Not All Errors Are Equal: A Systematic Study of Error Propagation in Large Language Model Inference

Large language models (LLMs) are increasingly integrated into high-performance computing (HPC) workflows, accelerating scientific discovery through diverse perspectives such as code generation and domain-specific decision-making. Yet, how soft errors propagate and affect LLM inference remains largely unexplored. To bridge this gap, we present a comprehensive study on error propagation in LLM inference, enabled by our proposed LLMFI, a configurable and deterministic fault-injection framework. Using LLMFI, we systematically inject faults across three open-weighted LLMs and thirteen representative tasks, covering reasoning, multilingual, mathematical, and coding domains. In addition, we conduct fine-grained case studies that reveal critical vulnerability patterns. Overall, our study yields 17 takeaways that advance the understanding of error propagation in LLM inference and introduces four low-overhead directions to improve reliability through software-only modification, offering practical guidance for future error detection and mitigation.

Faster Synchronous On-Policy RL via Straggler-Aware Group Sizing

Synchronous reinforcement learning methods such as Group Relative Policy Optimization (GRPO) provide stable and reproducible on-policy training, but they are highly vulnerable to stragglers, a single unusually long rollout can delay reward computation and parameter updates for the entire group. This problem becomes more severe as group size increases, creating a tension between the benefits of larger groups and the wall-clock cost of synchronization stalls. We propose Straggler-Aware Group Control (SAGC), a dynamic group-size controller that adapts the training group online based on observed rollout behavior. SAGC formulates group-size selection as an online constrained optimization problem, seeking to retain the benefits of larger groups while controlling the long-term rate of straggler events. Across synchronous GRPO and DAPO training, and on top of both vanilla and strong engineered baselines, SAGC consistently reduces straggler incidence and improves wall-clock efficiency while achieving competitive or better training reward. We further show that these gains transfer to final model quality: SAGC is competitive with or better than the strongest static group-size baseline on downstream reasoning benchmarks, and often produces shorter outputs without any explicit length penalty. These results position dynamic group control as a practical way to make synchronous on-policy RL more efficient and robust.

Preserving Clusters in Error-Bounded Lossy Compression of Particle Data

Lossy compression is widely used to reduce storage and I/O costs for large-scale particle datasets in scientific applications such as cosmology, molecular dynamics, and fluid dynamics, where clustering structures (e.g., single-linkage or Friends-of-Friends) are critical for downstream analysis; however, existing compressors typically provide only pointwise error bounds on particle positions and offer no guarantees on preserving clustering outcomes, and even small perturbations can alter cluster connectivity and compromise scientific validity. We propose a correction-based technique to preserve single-linkage clustering under lossy compression, operating on decompressed data from off-the-shelf compressors such as SZ3 and Draco. Our key contributions are threefold: (1) a clustering-aware correction algorithm that identifies vulnerable particle pairs via spatial partitioning and local neighborhood search; (2) an optimization-based formulation that enforces clustering consistency using projected gradient descent with a loss that encodes pairwise distance violations; and (3) a scalable GPU-accelerated and distributed implementation for large-scale datasets. Experiments on cosmology and molecular dynamics datasets show that our method effectively preserves clustering results while maintaining competitive compression performance compared with SZ3, ZFP, Draco, LCP, and space-filling-curve-based schemes.

LUMOS: Democratizing SciML Workflows with L0-Regularized Learning for Unified Feature and Parameter Adaptation

The rapid growth of scientific machine learning (SciML) has accelerated discovery across diverse domains, yet designing effective SciML models remains a challenging task. In practice, building such models often requires substantial prior knowledge and manual expertise, particularly in determining which input features to use and how large the model should be. We introduce LUMOS, an end-to-end framework based on L0-regularized learning that unifies feature selection and model pruning to democratize SciML model design. By employing semi-stochastic gating and reparameterization techniques, LUMOS dynamically selects informative features and prunes redundant parameters during training, reducing the reliance on manual tuning while maintaining predictive accuracy. We evaluate LUMOS across 13 diverse SciML workloads, including cosmology and molecular sciences, and demonstrate its effectiveness and generalizability. Experiments on 13 SciML models show that LUMOS achieves 71.45% parameter reduction and a 6.4x inference speedup on average. Furthermore, Distributed Data Parallel (DDP) training on up to eight GPUs confirms the scalability of

HII-DPO: Eliminate Hallucination via Accurate Hallucination-Inducing Counterfactual Images

Large Vision-Language Models (VLMs) have achieved remarkable success across diverse multimodal tasks but remain vulnerable to hallucinations rooted in inherent language bias. Despite recent progress, existing hallucination mitigation methods often overlook the underlying hallucination patterns driven by language bias. In this work, we design a novel pipeline to accurately synthesize Hallucination-Inducing Images (HIIs). Using synthesized HIIs, we reveal a consistent scene-conditioned hallucination pattern: models tend to mention objects that are highly typical of the scene even when visual evidence is removed. To quantify the susceptibility of VLMs to this hallucination pattern, we establish the Masked-Object-Hallucination (MOH) benchmark to rigorously evaluate existing state-of-the-art alignment frameworks. Finally, we leverage HIIs to construct high-quality preference datasets for fine-grained alignment. Experimental results demonstrate that our approach effectively mitigates hallucinations while preserving general model capabilities. Specifically, our method achieves up to a 38% improvement over the current state-of-the-art on standard hallucination benchmarks.

PackKV: Reducing KV Cache Memory Footprint through LLM-Aware Lossy Compression

Transformer-based large language models (LLMs) have demonstrated remarkable potential across a wide range of practical applications. However, long-context inference remains a significant challenge due to the substantial memory requirements of the key-value (KV) cache, which can scale to several gigabytes as sequence length and batch size increase. In this paper, we present \textbf{PackKV}, a generic and efficient KV cache management framework optimized for long-context generation. %, which synergistically supports both latency-critical and throughput-critical inference scenarios. PackKV introduces novel lossy compression techniques specifically tailored to the characteristics of KV cache data, featuring a careful co-design of compression algorithms and system architecture. Our approach is compatible with the dynamically growing nature of the KV cache while preserving high computational efficiency. Experimental results show that, under the same and minimum accuracy drop as state-of-the-art quantization methods, PackKV achieves, on average, \textbf{153.2}\% higher memory reduction rate for the K cache and \textbf{179.6}\% for the V cache. Furthermore, PackKV delivers extremely high execution throughput, effectively eliminating decompression overhead and accelerating the matrix-vector multiplication operation. Specifically, PackKV achieves an average throughput improvement of \textbf{75.7}\% for K and \textbf{171.7}\% for V across A100 and RTX Pro 6000 GPUs, compared to cuBLAS matrix-vector multiplication kernels, while demanding less GPU memory bandwidth. Code available on https://github.com/BoJiang03/PackKV

DeepCQ: General-Purpose Deep-Surrogate Framework for Lossy Compression Quality Prediction

Error-bounded lossy compression techniques have become vital for scientific data management and analytics, given the ever-increasing volume of data generated by modern scientific simulations and instruments. Nevertheless, assessing data quality post-compression remains computationally expensive due to the intensive nature of metric calculations. In this work, we present a general-purpose deep-surrogate framework for lossy compression quality prediction (DeepCQ), with the following key contributions: 1) We develop a surrogate model for compression quality prediction that is generalizable to different error-bounded lossy compressors, quality metrics, and input datasets; 2) We adopt a novel two-stage design that decouples the computationally expensive feature-extraction stage from the light-weight metrics prediction, enabling efficient training and modular inference; 3) We optimize the model performance on time-evolving data using a mixture-of-experts design. Such a design enhances the robustness when predicting across simulation timesteps, especially when the training and test data exhibit significant variation. We validate the effectiveness of DeepCQ on four real-world scientific applications. Our results highlight the framework's exceptional predictive accuracy, with prediction errors generally under 10\% across most settings, significantly outperforming existing methods. Our framework empowers scientific users to make informed decisions about data compression based on their preferred data quality, thereby significantly reducing I/O and computational overhead in scientific data analysis.

An Efficient Gradient-Aware Error-Bounded Lossy Compressor for Federated Learning

Federated learning (FL) enables collaborative model training without exposing clients' private data, but its deployment is often constrained by the communication cost of transmitting gradients between clients and the central server, especially under system heterogeneity where low-bandwidth clients bottleneck overall performance. Lossy compression of gradient data can mitigate this overhead, and error-bounded lossy compression (EBLC) is particularly appealing for its fine-grained utility-compression tradeoff. However, existing EBLC methods (e.g., SZ), originally designed for smooth scientific data with strong spatial locality, rely on generic predictors such as Lorenzo and interpolation for entropy reduction to improve compression ratio. Gradient tensors, in contrast, exhibit low smoothness and weak spatial correlation, rendering these predictors ineffective and leading to poor compression ratios. To address this limitation, we propose an EBLC framework tailored for FL gradient data to achieve high compression ratios while preserving model accuracy. The core of it is an innovative prediction mechanism that exploits temporal correlations across FL training rounds and structural regularities within convolutional kernels to reduce residual entropy. The predictor is compatible with standard quantizers and entropy coders and comprises (1) a cross-round magnitude predictor based on a normalized exponential moving average, and (2) a sign predictor that leverages gradient oscillation and kernel-level sign consistency. Experiments show that this new EBLC yields up to 1.53x higher compression ratios than SZ3 with lower accuracy loss. Integrated into a real-world FL framework, APPFL, it reduces end-to-end communication time by 76.1%-96.2% under various constrained-bandwidth scenarios, demonstrating strong scalability for real-world FL deployments.

MoE-Compression: How the Compression Error of Experts Affects the Inference Accuracy of MoE Model?

With the widespread application of Mixture of Experts (MoE) reasoning models in the field of LLM learning, efficiently serving MoE models under limited GPU memory constraints has emerged as a significant challenge. Offloading the non-activated experts to main memory has been identified as an efficient approach to address such a problem, while it brings the challenges of transferring the expert between the GPU memory and main memory. We need to explore an efficient approach to compress the expert and analyze how the compression error affects the inference performance. To bridge this gap, we propose employing error-bounded lossy compression algorithms (such as SZ3 and CuSZp) to compress non-activated experts, thereby reducing data transfer overhead during MoE inference. We conduct extensive experiments across various benchmarks and present a comprehensive analysis of how compression-induced errors in different experts affect overall inference accuracy. The results indicate that experts in the shallow layers, which are primarily responsible for the attention mechanism and the transformation of input tokens into vector representations, exhibit minimal degradation in inference accuracy when subjected to bounded errors. In contrast, errors in the middle-layer experts, which are central to model reasoning, significantly impair inference accuracy. Interestingly, introducing bounded errors in the deep-layer experts, which are mainly responsible for instruction following and output integration, can sometimes lead to improvements in inference accuracy.

Systematic Evaluation of Optimization Techniques for Long-Context Language Models

Large language models (LLMs) excel across diverse natural language processing tasks but face resource demands and limited context windows. Although techniques like pruning, quantization, and token dropping can mitigate these issues, their efficacy in long-context scenarios and system evaluation remains underexplored. This paper systematically benchmarks these optimizations, characterizing memory usage, latency, and throughput, and studies how these methods impact the quality of text generation. We first analyze individual optimization methods for two LLM architectures supporting long context and then systematically evaluate combinations of these techniques to assess how this deeper analysis impacts performance metrics. We subsequently study the scalability of individual optimization methods on a larger variant with 70 billion-parameter model. Our novel insights reveal that naive combination inference optimization algorithms can adversely affect larger models due to compounded approximation errors, as compared to their smaller counterparts. Experiments show that relying solely on F1 obscures these effects by hiding precision-recall trade-offs in question answering tasks. By integrating system-level profiling with task-specific insights, this study helps LLM practitioners and researchers explore and balance efficiency, accuracy, and scalability across tasks and hardware configurations.