Searcharxiv⌕ Search

arXiv subjects

Michael W. Mahoney

Publications and source records attributed to Michael W. Mahoney.

At least 19 recordsLinked to original sources
arXiv2026

LLM Inference in a Flash!

Large Language Models (LLMs) have shown impressive capabilities across a range of natural language processing tasks, and LLM inference has emerged as a critical workload for enabling downstream applications. The demands of serving LLM inference are becoming increasingly challenging as requests shift toward longer sequences and heavier inference, driven by retrieval-augmented generation, inference-time compute scaling, and long-context applications. Additionally, these challenges are compounded by hardware trends, as memory capacity and communication bandwidth are not scaling as fast as increases in workload complexity. Compute-in-Flash is a promising solution to address memory bandwidth limitations by moving computation close to memory, and to exploit the large capacity of SSD technologies. However, it is challenging to deploy LLMs on these systems as they lack support for high-precision floating point operations and have limited write endurance. In our work, we aim to address these challenges by designing inference algorithms to enable LLM inference on Flash compute-in-memory devices. We present an end-to-end integer-only quantization approach to eliminate expensive floating-point computations. To address the limited write endurance, we design a dictionary-based KV cache compression strategy based on sparse dictionary coding that represents each KV vector as a linear combination of static dictionary vectors. These algorithmic improvements enable us to exploit the benefits of Compute-in-Flash for both model weights and KV cache, and to minimize expensive data transfer operations. Across Llama-3.1-8B and Qwen-2.5-7B, our combined method exhibits limited accuracy degradation while reducing dynamic KV cache traffic by 15$\times$.

cs.LG↗
arXiv2026

Eigenanalysis framework for autoregressive neural emulators of multi-scale chaotic dynamics

Neural autoregressive models have rapidly emerged as powerful emulators of high-dimensional chaotic systems, yet their long-term instability and error growth remain poorly understood, leading to ad-hoc solutions. Here, we develop an eigenanalysis framework that reveals the dynamical origin of this error growth. By analyzing the Jacobian of the learned one-step update map with respect to the state, we show how inference-time error growth, and thus model stability, is governed by its spectral radius. Direct-step architectures (models that predict the next state from the previous one) generically admit unstable eigenvalues with magnitudes exceeding one, explaining the rapid divergence of these widely used models. In contrast, integration-constrained models (where the time derivative is estimated and integrated with a higher-order integrator) collapse their eigenspectrum onto the unit circle, yielding neutral stability and a universal linear error-scaling law. The largest eigenvalue of this Jacobian provides an architecture-agnostic, a priori diagnostic of short-term skill, long-term stability, and spectral bias, without requiring an expensive rollout. Leveraging this theory, we introduce a stability-promoting loss that explicitly regularizes Jacobian-driven error amplification, improving both forecast accuracy and dynamical robustness. Demonstrated across $29$ models spanning two architectures, several explicit and implicit integrators, and multiple loss functions on the Kuramoto-Sivashinsky system, our results establish a theoretical foundation for the design and evaluation of neural emulators of chaotic multi-scale dynamics. More broadly, our framework is a step toward the kind of a priori stability analysis that numerical analysis provides for discretizations of differential equations and that scientific machine learning currently lacks.

cs.AI↗
arXiv2026

Agentic Test-Time Scaling for WebAgents

Test-time scaling has become a standard way to improve performance and boost reliability of neural network models. However, its behavior on agentic, multi-step tasks remains less well-understood: small per-step errors can compound over long horizons; and we find that naive policies that uniformly increase sampling show diminishing returns. In this work, we present CATTS, a simple technique for dynamically allocating compute for multi-step agents. We first conduct an empirical study of inference-time scaling for web agents. We find that uniformly increasing per-step compute quickly saturates in long-horizon environments. We then investigate stronger aggregation strategies, including an LLM-based Arbiter that can outperform naive voting, but that can overrule high-consensus decisions. We show that uncertainty statistics derived from the agent's own vote distribution (entropy and top-1/top-2 margin) correlate with downstream success and provide a practical signal for dynamic compute allocation. Based on these findings, we introduce Confidence-Aware Test-Time Scaling (CATTS), which uses vote-derived uncertainty to allocate compute only when decisions are genuinely contentious. CATTS improves performance on WebArena-Lite, Online-Mind2Web, and GoBrowse by up to 11.8% over majority voting while using fewer tokens than uniform scaling, providing both efficiency gains and an interpretable decision rule.

cs.AI↗
arXiv2026

LatentFlow: Visual Analytics for Latent Space Analysis in Molecular Graph Neural Networks

Chemists and materials scientists increasingly use machine learning models, such as graph neural networks (GNNs), to predict properties of molecules and the outcomes of their reactions. Beyond predictive performance, understanding how these models organize chemical information internally in their latent spaces, i.e., the embeddings of the molecules, is critical. Analyzing latent spaces helps diagnose model behavior and assess whether the learned embeddings are organized in ways that reflect meaningful chemical relationships. Unfortunately, existing methods provide limited support for analyzing latent spaces across layers and across different model states (e.g., training epochs, model configurations, and input data), making it difficult to understand how these latent spaces evolve throughout a model or relate to chemical concepts. We present LatentFlow, a visual analytics system developed in collaboration with a domain expert for analyzing latent spaces in molecular GNNs. LatentFlow groups embeddings into clusters and supports exploration of latent spaces by tracking how these clusters change across layers and model states using a modified Sankey diagram. To support interpretation, LatentFlow links these clusters to representative molecules and their shared substructures, and it allows scientists to introduce their own domain knowledge and compare it with the patterns found in the latent spaces. We evaluate LatentFlow through two case studies. The results show that LatentFlow helps scientists understand how latent spaces evolve, identify meaningful molecular patterns, and better interpret model behavior.

cs.LG↗
arXiv2026

Time-Aware Prior Fitted Networks for Zero-Shot Forecasting with Exogenous Variables

In many time series forecasting settings, the target time series is accompanied by exogenous covariates, such as promotions and prices in retail demand; temperature in energy load; calendar and holiday indicators for traffic or sales; and grid load or fuel costs in electricity pricing. Ignoring these exogenous signals can substantially degrade forecasting accuracy, particularly when they drive spikes, discontinuities, or regime and phase changes in the target series. Most current time series foundation models (e.g., Chronos, Sundial, TimesFM, TimeMoE, TimeLLM, and LagLlama) ignore exogenous covariates and make forecasts solely from the numerical time series history, thereby limiting their performance. In this paper, we develop ApolloPFN, a prior-data fitted network (PFN) that is time-aware (unlike prior PFNs) and that natively incorporates exogenous covariates (unlike prior univariate forecasters). Our design introduces two major advances: (i) a synthetic data generation framework that injects realistic temporal patterns, structural changes, and exogenous dependencies into the PFN prior; and (ii) time-aware architectural modifications that embed inductive biases needed to exploit temporal context. We demonstrate that ApolloPFN outperforms existing baselines across several forecasting benchmarks with exogenous information, including M5, electric price forecasting, UCI Air Quality, and Solar Energy datasets.

cs.LG↗
arXiv2026

Forking-Sequences: Statistically and Computationally Efficient Multi-Horizon Forecasting with Reduced Volatility

While accuracy is a critical requirement for time series forecasting, an equally important desideratum is reasonable forecast volatility across forecast creation dates (FCDs). Even highly accurate models can produce erratic revisions between FCDs, undermining trust and disrupting downstream decision-making. To improve the volatility of forecast revisions, state-of-the-art models like MQCNN, MQT, and SPADE employ a powerful yet underexplored neural network architectural design: forking-sequences. This architectural design jointly encodes and decodes the entire time series across all FCDs, producing an entire multi-horizon forecast grid in a single forward pass. This approach contrasts with conventional neural forecasting methods that process FCDs independently, generating only a single multi-horizon forecast per forward pass. In this work, we formalize the forking-sequences design and motivate its broader adoption by introducing a metric for quantifying excess volatility in forecast revisions and by providing theoretical and empirical analysis. We theoretically motivate three key benefits of forking-sequences: (i) reduced forecast volatility through ensembling; (ii) gradient variance reduction, improving the statistical efficiency of the training procedure; and (iii) improved inference computational efficiency. We validate the benefits of forking-sequences compared to baseline window-sampling on the M-series benchmark, using 16 datasets from the M1, M3, M4, and Tourism competitions. We observe median sCRPS improvements across datasets of 46.2%, 49.3%, 28.6%, 24.7%, and 6.4% for RNN, LSTM, CNN, Transformer, and State Space-based architectures, respectively. We then show that forecast ensembling during inference can reduce median forecast volatility by 13.2%, 13.0%, 10.9%, 10.2%, and 11.2% for these respective models trained with forking-sequences, while maintaining accuracy.

cs.LG↗
arXiv2026

Learning to Discover Iterative Spectral Algorithms

We introduce AutoSpec, a neural network framework for discovering iterative spectral algorithms for large-scale numerical linear algebra and numerical optimization. Our self-supervised models adapt to input operators using coarse spectral information (e.g., eigenvalue estimates and residual norms), and predict recurrence coefficients for computing or applying a matrix polynomial tailored to a downstream task. The effectiveness of AutoSpec relies on three ingredients: an architecture whose inference pass implements short, executable numerical linear algebra recurrences; efficient training on small synthetic problems with transfer to large-scale real-world operators; and task-defined objectives that enforce the desired approximation or preconditioning behavior across the range of spectral profiles represented in the training set. We apply AutoSpec to discovering algorithms for representative tasks on spd matrices: accelerating matrix function approximation; accelerating sparse linear solvers; and spectral filtering/preconditioning for eigenvalue computations. On real-world matrices, the learned procedures deliver up to order-of-magnitude improvements in accuracy and/or reductions in iteration count, relative to spectrum-agnostic baselines. We find clear connections to classical theory: the induced polynomials may exhibit equioscillation behavior characteristic of Chebyshev polynomial approximation. The code is available at: https://github.com/zihanghliu/AutoSpec .

cs.LG↗
arXiv2026

Random Matrix Theory for Deep Learning: Beyond Eigenvalues of Linear Models

Modern Machine Learning (ML) and Deep Neural Networks (DNNs) often operate on high-dimensional data and rely on overparameterized models, where classical low-dimensional intuitions break down. In particular, the proportional regime where the data dimension, sample size, and number of model parameters are all large and comparable, gives rise to novel and sometimes counterintuitive behaviors. This paper extends traditional Random Matrix Theory (RMT) beyond eigenvalue-based analysis of linear models to address the challenges posed by nonlinear ML models such as DNNs in this regime. We introduce the concept of High-dimensional Equivalent, which unifies and generalizes both Deterministic Equivalent and Linear Equivalent, to systematically address three technical challenges: high dimensionality, nonlinearity, and the need to analyze generic eigenspectral functionals. Leveraging this framework, we provide precise characterizations of the training and generalization performance of linear models, nonlinear shallow networks, and deep networks. Our results capture rich phenomena, including scaling laws, double descent, and nonlinear learning dynamics, offering a unified perspective on the theoretical understanding of deep learning in high dimensions.

stat.ML↗
arXiv2026

Eigen-Spike Emergence and Quadratic Equivalents for Conjugate Kernels on Nonlinearly Separable Data

Recent work in random matrix theory (RMT) has developed the notion of deterministic equivalents: typically linear surrogate models that approximate the spectral behavior of large nonlinear random matrices, such as nonlinear feature maps in neural networks (NNs). Such equivalents make theoretical predictions tractable by reducing a complex model to a simpler one with properties that fall under the umbrella of classical RMT tools. However, this leaves open the question of whether this idealized linear equivalence remains meaningful for classification of high-dimensional nonlinearly separable data. Motivated by this, we consider the conjugate kernel (CK), which is the nonlinear feature map of a one-layer feedforward NN, under a canonical nonlinearly separable dataset for the XOR problem; and we use the study of informative outlier eigenvalues in the CK and whether their corresponding eigenvectors asymptotically align with XOR labels as a proxy for nonlinear learnability. We develop a robust quadratic equivalent of the CK matrix that enables a precise analysis of emergent informative spikes, as one modifies various knobs common in ML practice: sample complexity, signal-to-noise ratio (SNR), nonlinear activation choice, and pretrained features. We identify regimes in which these knobs move the CK beyond the linear equivalent and produce BBP-type transitions to label-aligned outlier eigenspaces. Our analysis helps bring deterministic-equivalence tools from RMT to bear on problems of practical relevance in ML.

stat.ML↗
arXiv2026

Residual Context Diffusion Language Models

Diffusion Large Language Models (dLLMs) have emerged as a promising alternative to purely autoregressive language models because they can decode multiple tokens in parallel. However, state-of-the-art block-wise dLLMs rely on a "remasking" mechanism that decodes only the most confident tokens and discards the rest, effectively wasting computation. We demonstrate that recycling computation from the discarded tokens is beneficial, as these tokens retain contextual information useful for subsequent decoding iterations. In light of this, we propose Residual Context Diffusion (RCD), a module that converts these discarded token representations into contextual residuals and injects them back for the next denoising step. RCD uses a decoupled two-stage training pipeline to bypass the memory bottlenecks associated with backpropagation. We validate our method on both long CoT reasoning (SDAR) and short CoT instruction following (LLaDA) models. We demonstrate that a standard dLLM can be efficiently converted to the RCD paradigm with merely ~300 million tokens. RCD consistently improves frontier dLLMs by 4-11 percentage points in accuracy with minimal extra computation overhead across a wide range of benchmarks. Notably, on the most challenging AIME tasks, RCD nearly doubles baseline accuracy and attains up to 4-5x fewer denoising steps at baseline's peak accuracy.

cs.CL↗
arXiv2026

AlphaQ: Calibration-Free Bit Allocation for Mixture-of-Experts Quantization

Mixture-of-Experts (MoE) architectures scale model capacity through sparse expert activation, but their deployment remains memory-bound because all expert weights must reside in memory. Mixed-precision quantization can substantially reduce this footprint by assigning different bit-widths to different experts. Existing approaches, however, typically rely on calibration data to estimate expert importance and determine bit allocation. For frontier MoE LLMs, the original training data, and hence the true training distribution, is proprietary and inaccessible. As a result, calibration sets are inevitably imperfect surrogates, and this can misestimate expert utilization and lead to suboptimal bit allocation. Motivated by the substantial cross-expert quality variability observed in modern MoE models, and by the success of Heavy-Tailed Self-Regularization (HT-SR) theory at predicting neural network model quality without access to training or testing data, we propose AlphaQ, a calibration-free bit-allocation method for MoE quantization. AlphaQ draws on HT-SR theory and follows a simple principle: experts with more heavy-tailed weight spectra are typically better trained and hence should receive higher bit-widths, while experts with weaker heavy-tailed structure can be quantized more aggressively. AlphaQ operationalizes this principle by measuring expert-wise spectral heavy-tailedness and solving a budget-constrained optimization problem that minimizes total quantization error under a global bit-budget constraint. Across several MoE models, AlphaQ consistently outperforms calibration-based baselines under matched bit budgets. Notably, on Qwen1.5-MoE, AlphaQ achieves near full-precision accuracy with an average expert precision of only 3.5 bits, while delivering more than 4$\times$ memory compression. Our code is available at https://github.com/Superone77/AlphaQ.

cs.LG↗
arXiv2026

Unveiling Multi-regime Patterns in SciML: Distinct Failure Modes and Regime-specific Optimization

Neural networks trained under different hyperparameter settings can fall into distinct training "regimes," with consistent behavior within regimes and qualitative differences across regimes. In this paper, we study such multi-regime behavior in scientific machine learning (SciML) models through a regime-aware diagnostic framework that jointly analyzes performance, training dynamics, and loss-landscape geometry. We identify three key findings: (i) a consistent three-regime structure emerges across many standard SciML models, different constraint enforcements, and various optimizer designs; (ii) optimization effectiveness is regime-specific, with no single method performing well across all regimes; and (iii) SciML models can exhibit fine-grained failure modes that can challenge conventional interpretations of standard loss-landscape metrics. Our results provide an approach to establish a unified, task-oblivious perspective on failure modes in SciML and to inform regime-aware guidance for improving robustness. We validate these findings across widely-used SciML models, including physics-informed neural networks, neural operators, and neural ordinary differential equations, on benchmarks spanning representative ordinary and partial differential equations.

cs.LG↗
arXiv2026

Neural equilibria for long-term prediction of nonlinear conservation laws

Nonlinear conservation laws govern a broad class of important physical systems in science and industry and are central to scientific machine learning (SciML). Large general-purpose models offer speed, but replacing the numerical and physical structure of solvers often compromises stability, accuracy, and physical faithfulness. Here, we aim to balance the general inductive bias of conservation with the flexibility and speed of neural networks through a conservation-aware SciML backbone, which we call Neural Discrete Equilibrium (NeurDE). NeurDE places machine learning inside a kinetic solver by learning the local equilibrium closure of a Boltzmann formulation. The kinetic solver still performs transport, relaxation, moment recovery, and conservation; the neural network provides only the nonlinear equilibrium target. We test NeurDE on $6$ conserved systems, including three very challenging subsonic, transonic, and supersonic shock systems. NeurDE outperforms state-of-the-art SciML methods, including neural operators and pretrained SciML foundation models that are $10^4$ and $10^6$ times larger, respectively. Most notably, NeurDE improves upon the numerical method from which it is derived. NeurDE therefore provides a compact target for scientific machine learning in conservative simulation: learn the equilibrium law toward which the system relaxes, not the evolution law itself.

cs.LG↗
arXiv2026

Speculative Interaction Agents: Building Real-Time Agents with Asynchronous I/O and Speculative Tool Calling

There is a growing demand for agentic AI technologies for a range of downstream applications like customer service and personal assistants. For applications where the agent needs to interact with a person, real-time low-latency responsiveness is required; for example, with voice-controlled applications, under 1 second of latency is typically required for the interaction to feel seamless. However, if we want the LLM to reason and execute an agentic workflow with tool calling, this can add several seconds or more of latency, which is prohibitive for real-time latency-sensitive applications. In our work, we propose Speculative Interaction Agents to enable real-time interaction even for agents with complex multi-turn tool calling. We propose Asynchronous I/O, which decouples the core agent reason-and-act thread from waiting for additional information from either the user or environment, thereby allowing for overlapping agentic processing while waiting on external delays. We also propose Speculative Tool Calling as a method to manage task execution when the agent is still unsure if it has received the full information or if additional user information may later be provided. For strong cloud models, our method can be applied out-of-the-box to existing real-time cloud APIs, providing 1.3-1.7$\times$ speedups with minor accuracy loss. To enable real-time interaction with small edge-scale models, we also present a clock-based training methodology that adapts the model to handle streaming inputs and asynchronous responses, and demonstrate a synthetic data generation strategy for SFT. Altogether, this approach provides 1.6-2.2$\times$ speedups with the Qwen2.5-3B-Instruct and Llama-3.2-3B-Instruct models across multiple tool calling benchmarks.

cs.LG↗
arXiv2026

Zatom-1: Towards a Multimodal Foundation Model for 3D Molecules and Materials

General-purpose 3D modeling in chemistry encompasses molecules and materials, requiring both generative and predictive capabilities. However, most existing AI approaches are optimized for a single domain (molecules or materials) and a single task (generation or prediction), which limits representation sharing and transfer. We introduce Zatom-1, a cross-domain, general-purpose model architecture that unifies generative and predictive learning of 3D molecules and materials. Zatom-1 is a deliberately simplified Transformer trained with a multimodal flow matching objective that jointly models discrete atom types and continuous 3D geometries. This approach supports scalable pretraining with predictable gains as model capacity increases, while enabling fast and stable sampling. We use cross-domain generative pretraining as a universal initialization for downstream multi-task prediction of properties, energies, and forces. Empirically, Zatom-1 outperforms or competes with specialized baselines on both multi-task generative and predictive benchmarks in data-controlled settings, while improving generative inference speed by more than an order of magnitude. Our experiments demonstrate positive predictive transfer between data domains from joint generative pretraining: modeling materials during generative pretraining improves molecular property prediction accuracy. Open-source code and model weights are freely available at https://github.com/Zatom-AI/zatom.

cs.LG↗
arXiv2026

Is Flow Matching Just Trajectory Replay for Sequential Data?

Flow matching (FM) is increasingly used in scientific domains for time series generation and forecasting, where data often arise from underlying dynamical systems. However, it is not well-understood whether it learns transferable dynamical structure or simply performs an effective "trajectory replay". We study this question by deriving the velocity field targeted by the empirical FM objective on sequential data in the limit of perfect function approximation. For the Gaussian conditional paths commonly used in practice, we show that the implied sampler is an ODE whose dynamics constitutes a nonparametric, memory-augmented continuous-time dynamical system. The optimal field admits a closed-form expression as a similarity-weighted mixture of instantaneous velocities induced by observed transitions, making the dataset dependence explicit and interpretable. This characterization positions neural FM models as parametric surrogates of an ideal nonparametric solution and suggests practical approximation schemes for robust ODE-based generation. As a byproduct of our analysis, the resulting closed-form sampler, FreeFM, provides strong probabilistic forecasts on nonlinear dynamical system benchmarks directly from historical transitions, without training.

stat.ML↗
arXiv2026

Free Decompression with Algebraic Spectral Curves

Tools from random matrix theory have become central to deep learning theory, using spectral information to provide mechanisms for modeling generalization, robustness, scaling, and failure modes. While often capable of modeling empirical behavior, practical computations are limited by matrix size, often imposing a restriction to models that are too small to be realistic. This motivates the inference of properties of larger models from the behavior of smaller ones. Free decompression (FD) is a recently proposed method for extrapolating spectral information across matrix sizes, but its utility is currently limited by strong assumptions that preclude its implementation on more realistic machine learning (ML) models. We use algebraic spectral curve theory to provide a general FD methodology for spectral densities whose Stieltjes transform satisfies an algebraic relation, a modeling assumption that is more likely to hold in practice. This recasts FD as an evolution along spectral curves which can be readily integrated. Our framework enables the expansion of spectral densities that have multiple or multi-modal bulks, that exist at multiple scales, and that contain atoms, all characteristic of real-world data and popular ML models. We demonstrate the efficacy of our framework on models of interest in modern ML, including Hessian and activation matrices associated with neural networks and large-scale diffusion models.

stat.ML↗
arXiv2026

Foundation Models for Discovery and Exploration in Chemical Space

Accurate prediction of atomistic, thermodynamic, and kinetic properties from molecular structures underpins materials innovation. Existing computational and experimental approaches lack the scalability required to navigate chemical space efficiently. Scientific foundation models trained on large unlabelled datasets offer a path towards navigating chemical space across application domains. Here, we develop MIST, a family of molecular foundation models with up to an order of magnitude more parameters and data than prior works. Trained using a novel tokenizer, Smirk, which comprehensively captures nuclear, electronic, and geometric information, MIST learns a diverse range of molecules. MIST models have been fine-tuned to predict more than 400 structure-property relationships and have been shown to match or exceed state-of-the-art performance across diverse benchmarks, from physiology to electrochemistry. We demonstrate the ability of these models to solve real-world problems across chemical space from multiobjective electrolyte solvent screening to stereochemical reasoning for organometallics and mixture property prediction. The clearest demonstration of a foundation model is its ability to solve problems that were neither explicit targets of training nor central to the intentions of its developers. We identify olfactory perception mapping as such a problem, and show that MIST accurately predicted scent profiles and learned a hierarchical representation of olfactory space consistent with hyperbolic geometry. We formulated hyperparameter aware Bayesian neural scaling laws which eliminate the need for hyperparameter sweeps at every scale, making training large compute-optimal models feasible on a limited compute budget. The methods and findings presented here represent a significant step towards accelerating materials discovery, design, and optimization using foundation models.

physics.chem-ph↗