Search arXiv⌕ Search

arXiv subjects

:

Publications and source records attributed to :.

At least 19 recordsLinked to original sources
arXiv2026

First measurement of the masses of the $Υ_1(1D)$ and $Υ_3(1D)$ states and the energy dependence of the cross sections for $e^+e^-\toΥ_J(1D)η$ and $e^+e^-\toΥ_J(1D)π^+π^-$

We study the processes $e^+e^-\toΥ_J(1D)η$ and $e^+e^-\toΥ_J(1D)π^+π^-$ at center-of-mass energies $\sqrt{s}$=(10.73 -- 11.02) GeV using a $142.5\,\mathrm{fb}^{-1}$ data sample, including 122~fb$^{-1}$ near the $Υ$(10860) peak ($\sqrt{s}$ = 10.866 GeV), collected with the Belle detector at the KEKB asymmetric-energy $e^+e^-$ collider. From the peak sample, the products of Born cross section times branching fraction are obtained for $σ_{\rm Born}(e^+e^-\toΥ_J(1D)η)$ or $σ_{\rm Born}(e^+e^-\toΥ_J(1D)π^+π^-)$ and ${\cal B}(Υ_J(1D)\toχ_{b1}γ)$ or ${\cal B}(Υ_J(1D)\toχ_{b2}γ)$ for each $Υ_J(1D)$ state. The corresponding branching fractions for $Υ(10860)$ decays are also obtained. The significances of the $Υ_1(1D)$, $Υ_2(1D)$, and $Υ_3(1D)$ signals are 4.8$σ$, ${>}10σ$, and 3.0$σ$, respectively, including systematic uncertainties. The mass for $Υ_2(1D)$ is measured to be $(10167.0\pm 1.0\pm 0.2)$ MeV/$c^2$, where the first and second uncertainties are statistical and systematic. The mass splittings $Δm_{12}=m(Υ_2(1D))-m(Υ_1(1D))$ and $Δm_{23}=m(Υ_3(1D))-m(Υ_2(1D))$ are $(11.8\pm1.5\pm0.4)$ MeV/$c^2$ and $(7.6\pm2.4\pm0.6)$ MeV/$c^2$, respectively.~We determine the energy dependence of the cross sections for $e^+e^-\toΥ_J(1D)η$ and $e^+e^-\toΥ_J(1D)π^+π^-$ for the $Υ_1(1D)$, $Υ_2(1D)$, and $Υ_3(1D)$ states, combined.

hep-ex↗
arXiv2026

Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI

We introduce Nemotron 3.5 Super and Nemotron 3.5 Lightning, open and efficient Mixture-of-Experts Hybrid Mamba-Transformer models designed for reasoning, long-context understanding, and agentic workloads. Super is a 120B-parameter, 12B-active multimodal model, while Lightning is a smaller 30B-parameter, 3B-active text model; both designed for efficient deployment. Both models combine large-scale distillation, long-context training, and native speculative decoding, while Super further incorporates Indic continued pretraining and agentic software-engineering mid-training. We develop a comprehensive post-training pipeline spanning supervised fine-tuning and reinforcement learning (RLVR and RLHF), complemented in Super by specialized teacher training and multi- teacher on-policy distillation (MOPD). This transforms Super from a text-only base model into a multimodal agent capable of reasoning over text, images, video, and interactive tool environments. We further explore test-time scaling through parallel agent execution and self-verification. Across training and deployment, we jointly optimize capability and efficiency. Indic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points, while agentic mid-training strengthens software-engineering capabilities and robustness across agent harnesses. The post-trained Super model achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra despite using approximately 4.6x fewer active parameters. Post-training MTP boosting improves speculative- decoding acceptance length by up to 21.8% under probabilistic draft sampling, while NVFP4 quantization achieves 99.62% median accuracy recovery relative to BF16. Inference optimizations enable up to 4,856 output tokens/s/GPU on GB200 while sustaining at least 50 tokens/s/user.

cs.CL↗
arXiv2026

Measurements of the branching fractions of $Λ_c^+\toΛK^+ π^0$ and $Λ_c^+\toΛK^+ η$ decays at Belle and Belle II

We present a study of the three-body, singly Cabibbo-suppressed decays $Λ_c^+ \to ΛK^+ π^0$ and $Λ_c^+ \to ΛK^+ η$ using data recorded with the Belle and Belle II detectors at the KEKB and SuperKEKB $e^+ e^-$ colliders. The data samples correspond to integrated luminosities of 980$\,\mathrm{fb}^{-1}$ and 570$\,\mathrm{fb}^{-1}$, respectively. We measure their branching-fraction ratios relative to the Cabibbo-favored decay $Λ_c^+ \to Λπ^+ π^0$ as $\mathcal{B}(Λ_c^+ \to ΛK^+ π^0)/\mathcal{B}(Λ_c^+ \to Λπ^+ π^0)=(1.82 \pm 0.07 \pm 0.07)\%$, with significantly improved precision compared to the previous measurement, and $\mathcal{B}(Λ_c^+ \to ΛK^+ η)/\mathcal{B}(Λ_c^+ \to Λπ^+ π^0)=(8.73 \pm 0.35 \pm 0.43)\times10^{-3}$, corresponding to the first observation of the $Λ_c^+ \to ΛK^+ η$ decay with a statistical significance exceeding 10 standard deviations. The first and second uncertainties are statistical and systematic, respectively. Using the world-average value of the branching fraction for $Λ_c^+ \to Λπ^+ π^0$, we obtain the branching fractions $\mathcal{B}(Λ_c^+ \to ΛK^+ π^0)=(1.29\pm 0.05\pm 0.05\pm 0.06)\times10^{-3}$ and $\mathcal{B}(Λ_c^+ \to ΛK^+ η)=(6.20\pm 0.25\pm 0.31\pm 0.30)\times10^{-4}$, where the third uncertainty arises from the uncertainty in the world-average value of $\mathcal{B}(Λ_c^+ \to Λπ^+ π^0)$. The measured $\mathcal{B}(Λ_c^+\toΛK^+η)$ is significantly larger than predictions based on $SU(3)$ flavor symmetry.

hep-ex↗
arXiv2026

Kolibri: A Sovereign European Model on the Pareto Frontier

This report introduces Kolibri, an English-German mixture-of-experts transformer with 78.1B total parameters and 3.46B active parameters per token, released as open weights under the Apache 2.0 licence. We built Kolibri for sovereign and specialised deployment, with a particular focus on German and on regulated domains such as public administration, industry, and aerospace. To support these settings, we control the full model-development process, including data, architecture, training infrastructure, post-training, and evaluation. We train Kolibri on 24T tokens across pre-training, mid-training, and long-context extension, with German accounting for more than 20% of the data mix by bytes. To reach this share, we complement German datasets with more than 2T German tokens that we curate from the web with our own pipeline and generate synthetically. Post-training combines supervised fine-tuning on curated German and English data with reinforcement learning on more than 1.2M internally curated tasks, which cover reasoning, tool use, instruction following, code, and retrieval. The model activates 4.4% of its parameters per token, and among the open models evaluated in this report, it lies on the Pareto frontier of quality and serving cost in English and German. Kolibri builds on Kolibri Origin, an earlier iteration that validated our full pipeline at scale. Both models come from our Model Factory, which connects data, experimentation, training, evaluation, and deployment through version-controlled workflows.

cs.CL↗
arXiv2026

Foundation4D: Towards a unified perception and reconstruction model for autonomous vehicles

We introduce Foundation4D, a unified feed-forward model for geometry estimation, 3D detection, semantic occupancy, and Gaussian reconstruction from multi-camera driving video. A shared backbone uses factorized spatiotemporal attention to efficiently combine observations across cameras and time. Our video inference pipeline uses pose graph optimization to refine camera trajectories and assembles object tracks across input windows. We also develop a data annotation pipeline and release PAI-Foundation4D, a benchmark with dense depth and semantic occupancy labels for 1,253 driving clips. With benchmark-specific fine-tuning, Foundation4D achieves the lowest depth AbsRel on Waymo and nuScenes, the highest detection mAP on nuScenes, and the highest occupancy mIoU on OpenOccupancy among the compared methods. Our task-combination study shows benefits for both detection and geometry: co-training detection with geometry and occupancy improves top-class mAP by 1.2 points over detection-only training, while co-training geometry with detection reduces camera-to-rig translation error by $13.4\%$ relative to geometry-only training. We further use the predicted geometry and trajectories as priors for lidarless NuRec, supporting reconstruction of driving scenes for simulation.

cs.CV↗
arXiv2026

LLM-jp-4 Technical Report

This paper presents LLM-jp-4, a family of Japanese sovereign large language models (LLMs) developed with an emphasis on transparency, reproducibility, and practical usability. LLM-jp-4 comprises an 8B dense model and a 32B-A3B mixture-of-experts model, both trained from scratch on approximately 12 trillion tokens. To support independent verification and reproduction, we establish explicit criteria for selecting training data, external models, and software, and make the development process publicly available. Evaluation results demonstrate substantial improvements over previous LLM-jp models and competitive performance against open-weight models of similar scale. By documenting the development principles, training procedures, and empirical findings of LLM-jp-4, we aim to provide a reference for future sovereign AI development.

cs.CL↗
arXiv2026

Search for the $\boldsymbol{B^0 \to K^0_{\rm S} τ^+ τ^-}$ decay

We present the first search for $B^0 \to K^0_{\rm S} τ^+τ^-$ decays. We look for signal decays in $B^0\bar B^0$ events produced in asymmetric-energy electron-positron collisions. This work uses samples from the Belle and Belle~II detectors, comprising 1.16 billion $Υ(4S)$ events. In $Υ(4S)\to B^0\bar{B}^0$ decays, the non-signal $\bar{B}^0$ meson is fully reconstructed in a hadronic channel. For the signal $B^0$ meson, $τ$-lepton decays into final states with a single charged particle are selected. We observe no evidence for the signal and set an upper limit on the branching fraction $\mathcal{B}(B^0\to K^0_{\rm S} τ^+τ^-) < 8.3 \times 10^{-4}$ at the 90\% confidence level. Combining this with the recent measurement of the isospin-partner decay $B^+\to K^+τ^+τ^-$, we determine an upper limit $\mathcal{B}(B\to Kτ^+τ^-) < 5.4\times10^{-4}$ at the 90\% confidence level.

hep-ex↗
arXiv2026

Apertus 1.5: From Open Text Models to Open Multimodal Reasoners

Language models are becoming infrastructure, and infrastructure built in the open, by public institutions, on documented data, is the alternative to dependence on a few providers; Apertus shows this is viable at scale. Apertus 1.5 extends the fully open Apertus family from text to multimodal reasoning. Continued pretraining of the Apertus 1.0 models, on 4T additional tokens for the 8B and 2T for the 70B, adds image and audio understanding and extends the context window fourfold to 262,144 tokens; multimodality enters through the tokenizer, the language-model head remains restricted to text, and the architecture is unchanged. Supervised finetuning, long-context finetuning, reinforcement learning with verifiable rewards, and preference alignment complete the released models. Thinking is a flag of the chat template rather than a separate model, and lifts mathematical reasoning by up to 20 points on the 8B model. The released models improve over their predecessors on nearly every benchmark; among fully open models of its class, the 8B leads on knowledge, instruction following, and cultural benchmarks, and its instruction following surpasses all compared models, including open-weight ones. On image and audio understanding, a first for the family, the models are competitive with fully open multimodal models, strongest on document understanding and long-form and lower-resource speech recognition, while specialized vision and audio models remain ahead on perception and transcription of English and other high-resource languages. As in Apertus 1.0, we train only on openly available data, respect crawler opt-outs retroactively, and exclude restrictively licensed material. We release the model weights, data preparation and training recipes, and evaluation suites under permissive licenses; openness extends to values, with the Apertus charter guiding alignment stages published alongside the models.

cs.CL↗
arXiv2026

Benchmarking exchange-only control of a 48-spin singlet manifold

Exchange-only quantum computing benefits from high fidelity and straightforward control afforded by the exchange interaction. However, independently controllable qubits must be encoded into subsystems of at least three electron spins, restricting computation to only a fraction of the available spin Hilbert space. The remaining states are treated as leakage and used only transiently in gate sequences. By contrast, a quantitative, system-wide measure of exchange-only performance can exploit the full available Hilbert space, including states conventionally treated as leakage. Here, we apply this approach to arrays of up to 48 electron spins, accessing a total-spin-zero Hilbert space with dimension exceeding $2^{40}$. Measurements of out-of-time-order correlators (OTOCs) reveal rich scrambling dynamics and demonstrate access to regimes relevant to quantum computational advantage. Using generalized forms of cross-entropy and mirror randomized benchmarking, we also assess the aggregate performance of the full exchange-only system per fundamental two-body interaction: the two-spin exchange. We obtain an effective system-level error of $3 \times 10^{-4}$ per exchange, incorporating the complete experimental control sequence and all associated noise sources. This value is up to an order of magnitude lower than those reported from comparable benchmarks on other platforms at the time of writing.

quant-ph↗
arXiv2026

Observation of the decays $B^{+} \to Σ_{c}(2455)^{++} \barΞ_{c}^{\prime-}$ and $B^{0} \to Σ_{c}(2455)^{0} \barΞ_{c}^{\prime0}$

We report the first observation of the decays $B^{+} \to Σ_{c}(2455)^{++} \barΞ_{c}^{\prime-}$ and $B^{0} \to Σ_{c}(2455)^{0} \barΞ_{c}^{\prime0}$, with significances of $6.4\,σ$ and $5.3\, σ$, respectively, including systematic uncertainties. This analysis is based on data samples containing $771.6 \times 10^{6}$ $Υ(4S)$ decays collected with the Belle detector at the KEKB collider and $520.6 \times 10^{6}$ $Υ(4S)$ decays collected with the Belle~II detector at the SuperKEKB collider. The branching fractions are measured to be $\mathcal{B}(B^+ \to Σ_c(2455)^{++} \barΞ_c^{\prime -}) = (1.69 \pm 0.31 \pm 0.12^{+1.39}_{-0.56}) \times 10^{-3}$ and $\mathcal{B}(B^0 \to Σ_c(2455)^{0} \barΞ_c^{\prime 0}) = (1.28 \pm 0.32 \pm 0.10^{+0.26}_{-0.22}) \times 10^{-3}$, where the first and second uncertainties are statistical and systematic, respectively, and the third arises from the uncertainties in the absolute branching fractions of $\barΞ_{c}^{-}$ and $\barΞ_{c}^{0}$ decays. This result represents the first observation of $B$-meson decays into a pair of charmed baryon-antibaryon states belonging to the same $SU(3)$ flavor sextet.

hep-ex↗
arXiv2026

Measurement of time-dependent $CP$ violation parameters in $B^{0} \to K_{S}^{0} π^{0} γ$ decays at Belle and Belle II

We perform a measurement of time-dependent $CP$ violation parameters in $B^{0} \to K_{S}^{0} π^{0} γ$ decays using a dataset of approximately $772 \times 10^6$ and $521 \times 10^6$ $Υ(4S)$ decays collected by the Belle and Belle II experiments, respectively. The measured parameters for the combined dataset in the $K^{*0}(892)$ dominated region ($M_{K_{S}^{0} π^{0}} \in [0.8,1.0] \mathrm{GeV}/c^2$) are $S = 0.09 \pm 0.16 \pm 0.02$ and $C = -0.09 \pm 0.08 \pm 0.04$. For the non-$K^{*0}(892)$ region ($M_{K_{S}^{0} π^{0}} \in [1.0,1.8] \mathrm{GeV}/c^2$), the corresponding values are $S = -0.32 \pm 0.33 \pm 0.09$ and $C = -0.07 \pm 0.17 \pm 0.08$. The first quoted uncertainties are statistical, while the second ones are systematic. These results are consistent with Standard Model predictions and more precise than previous measurements.

hep-ex↗
arXiv2026

The Platonic Universe: Do Foundation Models See the Same Sky?

We investigate when foundation models converge towards shared representations, and how this convergence depends on model capacity, training regime, and model architecture. We take a `science-for-AI' approach, using astronomy as an experimental instrument to test the Platonic Representation Hypothesis and its Aristotelian refinement against an external physical reference. The historical success of astrophysics is evidence that a compact, modality-invariant description of galaxy observables exists, and so representation convergence toward reality should be measurable against the physical parameters astronomers already use. Given this framework, we evaluate eleven foundation model families (spanning classification, self-distillation, joint-embedding prediction, autoencoding, vision-language pre-training, and astro-specific architectures from $\mathcal{O}$(10M)${\to}\mathcal{O}$(10B) parameters) on crossmatched JWST, HSC, and Legacy imagery, and DESI spectroscopy. All models are evaluated frozen, with no astronomy-specific fine-tuning. We probe redshift, stellar mass, and sSFR via linear probes, and local (MKNN) and global (CKA) embedding geometry within families, between modalities, and across architectures. We find that physics performance scales predictably with capacity; probe directions align consistently with expected astrophysical correlations and selection effects; and local (not global) embedding alignment tracks physics performance, including between DESI spectra and HSC imagery---modalities that share essentially no low-level statistics. Our results support the ARH over the strict PRH, demonstrate astronomy's value as an experimental framework for neural representation learning, and suggest that astro-foundation models can build on general-purpose pre-trained architectures, capitalizing on the broader open machine learning community's already-spent computational investment.

astro-ph.IM↗
arXiv2026

A Search For Stellar-mass Black Holes Via Astrometric Microlensing II: 2012-2015 Keck Candidates

The Milky Way is expected to host $\sim$10$^8$ stellar-mass black holes with an uncertain binary fraction. The only proven method to detect isolated stellar mass black holes is gravitational microlensing. Here we report the results of a microlensing search for black holes with photometry and astrometry. By combining 10 years of seeing-limited photometry from OGLE and MOA with diffraction-limited photometry and astrometry from adaptive optics imagers at the W.~M.~Keck Observatory, we constrain lens masses for OGLE-2012-BLG-0169, OGLE-2014-BLG-0613/MOA-2015-BLG-041, OGLE-2015-BLG-0029/MOA-2015-BLG-170, and OGLE-2015-BLG-0211. Of the four long-duration microlensing events monitored, we ruled out black hole lenses in 3 events, which likely have stellar or white dwarf lenses. OGLE-2015-BLG-0211 remains a black hole candidate with a poorly constrained lens mass with a 1$σ$ upper mass limit of 3.2$M_\odot$ and a 3$σ$ upper mass limit of 21.6$M_\odot$. This event suffered from poor observing conditions and significant astrometric reference frame uncertainties, but its analysis may benefit from additional astrometric data in the upcoming Gaia Data Release 4. Of the six long-timescale ($t_E>100$ days) microlensing events from this work and previous studies, one black hole has been confirmed with a second not ruled out. We briefly examine Galactic model simulations and find that our result agrees with current expectations. Ultimately, we need a larger sample of isolated black holes to constrain their formation processes. This will be possible in the coming years with Rubin and Roman, as well as improved astrometry from JWST and large, ground-based telescopes equipped with adaptive optics.

astro-ph.SR↗
arXiv2026

Environmental sustainability in basic research: a perspective from HECAP+

The climate crisis and the degradation of the world's ecosystems require humanity to take immediate action. The international scientific community has a responsibility to limit the negative environmental impacts of basic research. The HECAP+ communities (High Energy Physics, Cosmology, Astroparticle Physics, and Hadron and Nuclear Physics) make use of common and similar experimental infrastructure, such as accelerators and observatories, and rely similarly on the processing of big data. Our communities therefore face similar challenges to improving the sustainability of our research. This document aims to reflect on the environmental impacts of our work practices and research infrastructure, to highlight best practice, to make recommendations for positive changes, and to identify the opportunities and challenges that such changes present for wider aspects of social responsibility.

physics.soc-ph↗
arXiv2026

Search for the decay $B^+ \rightarrow K^+τ^+τ^-$ using data from the Belle and Belle II experiments

We report a search for the rare decay $B^{+} \rightarrow K^{+} τ^{+} τ^{-}$ using $1.2 \times 10^9$ $Υ(4S)$ mesons produced near threshold in electron-positron collisions and collected by the Belle and Belle~II experiments. We fully reconstruct the hadronic decay of one $B$ meson produced in the $Υ(4S)\rightarrow B^{+} B^{-}$ decay, and search for $B^{\pm}\rightarrow K^{\pm} τ^{+}τ^{-}$ candidates among the remaining collision products, reconstructing a charged kaon and leptonic decays of the $τ$ leptons. We optimize the selection for best sensitivity and look for an excess over background at low values of the residual energy detected in the calorimeter after full event reconstruction. We observe no significant excess and set the limit $\mathcal{B}(B^{+}\rightarrow K^{+}τ^{+}τ^{-})< 0.56\times 10^{-3}$ at the 90% confidence level, improving on the only previous result by a factor of four.

hep-ex↗
arXiv2026

DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression

The widespread adoption of long-horizon agents has made model workloads increasingly input-heavy. Although prior work has substantially reduced the cost of long-context computation, prefill remains computationally expensive, and large KV caches continue to strain HBM and SSD capacity and data-transfer bandwidth. Together, these compute, storage, and bandwidth demands constitute the primary bottleneck to further lowering deployment costs. To address this challenge, we introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. With its Causal Encoder-Decoder (CED) architecture, the model activates 16B parameters per token during decode but only 8B parameters during prefill, substantially improving cost efficiency for agentic workloads. To push the limits of KV cache compression, DeepSeek-V4.1-Flash combines cross-layer KV cache reuse in Compressed Sparse Attention 2 (CSA2) with FP4 KV caching. These designs reduce its global KV cache footprint (always in HBM) to 890 bytes per token, roughly 1/4 of the corresponding footprint of DeepSeek-V4-Flash. Further, through a dedicated deployment optimization known as SWA Bounded Replay, DeepSeek-V4.1-Flash reduces its persistent KV cache footprint (always on SSD or in host memory) to roughly 1/8 of that of DeepSeek-V4-Flash. Despite its much smaller KV cache footprint, the model delivers substantially better performance than the baseline. In addition, we streamline the DeepSeek-V4 architecture and introduce several efficient architectural extensions. We pretrain DeepSeek-V4.1-Flash on a multimodal corpus comprising 45T tokens and conduct comprehensive post-training, yielding strong performance across diverse text-based and multimodal agentic scenarios. Model checkpoints are available at https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash.

cs.CL↗
arXiv2026

Improving constraints on the Yukawa correction at the Galactic Center with multiple stellar orbits

We investigate the presence of a Yukawa-like correction ($\propto \, αe^{- r/λ}$) to Newtonian gravity at the Galactic Center, using a multi-star fitting code, including the newly discovered star S$301$, to improve the constraints obtained using S$2$ orbit alone. We perform a Markov Chain Monte Carlo analysis using the astrometric and spectroscopic data of stars S$2$, S$55$, S$29$, S$38$ and S$301$ collected by GRAVITY, GRAVITY$^+$, NACO and SINFONI instruments, covering the period from $1992$ to $2025$. Compared to GRAVITY Collaboration 2025 (Paper I), in which only S$2$ was fitted, the tightest bound on the Yukawa coupling is again reached at $λ= 3\cdot10^{13}\,\rm m\ (\sim 200\,AU)$, where we find $|α| < 6 \cdot10^{-4}$, an improvement of roughly a factor five. The addition of S$301$ extends the constraint to short ranges that were inaccessible to S$2$, yielding $|α| < 0.004$ at $λ=10^{12}\,\rm m$ and $|α| < 0.006$ at $λ=5\cdot 10^{11}\,\rm m$, while S$29$, with its larger apoapsis, gives $|α| < 0.1$ at $λ\sim 2\cdot10^{15}\,\rm m$. The latter two represent the tightest limits at these scale lengths on a Yukawa correction obtained to date around a supermassive black hole.

astro-ph.GA↗
arXiv2026

Miles v0.1: Production-Level Post-Training

We present Miles v0.1, a full-stack, production-ready system for frontier post-training. Building upon the clean design of slime, Miles designs each stage of the reinforcement-learning (RL) training loop around a single principle: components should be verified, clean, and customizable. With accuracy, efficiency, reliability, and scalability as first-class goals, Miles aims to make frontier-scale RL accessible to researchers and enterprises alike. This report walks through the system end to end: rollout engines built on SGLang, a trainer with a choice of two backends (NVIDIA Megatron-LM and PyTorch FSDP), and three weight-synchronization transports for different deployment topologies. Beyond full-parameter RL, Miles also supports LoRA RL, on-policy distillation, supervised fine-tuning, and true-on-policy rollout-training alignment, and extends the same architecture to diffusion models. We close with an end-to-end case study: fully asynchronous agentic RL on a GLM-5.2 744B-A40B model over terminal-use coding tasks, running on 64 NVIDIA GB300 GPUs with a median step time of 263 seconds over the first 30 measured steps. Miles is open-sourced at https://github.com/radixark/miles, with the project website at https://miles.radixark.com.

cs.LG↗