Hewlett Packard Enterprise（慧与）

Recent FPGAs have improved deep learning (DL) inference efficiency through dedicated tensor blocks and in-BRAM computation. ReRAM-based analog in-memory computing (IMC) pushes efficiency further, offering an order-of-magnitude improvement in compute density and energy efficiency over conventional digital logic by performing vector-matrix multiplication (VMM) directly within the ReRAM crossbar; prior work has integrated such IMC blocks into FPGAs for DL inference. However, conventional IMC designs support only static-weight VMM, leaving nonlinear operations and dynamic matrix-matrix multiplication (DIMM) to the FPGA fabric. As a result, the benefits of IMC are largely confined to static-weight models, whereas Transformer-based models, which rely on frequent nonlinear and DIMM operations, gain only limited improvement. Moreover, the ADCs within each IMC block consume more than 70% of its area and power, further limiting system efficiency and scalability. To address these limitations, we propose a novel FPGA architecture that integrates an ADC-free IMC block, replacing the conventional ADC with analog content-addressable memories (ACAMs) that natively perform nonlinear operations inside the block. To fully exploit this block, we conduct an FPGA-aware design-space exploration that determines optimal crossbar dimensions while balancing FPGA area, flexibility, and DL performance, and we develop an efficient mapping that leverages ACAMs to carry out DIMM operations, extending the applicability of IMC to attention computation. On CNN and Transformer-based benchmarks, the proposed architecture achieves up to 40x and 1.9x higher energy efficiency and 4.1x and 2.5x higher area efficiency, respectively. Overall, it significantly improves FPGA DL inference efficiency and sustains robust gains on Transformer-based workloads across long input sequences, advancing domain-specialized FPGA design.

LLMs can be conveniently adapted to a diverse set of tasks, e.g, prediction, question-answering tasks, etc, using appropriate prompts with few-shot examples. Biased or harmful concepts, e.g. gender or bio-weapons, present in pre-trained LLMs can lead to unsafe or unethical responses for many such prompts. Removing such undesirable concepts robustly across different prompt types remains a challenging problem, since existing unlearning methods typically ignore the impact of prompt variation. In this paper, we explore a novel adversarial approach to use a joint prompt for the main task and concept task prediction. We show that fine-tuning using the ``worst prompt type'' for concept prediction (with the highest concept accuracy) improves the average unlearning performance over a fine-tuning method that uses a combination of all prompt types. Our proposed method, MPSelectTune, is a two-stage approach that minimizes the concept accuracy of the highest accuracy-prompt type, after fine-tuning using a novel multi-task loss using multiple prompt types. Experimental results on four benchmarks show $2 - 15\%$ main task accuracy improvements over recent baselines and while reducing the worst-case concept accuracy by up to $17\%$ compared to recent baselines.

Dynamic sparse attention (DSA) accelerates long-context LLM decoding by attending to only the top-K KV blocks relevant to each query, but it introduces a serialized selection-to-attention dependency that emerges as a new latency bottleneck. We present PRR, a speculate-reuse-repair runtime that exploits temporal locality in DSA selections to predict likely blocks, speculate the attention over them while selection is in flight, and incrementally repair missed blocks once the true selected set is known. PRR uses a lightweight EMA-based predictor, a profiling-guided speculation budget that keeps speculative work off the critical path, and a FlashAttention-based repair kernel that folds missed blocks into the partial attention state using online-softmax statistics. Across long-context benchmarks and representative DSA methods, PRR reduces per-token decoding latency by up to 40% while preserving downstream task accuracy. Github: https://github.com/Tianyu9748/Incremental_FlashAttention

Short pretraining runs can reduce experimental cost, but they can also over-promote configurations that only look strong at tiny budgets. We study an auditable staged-promotion protocol for a fixed micro-pretraining runner on two heterogeneous host blocks: Windows A100 and Linux L40S. Starting from twelve prior-screened configurations, we use staged budgets of 2 minutes, 5 minutes, 10 minutes, 60 minutes, and 12 hours, with frozen promotion rules before expensive continuations. The early screens are intentionally treated as unstable: the 5- and 10-minute rankings are host-sensitive, and the eventual 12-hour top-ranked condition is not the mean-best condition at the replicated 10-minute gate. Because seed ranges differ across stages, these changes are operational promotion evidence, not within-seed curves. A replicated 60-minute gate keeps the Staged Factorial Screening bridge reference in the promoted set, where it ranks first in all four 60-minute host-seed cells. In the final 12-hour confirmation package, the bridge condition ranks first in all four host-seed cells across two seeds; the greedy comparator does not meet the frozen 0.010 val_bpb near-equivalence rule; and the cheaper d8/ar48 (depth-8, aspect-48) sentinel does not meet the frozen 0.020 mean-gap rule. The executed 12-hour branch spends 144 GPU-hours, and the full staged protocol records 169.2 training GPU-hours including screening stages. Continuing all four 60-minute candidates would spend 192 GPU-hours, while continuing all nine replicated 10-minute candidates would spend 432 GPU-hours. The latter numbers are accounting counterfactuals for unrun continuations, not evidence that skipped candidates could not have overtaken the reference. The result is a bounded cost-allocation finding, not a claim of global optimality, capacity-normalized superiority, or superiority over adaptive hyperparameter optimization methods.

Budget-constrained micro-pretraining often requires triaging many candidate recipes on a shared accelerator before larger search budgets are spent. We study whether a staged fractional-factorial workflow can recover stable early effect structure in this setting. On a fixed autoresearch-derived single-GPU training loop, we run 613 experiments across pilot and follow-up screens at 2, 5, and 10 minutes; full 16-condition seeded reruns at 5 and 10 minutes; targeted seeded anchor checks; same-host greedy and matched-cost random baselines; a 60-minute bridge package; and bounded Windows A100 and Linux L40S anchor continuations through 24 hours. Main penalties from total batch, depth, and width are largest at short budgets and relax as budget increases. Within the predeclared seeded full-screen families, D, A, B, and C retain non-zero estimates at 5 and 10 minutes after within-budget Benjamini-Hochberg correction, while E does not. Random search can reach strong incumbents in this 32-condition space, but repeatedly in the same low-penalty region and without factor attribution. The 60-minute bridge anchor has the lowest mean, although that package does not separate workflow refinement from the larger bridge model's capacity advantage. In bounded 12-hour and 24-hour three-anchor continuations on both hosts, the bridge has the lowest sample mean while the non-bridge ordering stays host-sensitive. We therefore present a bounded methods result: use short designed screens to identify high-penalty directions, confirm promising anchors under repeated runs, and refine locally inside the reduced space. The evidence supports a bridge-centered recommendation through 24 hours on two hosts, not hardware-invariant ranking or general hyperparameter-optimization superiority.

Large language models (LLMs) frequently achieve impressive scores on standardized benchmarks, yet accuracy alone offers a limited view of their capabilities. Evaluating open-source LLMs through leaderboards faces persistent issues like data contamination, narrow task scope, and weak alignment with real-world reliability. Benchmark-based evaluations such as MMLU PRO, BBH, or IFEval primarily capture what a model outputs on fixed test sets, not how it processes information, calibrates uncertainty, or structures internal knowledge. In this article, we advocate for a shift from benchmark-centric evaluation toward a complementary, state-centered intrinsic assessment of LLMs. To this end, we introduce Latent Performance Profiling (LPP) -- a framework that derives task-agnostic diagnostics from hidden activations and output distributions. LPP defines a set of scalar metrics on a model's latent representations and dynamics, revealing scale-independent traits that enable interpretable comparisons and uncover hidden vulnerabilities. Unlike static accuracy scores, LPP provides stable, architecture-sensitive signatures across models of similar size. With extensive empirical analyses across eight LLMs, spanning a size range of 0.5B-14B, we demonstrate that models with similar benchmark scores can exhibit contrasting latent profiles, such as differences in entropy or adaptability. Guided by these insights, we design synthetic probes for uncertainty and symbolic reasoning that align with intrinsic metrics while decoupling from leaderboard bias. We recommend that reporting LPP alongside benchmarks provides a deeper, interpretable understanding of model behavior, enabling more reliable model selection, safety assessment, and evaluation beyond surface-level accuracy.

Prefix KV caching has become a key mechanism in LLM serving: it reduces time to first token (TTFT) by avoiding redundant computation across requests that share a prefix (i.e., the system prompt). However, the accumulated KV cache is often larger than what GPU memory and local DRAM can hold. To preserve latency, current systems keep the KV cache in remote DRAM pools, increasing serving-cluster size and cost. In this paper, we explore a different approach: storing the KV cache in S3-compatible object storage so that capacity is no longer the constraint, while minimizing the impact on TTFT. We propose ObjectCache, which co-designs the storage protocol and transfer schedule so that the storage server delivers KV cache data in the order the GPU consumes it, overlapping data transfer with compute across concurrent requests. We prototype ObjectCache on a 100 Gbps RoCE cluster with NIXL (an inference library that abstracts storage and memory), Ceph RGW (an Object Gateway for clusters), and DAOS (an open source storage system). For 64K contexts, common in today's systems, ObjectCache adds only 5.6\% latency over local DRAM; for 4K contexts, where less compute is available to mask transfer, ObjectCache adds 56--75\,ms over the optimal local layerwise baseline. Under shared bandwidth caps, our scheduler reduces added TTFT by 1.2--1.8x compared with equal bandwidth sharing.

A key trait of stochastic optimizers is that multiple runs of the same optimizer in attempting to solve the same problem can produce different results. As a result, their performance is evaluated over several repeats, or runs, on the problem. However, the accuracy of the estimated performance metrics depends on the number of runs and should be studied using statistical tools. We present a statistical analysis of the common metrics, and develop guidelines for experiment design to measure the optimizer's performance using these metrics to a high level of confidence and accuracy. To this end, we first discuss the confidence interval of the metrics and how they are related to the number of runs of an experiment. We then derive a lower bound on the number of repeats in order to guarantee achieving a given accuracy in the metrics. Using this bound, we propose an algorithm to adaptively adjust the number of repeats needed to ensure the accuracy of the evaluated metric. Our simulation results demonstrate the utility of our analysis and how it allows us to conduct reliable benchmarking as well as hyperparameter tuning and prevent us from drawing premature conclusions regarding the performance of stochastic optimizers.

Large Language Models (LLMs) have become increasingly prevalent in cloud-based platforms, propelled by the introduction of AI-based consumer and enterprise services. LLM inference requests in particular account for up to 90% of total LLM lifecycle energy use, dwarfing training energy costs. The rising volume of LLM inference requests is increasing environmental footprints, particularly carbon emissions and water consumption. To improve sustainability for LLM inference serving in cloud datacenter environments, we propose a novel multi-agent game-theoretic reinforcement learning framework called MARLIN to co-optimize time-to-first token (TTFT), carbon emissions, water usage, and energy costs associated with LLM inference. MARLIN demonstrates a reduction of at least 18% in TTFT, 33% in carbon emissions, 43% in water usage, and 11% in energy costs compared to state-of-the-art LLM inference management frameworks.

Analog content-addressable memories (aCAMs) based on memristors provide a promising pathway toward energy-efficient large-scale associative computing for Edge AI and embedded intelligence applications. They have been successfully applied to decision-tree inference and extend the capabilities of compute-in-memory (CIM) architectures beyond conventional vector-matrix multiplication. However, conventional designs such as the 6T2M architecture suffer from static search power, limited voltage gain, and pronounced match-line crosstalk, constraining analog precision and scalability. We introduce a strong-arm latched memristor (SALM) aCAM cell that replaces static voltage division with a dynamic current-race comparator, enabling high regenerative gain, intrinsic result latching, and near-zero static search power. Compared to 6T2M, SALM reduces read energy by 33% at identical latency while eliminating the gain and crosstalk limitations that prevent 6T2M from scaling to large arrays. SALM further enables scalable sequential and parallel latch sharing, and a dataset-aware optimization framework exposes an explicit energy-latency tradeoff, achieving up to 50% energy reduction at 3x latency across representative workloads. To enable architectural exploration, we develop a circuit-accurate behavioral model derived from SPICE lookup tables in 22 nm FD-SOI technology, capturing match-line dynamics and crosstalk. Integrated into the X-TIME decision-tree compiler, this framework demonstrates that SALM maintains near-software accuracy for high-dimensional datasets, whereas baseline designs degrade due to limited gain and cumulative crosstalk.

Inertial Confinement Fusion (ICF) holds transformative promise for sustainable, near-limitless clean energy, yet remains constrained by prohibitively high costs and limited experimental opportunities. This paper presents Human-in-the-Loop Meta Bayesian Optimization (HL-MBO), a framework that integrates expert knowledge with few-shot, uncertainty-aware machine learning to accelerate discovery in data-scarce, high-stakes scientific domains. HL-MBO introduces a meta-learned surrogate model with an expert-informed acquisition function to recommend candidate experiments. To foster trust and enable informed decisions, HL-MBO also provides interpretable explanations of its suggestions. We show HL-MBO outperforms current BO methods on ICF energy yield optimization, as well as benchmarks in molecular optimization and critical temperature maximization for superconducting materials.

Subtle visual anomalies such as hairline cracks, sub-millimeter voids, and low-contrast inclusions are structurally atypical yet visually ambiguous, making them both difficult to annotate and easy to overlook during active learning. Standard acquisition heuristics based on discriminative uncertainty or feature diversity often overselect dominant patterns while underexploring sparse yet important regions of the data space. This failure mode is especially severe in industrial defect inspection, where anomalies may be both low-prevalence and difficult to distinguish from surrounding structure. To resolve this, we propose GSAL, an active learning framework for object detection that combines a diffusion-based difficulty signal with a hierarchical semantic coverage prior. The diffusion component scores images and proposals using reconstruction discrepancy and denoising variability, prioritizing visually atypical or ambiguous examples. However, diffusion alone does not prevent acquisition from repeatedly favoring hard samples within dominant semantic modes. The semantic component therefore organizes candidate samples in a three-level concept graph and promotes coverage of underrepresented semantic regions while providing interpretable acquisition rationales. By balancing visual difficulty with semantic coverage, GSAL improves retrieval of subtle and rare targets that are often missed by uncertainty-only selection. Experiments on a proprietary thin-film defect, Pascal VOC and MS COCO dataset show consistent gains in label efficiency and rare-class retrieval over uncertainty-, diversity-, and hybrid-based baselines

Bayesian optimization (BO) has for sequential optimization of expensive black-box functions demonstrated practicality and effectiveness in many real-world settings. Meta-Bayesian optimization (meta-BO) focuses on improving the sample efficiency of BO by making use of information from related tasks. Although meta-BO is sample-efficient when task structure transfers, poor alignment between meta-training and test tasks can cause suboptimal queries to be suggested during online optimization. To this end, we propose a simple meta-BO algorithm that utilizes related-task information when determined useful, falling back to lookahead otherwise, within a unified framework. We demonstrate competitiveness of our method with existing approaches on function optimization tasks, while retaining strong performance in low task-relatedness regimes where test tasks share limited structure with the meta-training set.

Large Language Models (LLMs) increasingly produce natural language explanations alongside their predictions, yet it remains unclear whether these explanations reference predictive cues present in the input text. In this work, we present an empirical study of how LLM-generated explanations align with predictive lexical evidence from an external model in text classification tasks. To analyze this relationship, we compare explanation content against interpretable feature importance signals extracted from transparent linear classifiers. These reference models allow us to partition predictive lexical cues into supporting and contradicting evidence relative to the predicted label. Across three benchmark datasets-WIKIONTOLOGY, AG NEWS, and IMDB-we observe a consistent empirical pattern that we term support-contra asymmetry. Explanations accompanying correct predictions tend to reference more supporting lexical cues and fewer contradicting cues, whereas explanations associated with incorrect predictions reference substantially more contradicting evidence. This pattern appears consistently across datasets, across reference model families (logistic regression and linear SVM), and across multiple feature retrieval depths. These results suggest that LLM explanations often reflect lexical signals that are predictive for the task when predictions are correct, while incorrect predictions are more frequently associated with explanations that reference misleading cues present in the input. Our findings provide a simple empirical perspective on explanation-evidence alignment and illustrate how external sources of predictive evidence can be used to analyze the behavior of LLM-generated explanations.

AI weather prediction has advanced rapidly, yet no unified mathematical framework explains what determines forecast skill. Existing theory addresses specific architectural choices rather than the learning pipeline as a whole, while operational evidence from 2023-2026 demonstrates that training methodology, loss function design, and data diversity matter at least as much as architecture selection. This paper makes two interleaved contributions. Theoretically, we construct a framework rooted in approximation theory on the sphere, dynamical systems theory, information theory, and statistical learning theory that treats the complete learning pipeline (architecture, loss function, training strategy, data distribution) rather than architecture alone. We establish a Learning Pipeline Error Decomposition showing that estimation error (loss- and data-dependent) dominates approximation error (architecture-dependent) at current scales. We develop a Loss Function Spectral Theory formalizing MSE-induced spectral blurring in spherical harmonic coordinates, and derive Out-of-Distribution Extrapolation Bounds proving that data-driven models systematically underestimate record-breaking extremes with bias growing linearly in record exceedance. Empirically, we validate these predictions via inference across ten architecturally diverse AI weather models using NVIDIA Earth2Studio with ERA5 initial conditions, evaluating six metrics across 30 initialization dates spanning all seasons. Results confirm universal spectral energy loss at high wavenumbers for MSE-trained models, rising Error Consensus Ratios showing that the majority of forecast error is shared across architectures, and linear negative bias during extreme events. A Holistic Model Assessment Score provides unified multi-dimensional evaluation, and a prescriptive framework enables mathematical evaluation of proposed pipelines before training.

This paper summarizes the opportunities of in-network collective operations (INC) for accelerated collective operations in AI workloads. We provide sufficient detail to make this important field accessible to non-experts in AI or networking, fostering a connection between these communities. Consider two types of INC: Edge-INC, where the system is implemented at the node level, and Core-INC, where the system is embedded within network switches. We outline the potential performance benefits as well as six key obstacles in the context of both Edge-INC and Core-INC that may hinder their adoption. Finally, we present a set of predictions for the future development and application of INC.

We introduce a novel, closed-form approach for selective unlearning in multimodal models, specifically targeting pretrained models such as CLIP. Our method leverages nullspace projection to erase the target class information embedded in the final projection layer, without requiring any retraining or the use of images from the forget set. By computing an orthonormal basis for the subspace spanned by target text embeddings and projecting these directions, we dramatically reduce the alignment between image features and undesired classes. Unlike traditional unlearning techniques that rely on iterative fine-tuning and extensive data curation, our approach is both computationally efficient and surgically precise. This leads to a pronounced drop in zero-shot performance for the target classes while preserving the overall multimodal knowledge of the model. Our experiments demonstrate that even a partial projection can balance between complete unlearning and retaining useful information, addressing key challenges in model decontamination and privacy preservation.

Pretrained models like CLIP have demonstrated impressive zero-shot classification capabilities across diverse visual domains, spanning natural images, artistic renderings, and abstract representations. However, real-world applications often demand the removal (or "unlearning") of specific object classes without requiring additional data or retraining, or affecting the model's performance on unrelated tasks. In this paper, we propose a novel training- and data-free unlearning framework that enables three distinct forgetting paradigms: (1) global unlearning of selected objects across all domains, (2) domain-specific knowledge removal (e.g., eliminating sketch representations while preserving photo recognition), and (3) complete unlearning in selective domains. By leveraging a multimodal nullspace through synergistic integration of text prompts and synthesized visual prototypes derived from CLIP's joint embedding space, our method efficiently removes undesired class information while preserving the remaining knowledge. This approach overcomes the limitations of existing retraining-based methods and offers a flexible and computationally efficient solution for controlled model forgetting.

Optimizing or sampling complex cost functions of combinatorial optimization problems is a longstanding challenge across disciplines and applications. When employing family of conventional algorithms based on Markov Chain Monte Carlo (MCMC) such as simulated annealing or parallel tempering, one assumes homogeneous (equilibrium) temperature profiles across input. This instance independent approach was shown to be ineffective for the hardest benchmarks near a computational phase transition when the so-called overlap-gap-property holds. In these regimes conventional MCMC struggles to unfreeze rigid variables, escape suboptimal basins of attraction, and sample high-quality and diverse solutions. In order to mitigate these challenges, Nonequilibrium Nonlocal Monte Carlo (NMC) algorithms were proposed that leverage inhomogeneous temperature profiles thereby accelerating exploration of the configuration space without compromising its exploitation. Here, we employ deep reinforcement learning (RL) to train the nonlocal transition policies of NMC which were previously designed phenomenologically. We demonstrate that the resulting solver can be trained solely by observing energy changes of the configuration space exploration as RL rewards and the local minimum energy landscape geometry as RL states. We further show that the trained policies improve upon the standard MCMC-based and nonlocal simulated annealing on hard uniform random and scale-free random 4-SAT benchmarks in terms of residual energy, time-to-solution, and diversity of solutions metrics.