arXiv subjects

Torsten Hoefler

Publications and source records attributed to Torsten Hoefler.

At least 37 records · Page 2Linked to original sources

arXiv2026

Can AI Weather Models Predict Beyond Two Weeks? A Quantitative Benchmark and Analysis of Long Rollouts

While AI weather models excel at short-to-medium range forecasts (up to 15 days), they frequently suffer from ill-defined "instabilities" when rolled out over longer horizons. This work addresses the lack of a formal taxonomy by categorizing these failures into three distinct regimes: blow-up, drift, and loss of seasonality, through year-long rollouts of nine state-of-the-art AI weather models. Our analysis reveals that stability hinges on the treatment of small spatio-temporal scales: unstable models amplify high-frequency energy, while stable models act as denoisers when noise is added to their inputs. Far from reducing these models to mere stochastic parrots, our findings highlight that stable models generate unique weather trajectories, conditioned on the initial state. We verify our findings through ablation studies on architectural design choices, conducted using state-of-the-art Vision Transformer (ViT) AI weather model architectures.

cs.LG↗

arXiv2026

Demystifying NVSHMEM: A System-Level Analysis on Symmetric Memory and Device-Initiated Operations in GPU Communication

NVSHMEM is NVIDIA's OpenSHMEM-based PGAS communication library for GPU clusters, enabling GPU-initiated, one-sided communication through symmetric memory. Despite its growing adoption, a system-level understanding of its design and behavior remains scattered across documentation, source code, and application experience. This paper presents a concise study of NVSHMEM's programming model, implementation, and performance characteristics, focusing on symmetric memory, one-sided operations, and device-side collectives. We also examine DeepEP as a case study of NVSHMEM in performance-critical sparse deep learning workloads. Our analysis shows that NVSHMEM pioneered a device-side symmetric-memory programming model that enables fine-grained GPU-driven communication and is important for approaching the hardware performance limit. Overall, this work defines NVSHMEM's role as a systems building block, highlights its design tradeoffs, and identifies opportunities for improving GPU communication runtimes.

cs.DC↗

arXiv2026

Process Reward Agents for Steering Knowledge-Intensive Reasoning

Reasoning in knowledge-intensive domains remains challenging as intermediate steps are often not locally verifiable: unlike math or code, evaluating step correctness may require synthesizing clues across large external knowledge sources. As a result, subtle errors can propagate through reasoning traces, potentially never to be detected. Prior work has proposed process reward models (PRMs), including retrieval-augmented variants, but these methods operate post hoc, scoring completed trajectories, which prevents their integration into dynamic inference procedures. Here, we introduce Process Reward Agents (PRA), an inference-time method for providing domain-grounded, online, step-wise rewards to a frozen policy. In contrast to prior retrieval-augmented PRMs, PRA enables search-based decoding to rank and prune candidate trajectories at every generation step. Experiments on multiple medical reasoning benchmarks demonstrate that PRA consistently outperforms strong baselines, achieving 81.9% accuracy on MedQA with Qwen3-4B, a new state of the art at the 4B scale. Importantly, PRA generalizes to unseen frozen policy models ranging from 0.5B to 8B parameters, improving their accuracy by up to 25.7% without any policy model updates. More broadly, PRA suggests a paradigm in which frozen reasoners are decoupled from domain-specific reward modules, allowing the deployment of new backbones in complex domains without retraining.

cs.AI↗

arXiv2026

When Data Is Scarce: Scaling Sparse Language Models with Repeated Training

Scaling laws for dense LLMs under infinite data are well explored, but how sparsity interacts with limited data is not. In this work, we study sparse training in data-constrained regimes where limited unique tokens require multi-epoch training. Our experiments span models up to 1.92B parameters in the fitting set, sparsity up to 93.75%, unique data budgets up to 2.6B tokens, and total training tokens up to 41.6B over 16 epochs; we further validate extrapolation on held-out dense-equivalent models up to 7.68B parameters. We find that: 1. Sparse scaling in data-limited settings: We introduce a scaling law that models loss as a function of active parameters, unique tokens, data repetition, and sparsity, accurately predicting performance across compute and data budgets. 2. Delayed data saturation: sparse training postpones diminishing returns from repeated data, making multi-epoch training more effective. 3. Resource trade-offs: With fixed data, loss-optimal sparsity is moderate ~ 50%, while compute-optimal sparsity is higher and grows with data scale. Overall, sparsity is not just a tool for efficiency, but a mechanism for improving scaling trade-offs under data scarcity. Our code is available at: https://github.com/boqian333/sparse-dc-scaling.

cs.LG↗

arXiv2026

WUSH: Near-Optimal Adaptive Transforms for LLM Quantization

Quantizing LLM weights and activations is a standard approach for efficient deployment, but a few extreme outliers can stretch the dynamic range and amplify low-bit quantization errors. Prior transform-based mitigations (e.g., Hadamard rotations) are fixed and data-agnostic, and their optimality for quantization has remained unclear. We derive closed-form optimal linear blockwise transforms for joint weight-activation quantization under standard RTN AbsMax-scaled block quantizers, covering both integer and floating-point formats. The resulting construction, WUSH, combines a Hadamard backbone with a data-dependent second-moment component to form a non-orthogonal transform that is provably near-optimal for FP and INT quantizers under mild assumptions while admitting an efficient fused GPU implementation. Empirically, WUSH improves W4A4 accuracy over the strongest Hadamard-based baselines (e.g., on Llama-3.1-8B-Instruct in MXFP4, it gains +2.8 average points with RTN and +0.7 with GPTQ) while delivering up to 5.8$\times$ per-layer throughput over BF16 via FP4 MatMul. Source code is available at https://github.com/IST-DASLab/WUSH.

cs.LG↗

arXiv2026

Memory-Efficient LLM Training with Dynamic Sparsity: From Stability to Practical Scaling

Dynamic Sparse Training (DST) offers a promising paradigm for improving the training and inference efficiency of deep neural networks; however, we find that in large language model training, DST can suffer from optimization instability, manifested as loss spikes after topology updates. In this work, we show that the naive use of standard Adam-based optimizers leads to a cold-start issue for newly regrown parameters, resulting in excessively large updates and disrupted training dynamics. To address this issue, we propose Sparse Memory-Efficient Training (SMET), which stabilizes DST with optimizer warm-up and improves training progress through density-aware learning-rate scaling. SMET further reduces memory consumption by storing gradients and optimizer states only for active parameters. We provide a theoretical analysis of the update behaviors under SMET, showing improved optimization stability. Extensive experiments demonstrate that SMET enables stable, scalable, and memory-efficient sparse pre-training of LLMs, paving the way for sparse training as a practical alternative to dense training. Our code is publicly available at: https://github.com/QiaoXiao7282/SMET.

cs.LG↗

arXiv2026

Confounder Detection via Treatment Intent: A New Observational Study Design

Understanding the effects of interventions is central to scientific progress, with randomized controlled trials (RCTs) regarded as the gold standard for causal inference in many applied fields. However, RCTs are costly, time-consuming, and often constrained by ethical or practical limitations, motivating the need for causal methods able to draw conclusions from observational data. While such data is collected at ever larger scale, making its use for causal inference is often hindered by the fact that not all variables affecting treatment allocation and the outcome are observed: an issue known as unobserved confounding. In this paper, we introduce a new study design called confounder detection via treatment intent. The idea is to query a human expert who makes treatment decisions, and ask them to compare pairs of units proposed by a principled matching strategy, with the goal of eliciting unobserved variables that explain why treatment decisions differ. We provide a theoretical basis for such a procedure, ascertaining conditions under which such a study design may elicit unobserved confounders. Building on this newly established foundations, we study treatment effects of interventions in the intensive care unit (ICU). First, we show empirical evidence strongly indicating that electronic health records (EHRs) collected in ICUs are subject to unobserved confounding. By using clinical text notes as a proxy for physicians' knowledge and leveraging natural language processing, we provide a proof of concept for our methodology in a semi-synthetic environment with a known ground truth.

stat.ME↗

arXiv2026

The Geometry of LLM Quantization: GPTQ as Babai's Nearest Plane Algorithm

Quantizing the weights of large language models (LLMs) from 16-bit to lower bitwidth is the de facto approach to deploy massive transformers onto more affordable accelerators. While GPTQ emerged as one of the standard methods for one-shot post-training quantization at LLM scale, its inner workings are described as a sequence of algebraic updates that obscure geometric meaning or worst-case guarantees. In this work, we show that, when executed back-to-front (from the last to first dimension) for a linear layer, GPTQ is mathematically identical to Babai's nearest plane algorithm for the classical closest vector problem (CVP) on a lattice defined by the Hessian matrix of the layer's inputs. This equivalence is based on a sophisticated mathematical argument, and has two analytical consequences: first, the GPTQ error propagation step gains an intuitive geometric interpretation; second, GPTQ inherits the error upper bound of Babai's algorithm under the assumption that no weights are clipped. Leveraging this bound, we design post-training quantization methods that avoid clipping, and outperform the original GPTQ. In addition, we provide efficient GPU inference kernels for the resulting representation. Taken together, these results place GPTQ on a firm theoretical footing and open the door to importing decades of progress in lattice algorithms towards the design of future quantization algorithms for billion-parameter models. Source code is available at https://github.com/IST-DASLab/GPTQ-Babai.

cs.LG↗

arXiv2026

Grid Games: The Power of Multiple Grids for Quantizing Large Language Models

A major recent advance in quantization is given by microscaled 4-bit formats such as NVFP4 and MXFP4, quantizing values into small groups sharing a scale, assuming a fixed floating-point grid. In this paper, we study the following natural extension: assume that, for each group of values, we are free to select the "better" among two or more 4-bit grids marked by one or more bits in the scale value. We formalize the power-of-two-grids (PO2) problem, and provide theoretical results showing that practical small-group formats such as MXFP or NVFP can benefit significantly from PO2 grids, while the advantage vanishes for very large groups. On the practical side, we instantiate several grid families, including 1) PO2(NF4), which pairs the standard NF4 normal grid with a learned grid, 2) MPO2, a grid pair that is fully learned over real weights and activations, 3) PO2(Split87), an explicit-zero asymmetric grid and 4) SFP4, a TensorCore-implementable triple which pairs NVFP4 with two shifted variants. Results for post-training quantization of standard open models and pre-training of Llama-like models show that adaptive grids consistently improve accuracy vs single-grid FP4 under both weight-only and weight+activation. Source code is available at https://github.com/IST-DASLab/GridGames.

cs.LG↗

arXiv2026

ADELIA: Automatic Differentiation for Efficient Laplace Inference Approximations

Spatio-temporal Bayesian inference drives environmental and health sciences using latent Gaussian models. Integrated Nested Laplace Approximations (INLA) enable inference for these models at HPC scale but rely on derivative-based optimization over $d$ hyperparameters. State-of-the-art INLA implementations approximate derivatives via central finite differences (FD), requiring $2d{+}1$ evaluations. These evaluations are embarrassingly parallel, but total work and energy grow with $d$, limiting time-to-solution under fixed budgets. Reverse-mode automatic differentiation (AD) computes exact gradients independently of $d$, but its efficient application to INLA's structured-sparse kernels is an open challenge. We present ADELIA, the first AD-enabled INLA implementation with a structure-exploiting multi-GPU backward pass leveraging model sparsity. We evaluate ADELIA on ten benchmark models, including real-world air-pollution monitoring. We achieve $4.2$--$7.9\times$ per-gradient speedups and reliable convergence on production-scale models with up to 1.9M latent variables, where FD struggles. Even when scaled to 16--32 GPUs to match ADELIA's wall-clock time, FD consumes $5$--$8\times$ more energy.

cs.DC↗

arXiv2026

Resilient AI Supercomputer Networking using MRC and SRv6

Tail latency dominates the performance of synchronous pretraining jobs when running at very large scales. We describe a three-pronged approach: (1) a new RDMA-based transport protocol, MRC, sprays across many paths and actively load-balances between them, eliminating the issue of flow collisions (2) the use of multi-plane Clos topologies to get the benefits of high switch radix and redundancy, allowing training clusters well over 100K GPUs to be built as two-tier topologies while increasing physical redundancy, and (3) the use of static source-routing using SRv6 to allow MRC the freedom to bypass failures by itself. We describe our experiences running MRC and static SRv6 routing in production in OpenAI and Microsoft's largest training clusters, where it has been used to train the latest frontier models. We demonstrate how MRC allows AI training jobs to ride out many network failures that previously would have interrupted training.

cs.NI↗

arXiv2026

SpaDA: A Spatial Dataflow Architecture Programming Language

Spatial dataflow architectures like the Cerebras Wafer-Scale Engine deliver exceptional performance in AI and scientific computing by distributing scratchpad memory across hundreds of thousands of processing elements (PEs). Yet programming these architectures remains difficult: with no shared memory, data movement requires explicit configuration, and asynchronous task management introduces substantial complexity. We present SpaDA, a programming language that offers precise control over data placement, dataflow patterns, and asynchronous operations while abstracting low-level architectural details. We design and implement a compiler targeting Cerebras CSL through multi-level lowering and unique optimization passes. SpaDA functions as a high-level programming interface and an intermediate representation for domain-specific languages (DSLs), demonstrated here with the GT4Py stencil DSL. SpaDA enables concise expression of operations with complex parallel patterns -- including pipelined collective operations, multi-dimensional stencils, and dense linear algebra -- in 14.09x fewer lines than CSL, achieving over 260 TFlop/s across 730,000 PEs on a single device.

cs.DC↗

arXiv2026

SFT-then-RL Outperforms Mixed-Policy Methods for LLM Reasoning

Recent mixed-policy optimization methods for LLM reasoning that interleave or blend supervised and reinforcement learning signals report improvements over the standard SFT-then-RL pipeline. We show that numerous recently published research papers rely on a faulty baseline caused by two distinct bugs: a CPU-offloaded optimizer bug in DeepSpeed that silently drops intermediate micro-batches during gradient accumulation (affecting multiple downstream frameworks including TRL, OpenRLHF and Llama-Factory), and a loss aggregation bug in OpenRLHF that incorrectly weights per-mini-batch losses. Together they suppress SFT performance, with the optimizer bug accounting for most of the gap and the loss aggregation bug contributing a smaller additional effect. Once corrected, the standard SFT-then-RL pipeline surpasses every published mixed-policy method we evaluate by +3.8 points on math benchmarks with Qwen2.5-Math-7B and by +22.2 points with Llama-3.1-8B. Even a truncated variant with just 50 RL steps outperforms mixed-policy methods on math benchmarks while using fewer FLOPs.

cs.LG↗

arXiv2026

Earth System Foundation Model (ESFM): A unified framework for heterogeneous data integration and forecasting

Foundation models (FMs) for the Earth system learn statistical relationships between physical variables across massive datasets to enable versatile downstream applications through finetuning, separating them from task-specific weather models. Here, we introduce Earth System Foundation Model (ESFM), a fully open model building on the 3D Swin UNet backbone of the pioneering Aurora model. ESFM introduces extensions that increase functionality and foster adoption in climate sciences. First, the encoding scheme and training protocols have been extended to handle diverse datasets, including those containing missing values across all spatio-temporal dimensions such as satellite data, as well as station data, all under one backbone. Axial attention is introduced to capture inter-variable dependencies. As a result ESFM skillfully predicts variables in regions or on pressure levels where no data is present at the initial time, while preserving inter-variable relationships, for example between temperature, pressure, and humidity. Individual variable tokenization enables different sets of variables to be shuffled during training and simplifies the process of building extensions for new downstream tasks. Adaptive layer norm-based ensembles allow for a simple yet effective way to transform deterministic ESFM to a probabilistic FM. We present findings using dense gridded data (ERA5, CMIP6), regionally masked dense data, sparse gridded MODIS satellite data, and station data. Results demonstrate competitive or superior performance relative to state-of-the-art benchmarks. Case studies of Super Typhoon Doksuri (2023) and 2024 sudden stratospheric warming events show accurate positional and magnitude estimations of extreme weather. ESFM retains the strengths of previous foundation models, such as long-term stability, but facilitates application to a variety of downstream tasks.

physics.ao-ph↗

arXiv2026

An Engineering Journey Training Large Language Models at Scale on Alps: The Apertus Experience

Large Language Models (LLMs) have surged as a transformative technology for science and society, prompting governments worldwide to pursue sovereign AI capabilities that ensure data compliance and cultural representation. However, the associated capital costs and engineering complexity required to train these models have largely restricted such capabilities to the private sector, leaving a significant gap for public institutions. This paper details the engineering journey behind training Apertus, a fully open multilingual foundation model, on the Alps supercomputer. Representing a first-of-its-kind achievement for academia at the 70B parameter scale, we successfully deployed a massive pre-training campaign on one of Europe's largest systems for open science, powered by NVIDIA GH200 Grace Hopper Superchips. We detail the challenges encountered in readying HPC infrastructure for training AI models, from overcoming storage bottlenecks to stabilizing large-scale interconnects, and the lessons learned in transforming a supercomputer into a resilient software-defined Machine Learning Platform. Finally, we discuss the post-training requirements and evolution of our Machine Learning platform, outlining how this initial release lays the groundwork for a sustained, iterative operational capability, in particular for fine tuning foundation models, that extends well beyond a single model training run.

cs.DC↗

arXiv2026

PICO: Performance Insights for Collective Operations

Collective operations are cornerstones of both HPC applications and large-scale AI training and inference, yet benchmarking them in a systematic and reproducible way remains difficult on modern systems due to the complexity of their hardware and software stacks. Existing suites primarily report end-to-end timings and offer limited support for controlled algorithm and configuration selection, fine-grained profiling, and capturing the runtime environment. We present PICO (Performance Insights for Collective Operations), an open-source framework that decouples portable experiment setup from platform execution, provides a backend-adaptive parameter selection interface across MPI and NCCL, supplies plain-MPI reference collective implementations, optionally instrumentable, and records the system configuration for reproducible comparisons. Evaluated on three major supercomputers, PICO shows that default collective algorithms and transport settings can be up to $5\times$ slower than the best available choice. It provides diagnostic evidence by isolating topology sensitive algorithmic choices and, through instrumentation, reveals detailed algorithmic breakdowns. To assess end-to-end effects of benchmark-informed tuning and evaluate application-level impacts, we replay open-source LLM training traces in ATLAHS simulator with optimized collective profiles identified by PICO, achieving reductions in training times of up to $44\%$.

cs.DC↗

arXiv2026

Benchmarking Filtered Approximate Nearest Neighbor Search Algorithms on Transformer-based Embedding Vectors

Advances in embedding models for text, image, audio, and video drive progress across multiple domains, including retrieval-augmented generation, recommendation systems, and others. Many of these applications require an efficient method to retrieve items that are close to a given query in the embedding space while satisfying a filter condition based on the item's attributes, a problem known as filtered approximate nearest neighbor search (FANNS). By performing an in-depth literature analysis on FANNS, we identify a key gap in the research landscape: publicly available datasets with embedding vectors from state-of-the-art transformer-based text embedding models that contain abundant real-world attributes covering a broad spectrum of attribute types and value distributions. To fill this gap, we introduce the arxiv-for-fanns dataset of transformer-based embedding vectors for the abstracts of over 2.7 million arXiv papers, enriched with 11 real-world attributes such as authors and categories. We benchmark eleven different FANNS methods on our new dataset to evaluate their performance across different filter types, numbers of retrieved neighbors, dataset scales, and query selectivities. We distill our findings into eight key observations that guide users in selecting the most suitable FANNS method for their specific use cases.

cs.DB↗

arXiv2026

Demystifying Chains, Trees, and Graphs of Thoughts

The field of natural language processing (NLP) has witnessed significant progress in recent years, with a notable focus on improving large language models' (LLM) performance through innovative prompting techniques. Among these, prompt engineering coupled with structures has emerged as a promising paradigm, with designs such as Chain-of-Thought, Tree of Thoughts, or Graph of Thoughts, in which the overall LLM reasoning is guided by a structure such as a graph. As illustrated with numerous examples, this paradigm significantly enhances the LLM's capability to solve numerous tasks, ranging from logical or mathematical reasoning to planning or creative writing. To facilitate the understanding of this growing field and pave the way for future developments, we devise a general blueprint for effective and efficient LLM reasoning schemes. For this, we conduct an in-depth analysis of the prompt execution pipeline, clarifying and clearly defining different concepts. We then build the first taxonomy of structure-enhanced LLM reasoning schemes. We focus on identifying fundamental classes of harnessed structures, and we analyze the representations of these structures, algorithms executed with these structures, and many others. We refer to these structures as reasoning topologies, because their representation becomes to a degree spatial, as they are contained within the LLM context. Our study compares existing prompting schemes using the proposed taxonomy, discussing how certain design choices lead to different patterns in performance and cost. We also outline theoretical underpinnings, relationships between prompting and other parts of the LLM ecosystem such as knowledge bases, and the associated research challenges. Our work will help to advance future prompt engineering techniques.

cs.CL↗