Test-time scaling evaluates reasoning LLMs by sampling multiple outputs per prompt, but ranking models in this regime remains underexplored.We formalize dense benchmark ranking under test-time scaling and introduce Scorio, a library that implements statistical ranking methods such as paired-comparison models, item response theory (IRT) models, voting rules, and graph-and spectral-based methods.Across 20 reasoning models on four Olympiadstyle math benchmarks (AIME'24, AIME'25, HMMT'25, and BrUMO'25; up to N = 80 trials), most full-trial rankings agree closely with the Bayesian gold standard Bayes U @80 (mean Kendall's τ b = 0.93-0.95), and 19-34 methods recover exactly the same ordering.In the single-trial regime, the best methods reach τ b ≈ 0.86.Using greedy decoding as an empirical prior (Bayes R0 @N ) reduces variance at N = 1 by 16-52%, but can bias rankings when greedy and stochastic sampling disagree.These results identify reliable ranking methods for both high-and low-budget test-time scaling.We release Scorio as an open-source library at 1 .

Log anomaly detection (LogAD) is crucial for identifying failures and threats in large-scale computing and cyber-infrastructure systems. However, most existing LogAD approaches suffer from key limitations: they depend on slow and error-prone log parsing, employ tightly coupled end-to-end pipelines, often require supervision for improved detection performance, and rely on flawed single-pass evaluation protocols that fail to reflect the temporal dynamics of real-world online detection. These issues significantly hinder scalability, adaptability, and the practical deployment of solutions. To address these limitations, we introduce$\mathbf{K}^{\mathbf{4}}$(Knowing the Unknown by Knowing only the Known), a fully unsupervised, parser-independent, and representation-agnostic LogAD framework designed for high-performance online detection. At its core,$K^{4}$is grounded in a novel formulation based on representation-level typicality estimation, which transforms arbitrary log embeddings into compact and interpretable four-dimensional descriptors: Precision, Recall, Density, and Coverage (PRDC), which are swiftly computed via GPU-acceleration into geometric$k$-nearest neighbor statistics. These descriptors inform lightweight, modular detectors, including KDE, GMM, OCSVM, and a new adaptation of DeepSVDD, which enables efficient and accurate anomaly scoring without relying on structured formats or log representation retraining. To support realistic deployment scenarios, we also propose a principled streaming-faithful evaluation protocol that partitions datasets into fixed-size chunks and applies sliding-window sampling with strides to emulate online log ingestion, alleviating the performance overestimation and dataset undercoverage issues of prior single-pass evaluations and enabling reproducible benchmarking across datasets with varying anomaly densities. Using this setup, we conduct over$\mathbf{1 2 5, 0 0 0}$experiments across three real-world datasets (HDFS, BGL, Thunderbird), six pretrained embedding models, four detectors, and multiple training and log sampling configurations. Compared to six representative baseline methods spanning supervised, semi-supervised, self-supervised, and unsupervised paradigms,$K^{4}$consistently sets new state-of-the-art results (AUROC: 0.995-0.999, F1: 0.989-0.992) and outperforms all baselines by large margins, while keeping detector training under 4 seconds and per-sample inference latency as low as$4 \mu ~\mathrm{s}$, which are orders of magnitude faster than the most competitive alternatives.

Large-scale AI models, such as Large Language Models (LLMs) and Diffusion Models (DMs), have grown rapidly in size, creating significant challenges for efficient deployment on resource-constrained hardware. In this paper, we introduce Dynamic-Length Float (DFloat11), a lossless compression framework that reduces LLM and DM size by 30\% while preserving outputs that are bit-for-bit identical to the original model. DFloat11 is motivated by the low entropy in the BFloat16 weight representation of LLMs, which reveals significant inefficiency in the existing storage format. By applying entropy coding, DFloat11 assigns dynamic-length encodings to weights based on frequency, achieving near information-optimal compression without any loss of precision. To facilitate efficient inference with dynamic-length encodings, we develop a custom GPU kernel for fast online decompression. Our design incorporates the following: (i) compact, hierarchical lookup tables (LUTs) that fit within GPU SRAM for efficient decoding, (ii) a two-phase GPU kernel for coordinating thread read/write positions using lightweight auxiliary variables, and (iii) transformer-block-level decompression to minimize latency. Experiments on Llama 3.3, Qwen 3, Mistral 3, FLUX.1, and others validate our hypothesis that DFloat11 achieves around 30\% model size reduction while preserving bit-for-bit identical outputs. Compared to a potential alternative of offloading parts of an uncompressed model to the CPU to meet memory constraints, DFloat11 achieves 2.3--46.2$\times$ higher throughput in token generation. With a fixed GPU memory budget, DFloat11 enables 5.7--14.9$\times$ longer generation lengths than uncompressed models. Notably, our method enables lossless inference of Llama 3.1 405B, an 810GB model, on a single node equipped with 8$\times$80GB GPUs.