPCA) problem under the well-known spiked covariance model. Recent work has shown that the SPCA problem can be reformulated as a Mixed Integer Program (MIP) and can be solved to global optimality, leading to estimators that are known to enjoy optimal statistical properties. However, prior MIP algorithms for SPCA appear to be limited in terms of scalability to up to a thousand features or so. In this paper, we propose a new estimator for SPCA which can be formulated as a MIP. Different from earlier work, we make use of the underlying spiked covariance model and properties of the multivariate Gaussian distribution to arrive at our estimator. We establish statistical guarantees for our proposed estimator in terms of estimation error and support recovery. We derive guarantees under departures from the spiked covariance model, and for approximate solutions to the optimization problem. We propose a custom algorithm to solve the MIP, which scales better than off-the-shelf solvers, and demonstrate that our approach can be much more computationally attractive compared to earlier exact MIP-based approaches for the SPCA problem. Our numerical experiments on synthetic and real datasets show that our algorithms can address problems with up to 20,000 features in minutes; and generally result in favorable statistical properties compared to existing popular approaches for SPCA.

stat.ME↗

arXiv2026

Extracting Interpretable Models from Tree Ensembles: Computational and Statistical Perspectives

Tree ensembles are non-parametric methods widely recognized for their accuracy and ability to capture complex interactions. While these models excel at prediction, they are difficult to interpret and may fail to uncover useful relationships in the data. We propose an estimator to extract compact sets of decision rules from tree ensembles. The extracted models are accurate and can be manually examined to reveal relationships between the predictors and the response. A key novelty of our estimator is the flexibility to jointly control the number of rules extracted and the interaction depth of each rule, which improves accuracy. We develop a tailored exact algorithm to efficiently solve optimization problems underlying our estimator and an approximate algorithm for computing regularization paths, sequences of solutions that correspond to varying model sizes. We also establish novel non-asymptotic prediction error bounds for our proposed approach, comparing it to an oracle that chooses the best data-dependent linear combination of the rules in the ensemble subject to the same complexity constraint as our estimator. The bounds illustrate that the large-sample predictive performance of our estimator is on par with that of the oracle. Through experiments, we demonstrate that our estimator outperforms existing algorithms for rule extraction.

stat.ML↗

arXiv2026

Modeling with Categorical Features via Exact Fusion and Sparsity Regularisation

We study the high-dimensional linear regression problem with categorical predictors that have many levels. We propose a new estimation approach, which performs model compression via two mechanisms by simultaneously encouraging (a) clustering of the regression coefficients to collapse some of the categorical levels together; and (b) sparsity of the regression coefficients. We present novel mixed integer programming formulations for our estimator, and develop a custom row generation procedure to speed up the exact off-the-shelf solvers. We also propose a fast approximate algorithm for our method that obtains high-quality feasible solutions via block coordinate descent. As the main building block of our algorithm, we develop an exact algorithm for the univariate case based on dynamic programming, which can be of independent interest. We establish new theoretical guarantees for both the prediction and the cluster recovery performance of our estimator. Our numerical experiments on synthetic and real datasets demonstrate that our proposed estimator tends to outperform the state-of-the-art.

stat.ME↗

arXiv2026

Robust Batch-Level Query Routing for Large Language Models under Cost and Capacity Constraints

We study the problem of routing queries to large language models (LLMs) under cost, GPU resources, and concurrency constraints. Prior per-query routing methods often fail to control batch-level cost, especially under non-uniform or adversarial batching. To address this, we propose a batch-level, resource-aware routing framework that jointly optimizes model assignment for each batch while respecting cost and model capacity limits. We further introduce a robust variant that accounts for uncertainty in predicted LLM performance, along with an offline instance allocation procedure that balances quality and throughput across multiple models. Experiments on two multi-task LLM benchmarks show that robustness improves accuracy by 1-14% over non-robust counterparts (depending on the performance estimator), batch-level routing outperforms per-query methods by up to 24% under adversarial batching, and optimized instance allocation yields additional gains of up to 3% compared to a non-optimized allocation, all while strictly controlling cost and GPU resource constraints.

cs.LG↗

arXiv2026

DuaLip-GPU Technical Report

Large-scale linear programs (LPs) arise in many decision systems, including ranking, allocation, and matching problems that must be solved repeatedly at massive scale. Prior work such as ECLIPSE and LinkedIn's open-source DuaLip showed that ridge-regularized dual ascent with first-order methods can scale to these settings. However, the original implementation was tightly coupled to a small number of schemas and built on a CPU-centric Scala/Spark stack, limiting extensibility and preventing effective use of modern accelerators. We present a redesigned solver architecture that decouples problem specification from the optimization engine and targets GPU execution. The system uses an operator-centric programming model in which LP formulations are expressed through composable primitives for dual objective evaluation and blockwise projection operators for decomposable constraint families. This design allows new formulations to be added locally while reusing a shared optimization loop, diagnostics, and distributed infrastructure. To realize the available parallelism, we develop GPU execution techniques tailored to sparse matching constraints, including constraint-aligned sparse layouts, batched projection kernels, and a distributed design that communicates only dual variables. Further, we improve the underlying ridge-regularized dual ascent method with Jacobi-style row normalization, primal scaling, and a continuation scheme for the regularization parameter. On extreme-scale matching workloads, the GPU implementation achieves at least a 10x wall-clock speedup over the prior distributed CPU DuaLip solver under matched stopping criteria, while maintaining convergence guarantees.

cs.DC↗

arXiv2025

Theoretical Compression Bounds for Wide Multilayer Perceptrons

Pruning and quantization techniques have been broadly successful in reducing the number of parameters needed for large neural networks, yet theoretical justification for their empirical success falls short. We consider a randomized greedy compression algorithm for pruning and quantization post-training and use it to rigorously show the existence of pruned/quantized subnetworks of multilayer perceptrons (MLPs) with competitive performance. We further extend our results to structured pruning of MLPs and convolutional neural networks (CNNs), thus providing a unified analysis of pruning in wide networks. Our results are free of data assumptions, and showcase a tradeoff between compressibility and network width. The algorithm we consider bears some similarities with Optimal Brain Damage (OBD) and can be viewed as a post-training randomized version of it. The theoretical results we derive bridge the gap between theory and application for pruning/quantization, and provide a justification for the empirical success of compression in wide multilayer perceptrons.

cs.LG↗