SAKURA: Reused Name Across Diverse Fields

• SAKURA is a nominal label applied to distinct research artifacts across domains, including audio-language benchmarks, HPC clusters, side-channel hardware, and more.
• The SAKURA audio-language benchmark critically evaluates multi-hop reasoning in LALMs, revealing significant gaps between single-hop perception and inference.
• Distinct implementations of SAKURA in HPC, software engineering, astrophysics, and cryptographic hashing highlight the need for immediate domain disambiguation.

to=functions.shell 荣富js av不卡免费播放 北京赛车如何json {"command":"python - <<'PY'\nimport urllib.request,sys\nurl='https://export.arxiv.org/api/query?search_query=all:SAKURA&start=0&max_results=3'\ntry:\n data=urllib.request.urlopen(url,timeout=10).read(2000)\n print(data.decode('utf-8','ignore'))\nexcept Exception as e:\n print('ERR',e)\nPY"} акоронавирус code 0 stdout: ERR <urlopen error timed out>

In the cited arXiv literature, SAKURA denotes multiple independent research artifacts rather than a single unified system. The name appears in large audio-LLM evaluation, open Ethernet-based HPC, side-channel experimentation hardware, self-supervised sign-language recognition, software test generation, vocabulary difficulty prediction, gravitational $N$-body integration, educational radiation detection, Keccak tree hashing, and percolation-based magnetotail modeling. The shared label therefore functions as a cross-domain name reused by unrelated projects, with the most technically developed recent usages including the SAKURA audio-reasoning benchmark (Yang et al., 19 May 2025), the SAKURAONE AI-HPC platform (Konishi, 2 Jul 2025), and the sign-language representation-learning framework SAKURA (Zhao et al., 2024).

1. Major research usages

The literature treats the following SAKURA-named artifacts as separate objects with distinct goals, datasets, and technical stacks.

Usage	Domain	Core characterization
SAKURA benchmark (Yang et al., 19 May 2025)	Audio-language evaluation	Multiple-choice single-hop and multi-hop reasoning over speech and audio
SAKURAONE (Konishi, 2 Jul 2025)	HPC infrastructure	100-node, 800-GPU managed cluster with 800 GbE, SONiC, RoCEv2
SAKURA-X in MaskedNet (Dubey et al., 2019)	Hardware security	FPGA board used for DPA attacks and masking evaluation
SAKURA (Zhao et al., 2024)	Sign-language recognition	Self-supervised skeleton-based representation learning with spatial-temporal consistency
Sakura (Stennett et al., 30 May 2026)	Software engineering	Multi-agent framework for generating complex tests from natural-language descriptions
Sakura (Nohejl et al., 14 May 2026)	NLP evaluation	Systems for vocabulary difficulty prediction in the BEA 2026 shared task
Sakura integrator (Ferrari et al., 2014)	Computational astrophysics	Keplerian-based Hamiltonian splitting for gravitational $N$-body simulations
SAKURA detector (Matsushita et al., 8 Sep 2025)	Educational instrumentation	Palm-sized two-dimensional muon scintillation detector
Sakura model (Milovanov, 2012)	Space plasma theory	Percolation-based SOC model of the Earth’s magnetotail
Sakura coding (Atighehchi, 2016)	Cryptographic hashing	Tree-hash framing for parallel SHAKE constructions

A recurrent misconception is to assume technical continuity across these uses. The papers do not present a common codebase, hardware lineage, or scientific program linking the benchmark, the HPC cluster, the FPGA board, the integrator, and the detector; the commonality is nominal rather than architectural.

2. SAKURA as an audio-language reasoning benchmark

The SAKURA benchmark was introduced to evaluate multi-hop reasoning of large audio-LLMs (LALMs) on raw speech and audio inputs (Yang et al., 19 May 2025). It is organized into four tracks—Gender, Language, Emotion, and Animal—and each track is split into single-hop and multi-hop sub-tracks. Each sub-track contains 500 multiple-choice questions per track, for a total of 4,000 questions. The reported corpus-level statistics are an average audio duration of 4.79 s with $\sigma=2.23$ and range .22–20.78 s, and an average instruction length of 31.3 words with $\sigma=12.6$ and range 8–66. The benchmark defines

$$\mathrm{Acc}_{\rm single}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^s=y_i^s), \qquad \mathrm{Acc}_{\rm multi}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^m=y_i^m),$$

and also introduces the multi-hop reasoning score (MRS),

$$\mathrm{MRS} = \frac{\sum_{i=1}^N \mathbf{1}(\hat y_i^s=y_i^s \wedge \hat y_i^m=y_i^m)} {\sum_{i=1}^N \mathbf{1}(\hat y_i^s=y_i^s)}.$$

This formulation isolates failures of reasoning from failures of first-hop perception.

The headline empirical result is a large gap between single-hop perception and multi-hop reasoning in end-to-end LALMs. In the benchmark’s main comparison, Qwen2-Audio reaches 81.2 % single-hop average but only 49.1 % multi-hop average, while DeSTA2 reaches 63.0 % and 57.1 %, and GPT-4o Audio reaches 71.3 % and 54.3 %. A modular cascade, ASR+AAC+LLM, attains 79.1 % single-hop and 74.5 % multi-hop, outperforming most end-to-end systems on multi-hop reasoning. The authors interpret this as evidence that current LALMs can often extract the relevant attribute but fail to propagate that latent audio fact into downstream inference.

Model	Single-hop avg	Multi-hop avg
DeSTA2	63.0 %	57.1 %
Qwen2-Audio	81.2 %	49.1 %
GPT-4o Audio	71.3 %	54.3 %
ASR+AAC+LLM	79.1 %	74.5 %
Chance	31.3 %	31.3 %

Subsequent work used SAKURA as a diagnostic benchmark rather than only a leaderboard. DeSTA2.5-Audio reports SAKURA-Single accuracy = 76.65 % and SAKURA-Multi accuracy = 69.85 %, and the paper attributes the improvement chiefly to self-generated cross-modal alignment, emphasizing distributional consistency and preservation of pre-training knowledge (Lu et al., 3 Jul 2025). ORCA argues that Q-Former compression collapses audio information and reports Avg-M = 75.2 % on SAKURA multi-hop, compared with 48.8 % for an identically trained DeSTA2.5-Audio 4B baseline and 49.0 % for Audio Flamingo-3 (8B); the same paper reports a 12× reduction in mean pairwise query cosine similarity and a 75× increase in cross-speaker variance at the connector level (Chung et al., 7 Jul 2026). A separate faithfulness study perturbs chain-of-thoughts on SAKURA using filler token injection, paraphrasing, early answering, and adding mistakes, and concludes that LALM chain-of-thoughts on SAKURA are largely faithful, because answer consistency declines when the reasoning trace is semantically corrupted (Jain et al., 26 Sep 2025).

The benchmark is therefore significant less as a generic speech benchmark than as a probe of the interface between audio perception, latent fact retention, and reasoning composition. A plausible implication is that SAKURA has become a specialized stress test for whether multimodal systems can use internally recovered acoustic attributes as explicit premises in later inference.

3. SAKURA in HPC infrastructure and measurement hardware

SAKURAONE is a managed HPC cluster developed and operated by the SAKURA Internet Research Center and optimized for advanced workloads including LLM training (Konishi, 2 Jul 2025). The system comprises 100 compute nodes, each with 2× Intel Xeon Platinum 8580+, 1.5 TB DDR5-5600, and 8× NVIDIA H100 SXM 80 GB, for a total of 800 GPUs. Storage is a shared all-flash Lustre subsystem with 2 PB physical capacity. Internode communication uses a full-bisection bandwidth interconnect with a Rail-Optimized topology over 800 GbE (2DR4), RoCEv2, and SONiC on Edge-core AIS800-64O switches powered by Broadcom Tomahawk 5 ASICs. The system was ranked 49th in the ISC 2025 TOP500 by HPL and was reported as the only system within the top 100 using a fully open SONiC-based 800 GbE interconnect. Its benchmark results are 33.95 PFLOP/s on HPL, 396.295 TFLOP/s on HPCG, and 339.86 PFLOP/s on HPL-MxP in FP8.

A later operational study adds workload-level evidence about the same platform in a single-tenant LLM development environment (Konishi et al., 15 Apr 2026). It reports that CANCELLED jobs accounted for 73.5 % of GPU-occupied time, COMPLETED jobs for ~26.2 %, and FAILED jobs for 0.3 %. By node-count category, 1-node jobs represented 76.9 % of jobs but only 1.8 % of GPU-time, whereas jobs using $\ge 17$ nodes were only 3.3 % of jobs but consumed 73.3 % of GPU-time. The time series showed a phase shift from continued pretraining at 17–32 nodes/day in Jun ’24–Jan ’25 to a growing share of 3–16 node jobs in Feb ’25–Mar ’25, interpreted as a transition from large-scale training to iterative refinement. The paper’s broader claim is that open Ethernet fabrics with careful ECN/PFC tuning, multi-rail NCCL striping, and physical separation of training and I/O planes can sustain top-tier AI-HPC behavior without proprietary interconnect lock-in.

A different hardware usage of the name appears in SAKURA-X, the FPGA board used in the MaskedNet side-channel study (Dubey et al., 2019). The platform is based on a Xilinx Kintex-7 XC7K160T-1FBGC and includes a 10 mΩ shunt resistor in the FPGA supply line with on-board amplifier stages. In the reported setup, the shunt voltage was boosted with a Riscure HD24248 and digitized by a PicoScope 3206D at 250 MHz, while the design ran at $f_{\rm CLK}=24$ MHz. On this board, first-order DPA on the unprotected neural-network inference engine succeeded with ~200 traces for the stage-2 adder-tree attack, whereas the masked design with PRNG enabled failed under first-order DPA even at 100,000 traces; the masked implementation incurred ≈ 2.8× slowdown and ≈ 2.3× area overhead. SAKURA-X is thus a measurement platform for hardware side-channel analysis, not an HPC system.

These two hardware usages illustrate a sharp domain split. SAKURAONE represents open, GPU-dense, datacenter-scale infrastructure; SAKURA-X represents low-level, cycle-aligned measurement hardware for physical attacks on FPGA implementations.

4. SAKURA in machine learning and software engineering

In sign-language recognition, SAKURA is a self-supervised pre-training framework for isolated recognition from skeleton (pose) data (Zhao et al., 2024). Its two defining mechanisms are multi-granularity spatial consistency and multi-order temporal consistency with cross-modal transfer. The framework separates hand and trunk branches, computes a first-order motion modality

$m_{i,p}=x_{i,p}-x_{i-1,p},$

and couples joint and motion encoders through bidirectional KL-based knowledge transfer using memory queues of size $16\,384$ and top-$N$0 neighbor anchors with $N$1. Training uses MoCo with $N$2, temperatures $N$3, $N$4, $N$5, $N$6, 150 pre-training epochs, and GCN+Transformer backbones outputting 512-D features followed by an MLP to 128-D contrastive space. Reported benchmark results include MSASL100/200/1000 = 86.26 / 82.93 / 65.22 %, WLASL100/300/2000 = 86.43 / 77.84 / 51.98 %, NMFs_CSL = 75.6 %, and SLR500 = 96.9 %. Ablations show 69.9 % $N$7 86.3 % Top-1 when pre-training rises from 0 % to 100 % of the data, and 83.8 / 81.6 / 86.3 % for Joint-only / Motion-only / Joint+Motion.

In software engineering, Sakura is described as the first agent-based framework for generating structurally complex test cases from NL descriptions (Stennett et al., 30 May 2026). The architecture contains three main phases: code index creation, NL decomposition into a BDD-style intermediate form, and an agentic workflow with Localization, Composition, and Supervisor agents. The evaluation uses 20 Apache Commons projects and 1,464 test scenarios derived from 488 ground-truth tests. The principal metrics are compilability, coverage overlap, and structural similarity over instantiated types, assertion types, method calls, and focal methods. Against Gemini CLI, Sakura achieves 97.3 % compilability with G-2.5 Pro, compared to 54.6 % for the baseline; on coverage overlap, Sakura (Pro) reaches 79 % versus 48 % for Gemini CLI (Pro). With smaller open-source models, Sakura Qwen3-Coder reaches 88.4 % compilability and 73 % coverage overlap, while Sakura Devstral reaches 85.4 % compilability and 64 % coverage overlap. The paper further reports 83.8 % instantiated-type recall, 88.3 % assertion-type recall, 73.3 % method-call recall, and 69.6 % focal-method recall under G-2.5 Pro.

In NLP evaluation, Sakura at BEA 2026 names two systems for vocabulary difficulty prediction (Nohejl et al., 14 May 2026). The high-accuracy black-box model fine-tunes GLM-4-32B, Qwen2.5-32B, and Mistral-3-14B with QLoRA and a soft-target cross-entropy loss

$N$8

followed by probability-weighted score decoding. The open-track ensemble reports mean Pearson correlation $N$9 and RMSE $\sigma=2.23$0, with individual LLMs at $\sigma=2.23$1. The explainable model uses XGBoost regression and SHAP, reaching mean Pearson $\sigma=2.23$2 and RMSE $\sigma=2.23$3 on the closed-track test, outperforming a fine-tuned XLM-RoBERTa baseline at $\sigma=2.23$4 and RMSE $\sigma=2.23$5. The feature analysis identifies production frequency as the strongest predictor and highlights the roles of spelling difficulty, lexical ambiguity, CEFR level, and test-item trickiness.

Taken together, these ML and software-engineering uses of SAKURA are methodologically heterogeneous. One is a contrastive self-supervised representation learner over pose sequences, one is a multi-agent code-generation framework with execution feedback, and one is a shared-task system for scalar difficulty prediction. The shared name does not correspond to a shared modeling paradigm.

5. SAKURA in scientific computing, cryptography, and instrumentation

In computational astrophysics, Sakura denotes a Keplerian-based Hamiltonian splitting for the gravitational $\sigma=2.23$6-body problem (Ferrari et al., 2014). The method rewrites pairwise gravitational terms as the difference between a two-body Kepler Hamiltonian and a reduced-mass kinetic term, yielding a first-order map that is not symplectic and then a time-symmetric second-order map

$\sigma=2.23$7

Its key practical property is that each pairwise orbit is evolved by an exact two-body Kepler solver. The method is reported as especially suitable for near-Keplerian systems and systems with intrinsic hierarchies, such as clusters with primordial binaries or stars orbiting a supermassive black hole. In parallel scaling tests, the authors report $\sigma=2.23$8 efficiency with only 8 particles per core, and close to perfect scaling for 16,384 particles on a 128-core distributed-memory machine. The paper also reports that for a 128-body Plummer + central mass system with $\sigma=2.23$9, Sakura was 5–6 orders more accurate and $\sigma=12.6$0 orders faster than Hermite at comparable energy-error levels.

In cryptographic hashing, Sakura coding is the framing scheme used to construct parallel SHAKE trees over Keccak permutations (Atighehchi, 2016). The paper on parallel SHAKE optimization treats Sakura as a fully specified method for turning RawSHAKE into a tree-hash mode that is automatically indifferentiable from a random oracle. The analysis emphasizes message hops, chaining hops, and kangaroo hops, then derives depth and processor formulas such as

$\sigma=12.6$1

For SHAKE256, the paper gives, for example,

$\sigma=12.6$2

and reports an 8-call depth with under 910 processors for a one-megabit message under the compacted ternary tree.

In educational instrumentation, SAKURA names a palm-sized two-dimensional muon scintillation detector built from a 5×5 array of 10 mm $\sigma=12.6$3 10 mm $\sigma=12.6$4 10 mm CsI(Tl) crystals and only four SiPMs (Matsushita et al., 8 Sep 2025). The detector uses a single 5 V USB line boosted to 29.5 V, RC-shaped outputs of width $\sigma=12.6$5, and a Red Pitaya STEMlab 125-14 for digitization. Position reconstruction is performed by a center-of-gravity rule,

$\sigma=12.6$6

and beam-test residuals at CERN T10 yield 13.4 mm resolution along $\sigma=12.6$7 and 7.48 mm along $\sigma=12.6$8. The full cost breakdown sums to $790, and the paper states that the system can be built for less than 1,000 USD. Its educational significance lies in achieving two-dimensional radiation imaging with four SiPM channels, a single FPGA-ADC board, and a workflow performed by high-school students.

These usages share a concern with efficient structure—pairwise analytic decomposition, unambiguous tree framing, or low-channel-count readout—but their technical content is otherwise unrelated.

6. SAKURA as a theoretical model and as a naming pattern

In space-plasma theory, the Sakura model is a phenomenological framework for the Earth’s stretched magnetotail current sheet built from percolation theory, fractional transport, and a weak-turbulence/linear-response picture (Milovanov, 2012). As the active-site fraction $\sigma=12.6$9 approaches the percolation threshold $$\mathrm{Acc}_{\rm single}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^s=y_i^s), \qquad \mathrm{Acc}_{\rm multi}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^m=y_i^m),$$0, the correlation length scales as

$$\mathrm{Acc}_{\rm single}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^s=y_i^s), \qquad \mathrm{Acc}_{\rm multi}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^m=y_i^m),$$1

while the ac conductivity obeys

$$\mathrm{Acc}_{\rm single}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^s=y_i^s), \qquad \mathrm{Acc}_{\rm multi}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^m=y_i^m),$$2

with $$\mathrm{Acc}_{\rm single}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^s=y_i^s), \qquad \mathrm{Acc}_{\rm multi}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^m=y_i^m),$$3 in two dimensions. The associated memory kernel leads to a fractional diffusion equation

$$\mathrm{Acc}_{\rm single}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^s=y_i^s), \qquad \mathrm{Acc}_{\rm multi}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^m=y_i^m),$$4

with $$\mathrm{Acc}_{\rm single}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^s=y_i^s), \qquad \mathrm{Acc}_{\rm multi}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^m=y_i^m),$$5. One of the model’s observational consequences is a magnetic fluctuation spectrum

$$\mathrm{Acc}_{\rm single}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^s=y_i^s), \qquad \mathrm{Acc}_{\rm multi}=\frac1N\sum_{i=1}^N \mathbf{1}(\hat y_i^m=y_i^m),$$6

which the chapter connects to AMPTE and INTERBALL-1 observations in the near-Earth tail. The model also predicts a transition from steady SOC to fishbone-like bursting when the driving rate exceeds a scaling threshold.

Across all of these works, the principal encyclopedic point is not that SAKURA designates one mature research ecosystem, but that it has become a reused technical name attached to sharply different artifacts: a benchmark exposing failures of multimodal latent-fact integration, an open Ethernet AI supercomputing platform, a side-channel testbed board, a self-supervised pose-learning objective, an agentic software-testing pipeline, a vocabulary-difficulty predictor, a Kepler-splitting integrator, a low-cost muon imager, a Keccak framing scheme, and a magnetotail SOC model. Any technical discussion of “SAKURA” therefore requires immediate domain disambiguation.