搜索结果：Demystifying

Machine learning predictions are increasingly used to supplement incomplete or costly-to-measure outcomes in fields such as biomedical research, environmental science, and social science. However, treating predictions as ground truth introduces bias while ignoring them wastes valuable information. Prediction-Powered Inference (PPI) offers a principled framework that leverages predictions from large unlabeled datasets to improve statistical efficiency while maintaining valid inference through explicit bias correction using a smaller labeled subset. Despite its potential, the growing PPI variants and the subtle distinctions between them have made it challenging for practitioners to determine when and how to apply these methods responsibly. This paper demystifies PPI by synthesizing its theoretical foundations, methodological extensions, connections to existing statistics literature, and diagnostic tools into a unified practical workflow. Using the Mosaiks housing price data, we show that PPI variants produce tighter confidence intervals than complete-case analysis, but that double-dipping, i.e. reusing training data for inference, leads to anti-conservative confidence intervals and c

The empirical success of large language model (LLM) pre-training relies heavily on heuristic stabilization techniques, such as explicit normalization layers and weight decay. While recent constrained optimization approaches that explicitly restrict weights may improve numerical stability and performance, the mechanism and motivation for adding constraints still remain elusive. This paper systematically demystifies the role of explicit manifold constraints in LLM pre-training. By introducing the Msign-Aligned Constrained Riemannian Optimizer (MACRO)-a provably convergent, single-loop optimization framework-our study disentangles weight regularization heuristics from interacting mechanisms like RMS normalization and decoupled weight decay. Theoretical analyses and comprehensive empirical evaluations reveal that manifold constraints independently bound forward activation scales and enforce stable rotational equilibrium, thereby subsuming the roles of these heuristic mechanisms. Evaluations on large-scale LLM architectures demonstrate that MACRO achieves highly competitive performance while rigorously preserving the theoretical guarantees of exact Riemannian optimization.

In the last few years, considerable research efforts have focused on measuring and improving Starlink network performance, especially for user terminals (UTs) in stationary scenarios. However, the performance of Starlink networks in mobility settings, particularly with frequent changes in the UT's orientation, and the impact of environmental factors, such as transient obstructions, has not been thoroughly studied, leaving gaps in understanding the causes of performance degradation. Recently, researchers have started identifying the communicating satellites to evaluate satellite selection strategies and the impact on network performance. However, existing Starlink satellite identification methods only work in stationary, obstruction-free scenarios, as they do not account for UT mobility, obstructions or detect dynamic beam switching events. In this paper, we reveal that the UT can perform multiple dynamic beam switching attempts to connect to different satellites when the UT-satellite link is degraded. This degradation can occur either due to the loss of line-of-sight (LoS) from changes in the FOV or obstructions, or due to poor signal quality, extending UT-satellite handovers beyon

The NVIDIA Collective Communication Library (NCCL) is a critical software layer enabling high-performance collectives on large-scale GPU clusters. Despite being open source with a documented API, its internal design remains largely opaque. The orchestration of communication channels, selection of protocols, and handling of memory movement across devices and nodes are not well understood, making it difficult to analyze performance or identify bottlenecks. This paper presents a comprehensive analysis of NCCL, focusing on its communication protocol variants (Simple, LL, and LL128), mechanisms governing intra-node and inter-node data movement, and ring- and tree-based collective communication algorithms. The insights obtained from this study serve as the foundation for ATLAHS, an application-trace-driven network simulation toolchain capable of accurately reproducing NCCL communication patterns in large-scale AI training workloads. By demystifying NCCL's internal architecture, this work provides guidance for system researchers and performance engineers working to optimize or simulate collective communication at scale.

FLUX.1 is a diffusion-based text-to-image generation model developed by Black Forest Labs, designed to achieve faithful text-image alignment while maintaining high image quality and diversity. FLUX is considered state-of-the-art in text-to-image generation, outperforming popular models such as Midjourney, DALL-E 3, Stable Diffusion 3 (SD3), and SDXL. Although publicly available as open source, the authors have not released official technical documentation detailing the model's architecture or training setup. This report summarizes an extensive reverse-engineering effort aimed at demystifying FLUX's architecture directly from its source code, to support its adoption as a backbone for future research and development. This document is an unofficial technical report and is not published or endorsed by the original developers or their affiliated institutions.

Off-policy reinforcement learning (RL) for large language models (LLMs) is attracting growing interest, driven by practical constraints in real-world applications, the complexity of LLM-RL infrastructure, and the need for further innovations of RL methodologies. While classic REINFORCE and its modern variants like Group Relative Policy Optimization (GRPO) are typically regarded as on-policy algorithms with limited tolerance of off-policyness, we present in this work a first-principles derivation for group-relative REINFORCE -- a REINFORCE variant that uses the within-group mean reward as the baseline for advantage calculation -- without assuming a specific training data distribution, showing that it admits a native off-policy interpretation. This perspective yields two general principles for adapting REINFORCE to truly off-policy settings: regularizing policy updates, and actively shaping the data distribution. Our analysis demystifies some myths about the roles of importance sampling and clipping in GRPO, unifies and reinterprets two recent algorithms -- Online Policy Mirror Descent and Asymmetric REINFORCE -- as regularized forms of the REINFORCE loss, and offers theoretical just

Proximal causal inference (PCI) has emerged as a promising framework for identifying and estimating causal effects in the presence of unobserved confounders. While many traditional causal inference methods rely on the assumption of no unobserved confounding, this assumption is likely often violated. PCI addresses this challenge by relying on an alternative set of assumptions regarding the relationships between treatment, outcome, and auxiliary variables that serve as proxies for unmeasured confounders. We review existing identification results, discuss the assumptions necessary for valid causal effect estimation via PCI, and compare different PCI estimation methods. We offer practical guidance on operationalizing PCI, with a focus on selecting and evaluating proxy variables using domain knowledge, measurement error perspectives, and negative control analogies. Through conceptual examples, we demonstrate tensions in proxy selection and discuss the importance of clearly defining the unobserved confounding mechanism. By bridging formal results with applied considerations, this work aims to demystify PCI, encourage thoughtful use in practice, and identify open directions for methodolog

Artificial Intelligence (AI) is rapidly being integrated into critical systems across various domains, from healthcare to autonomous vehicles. While its integration brings immense benefits, it also introduces significant risks, including those arising from AI misuse. Within the discourse on managing these risks, the terms "AI Safety" and "AI Security" are often used, sometimes interchangeably, resulting in conceptual confusion. This paper aims to demystify the distinction and delineate the precise research boundaries between AI Safety and AI Security. We provide rigorous definitions, outline their respective research focuses, and explore their interdependency, including how security breaches can precipitate safety failures and vice versa. Using clear analogies from message transmission and building construction, we illustrate these distinctions. Clarifying these boundaries is crucial for guiding precise research directions, fostering effective cross-disciplinary collaboration, enhancing policy effectiveness, and ultimately, promoting the deployment of trustworthy AI systems.

Graph neural networks (GNNs) are powerful tools for solving graph-related problems. Distributed GNN frameworks and systems enhance the scalability of GNNs and accelerate model training, yet most are optimized for node classification. Their performance on link prediction remains underexplored. This paper demystifies distributed training of GNNs for link prediction by investigating the issue of performance degradation when each worker trains a GNN on its assigned partitioned subgraph without having access to the entire graph. We discover that the main sources of the issue come from not only the information loss caused by graph partitioning but also the ways of drawing negative samples during model training. While sharing the complete graph information with each worker resolves the issue and preserves link prediction accuracy, it incurs a high communication cost. We propose SpLPG, which effectively leverages graph sparsification to mitigate the issue of performance degradation at a reduced communication cost. Experiment results on several public real-world datasets demonstrate the effectiveness of SpLPG, which reduces the communication overhead by up to about 80% while mostly preservi

Scaling inference compute enhances reasoning in large language models (LLMs), with long chains-of-thought (CoTs) enabling strategies like backtracking and error correction. Reinforcement learning (RL) has emerged as a crucial method for developing these capabilities, yet the conditions under which long CoTs emerge remain unclear, and RL training requires careful design choices. In this study, we systematically investigate the mechanics of long CoT reasoning, identifying the key factors that enable models to generate long CoT trajectories. Through extensive supervised fine-tuning (SFT) and RL experiments, we present four main findings: (1) While SFT is not strictly necessary, it simplifies training and improves efficiency; (2) Reasoning capabilities tend to emerge with increased training compute, but their development is not guaranteed, making reward shaping crucial for stabilizing CoT length growth; (3) Scaling verifiable reward signals is critical for RL. We find that leveraging noisy, web-extracted solutions with filtering mechanisms shows strong potential, particularly for out-of-distribution (OOD) tasks such as STEM reasoning; and (4) Core abilities like error correction are in

QUIC is an advanced transport layer protocol whose ubiquity on the Internet is now very apparent. Importantly, QUIC fuels the next generation of web browsing: HTTP/3. QUIC is a stateful and connection oriented protocol which offers similar features (and more) to the combination of TCP and TLS. There are several difficulties which readers may encounter when learning about QUIC: i.) its rapid evolution (particularly, differentiation between the QUIC standard and the now deprecated Google QUIC), ii.) numerous RFCs whose organization, language, and detail may be challenging to the casual reader, and iii.) the nature of QUIC's cross-layer and privacy-centric implementation, making it impossible to understand or debug by looking at packets alone. For these reasons, the aim of this paper is to present QUIC in a complete yet approachable fashion, thereby demystifying the protocol from its specifications.

Training data plays a crucial role in Large Language Models (LLM) scaling, yet high quality data is of limited supply. Synthetic data techniques offer a potential path toward sidestepping these limitations. We conduct a large-scale empirical investigation (>1000 LLMs with >100k GPU hours) using a unified protocol and scaling laws, comparing natural web data, diverse synthetic types (rephrased text, generated textbooks), and mixtures of natural and synthetic data. Specifically, we found pre-training on rephrased synthetic data \textit{alone} is not faster than pre-training on natural web texts; while pre-training on 1/3 rephrased synthetic data mixed with 2/3 natural web texts can speed up 5-10x (to reach the same validation loss) at larger data budgets. Pre-training on textbook-style synthetic data \textit{alone} results in notably higher loss on many downstream domains especially at small data budgets. "Good" ratios of synthetic data in training data mixtures depend on the model size and data budget, empirically converging to ~30% for rephrased synthetic data. Larger generator models do not necessarily yield better pre-training data than ~8B-param models. These results contr

Public cloud serverless platforms have attracted a large user base due to their high scalability, plug-and-play deployment model, and pay-per-use billing. However, compared to virtual machines and container hosting services, modern serverless offerings typically impose higher per-unit time and resource charges. Additionally, billing practices such as wall-clock time allocation-based billing, invocation fees, and usage rounding up can further increase costs. This work, for the first time, holistically demystifies these costs by conducting an in-depth, top-down characterization and analysis from user-facing billing models, through request serving architectures, and down to operating system scheduling on major public serverless platforms. We quantify, for the first time, how current billing practices inflate billable resources up to 4.35x beyond actual consumption. Also, our analysis reveals previously unreported cost drivers, such as operational patterns of serving architectures that create overheads, details of resource allocation during keep-alive periods, and OS scheduling granularity effects that directly impact both performance and billing. By tracing the sources of costs from b

Convolution remains the most compute-intensive operation in AI acceleration, often constituting over 80-90% of the workload. Existing approaches in spatial architectures such as coarse-grained reconfigurable arrays (CGRAs) and field-programmable gate arrays (FPGAs) frequently rely on loop unrolling or GEMM-based matrix transformations, introducing significant overhead in both data movement and instruction control. This paper presents a new framework designed to systematically demystify the 7-dimensional convolution loop nest by reinterpreting it as a hardware-centric data and instruction streaming problem. Instead of treating the loop nest as a fixed computational construct, our approach exposes its structure as a set of spatial and temporal mappings governed by hardware parameters such as compute element distribution, interconnect topology, and reconfigurability. This abstraction supports lightweight, flexible deployment of convolution without reliance on heavyweight transformations or reordering schemes. We demonstrate the application of our approach on the MAVeC accelerator. We detail the implementation of convolution operations in MAVeC and extend the framework to support full

With the advent of modern multi-chiplet FPGA architectures, vendors have begun integrating hardened NoC to address the scalability, resource usage, and frequency disadvantages of soft NoCs. However, as this work shows, effectively harnessing these hardened NoC is not trivial. It requires detailed knowledge of the microarchitecture and how it relates to the physical design of the FPGA. Existing literature has provided in-depth analyses for NoC in MPSoC devices, but few studies have systematically evaluated hardened NoC in FPGA, which have several unique implications. This work aims to bridge this knowledge gap by demystifying the performance and design trade-offs of hardened NoC on FPGA. Our work performs detailed performance analysis of hard (and soft) NoC under different settings, including diverse NoC topologies, routing strategies, traffic patterns and different external memories under various NoC placements. In the context of Versal FPGAs, our results show that using hardened NoC in multi-SLR designs can reduce expensive cross-SLR link usage by up to 30~40%, eliminate general-purpose logic overhead, and remove most critical paths caused by large on-chip crossbars. However, unde

This paper studies off-policy evaluation (OPE) in reinforcement learning with a focus on behavior policy estimation for importance sampling. Prior work has shown empirically that estimating a history-dependent behavior policy can lead to lower mean squared error (MSE) even when the true behavior policy is Markovian. However, the question of why the use of history should lower MSE remains open. In this paper, we theoretically demystify this paradox by deriving a bias-variance decomposition of the MSE of ordinary importance sampling (IS) estimators, demonstrating that history-dependent behavior policy estimation decreases their asymptotic variances while increasing their finite-sample biases. Additionally, as the estimated behavior policy conditions on a longer history, we show a consistent decrease in variance. We extend these findings to a range of other OPE estimators, including the sequential IS estimator, the doubly robust estimator and the marginalized IS estimator, with the behavior policy estimated either parametrically or non-parametrically.

Designing effective reward functions is critical for reinforcement learning-based biomechanical simulations, yet HCI researchers and practitioners often waste (computation) time with unintuitive trial-and-error tuning. This paper demystifies reward function design by systematically analyzing the impact of effort minimization, task completion bonuses, and target proximity incentives on typical HCI tasks such as pointing, tracking, and choice reaction. We show that proximity incentives are essential for guiding movement, while completion bonuses ensure task success. Effort terms, though optional, help refine motion regularity when appropriately scaled. We perform an extensive analysis of how sensitive task success and completion time depend on the weights of these three reward components. From these results we derive practical guidelines to create plausible biomechanical simulations without the need for reinforcement learning expertise, which we then validate on remote control and keyboard typing tasks. This paper advances simulation-based interaction design and evaluation in HCI by improving the efficiency and applicability of biomechanical user modeling for real-world interface dev

MuZero has achieved superhuman performance in various games by using a dynamics network to predict the environment dynamics for planning, without relying on simulators. However, the latent states learned by the dynamics network make its planning process opaque. This paper aims to demystify MuZero's model by interpreting the learned latent states. We incorporate observation reconstruction and state consistency into MuZero training and conduct an in-depth analysis to evaluate latent states across two board games: 9x9 Go and Gomoku, and three Atari games: Breakout, Ms. Pacman, and Pong. Our findings reveal that while the dynamics network becomes less accurate over longer simulations, MuZero still performs effectively by using planning to correct errors. Our experiments also show that the dynamics network learns better latent states in board games than in Atari games. These insights contribute to a better understanding of MuZero and offer directions for future research to improve the performance, robustness, and interpretability of the MuZero algorithm. The code and data are available at https://rlg.iis.sinica.edu.tw/papers/demystifying-muzero-planning.

Recently, the emergence of agentic RL has showcased that RL could also effectively improve the agentic reasoning ability of LLMs, yet the key design principles and optimal practices remain unclear. In this work, we conduct a comprehensive and systematic investigation to demystify reinforcement learning in agentic reasoning from three key perspectives: data, algorithm, and reasoning mode. We highlight our key insights: (i) Replacing stitched synthetic trajectories with real end-to-end tool-use trajectories yields a far stronger SFT initialization; high-diversity, model-aware datasets sustain exploration and markedly improve RL performance. (ii) Exploration-friendly techniques are crucial for agentic RL, such as clip higher, overlong reward shaping, and maintaining adequate policy entropy could improve the training efficiency. (iii) A deliberative strategy with fewer tool calls outperforms frequent tool calls or verbose self-reasoning, improving tool efficiency and final accuracy. Together, these simple practices consistently enhance agentic reasoning and training efficiency, achieving strong results on challenging benchmarks with smaller models, and establishing a practical baseline