We study post-training for dynamically routed mixture-of-experts (MoE) models, focusing on integrating independently optimized specialists while keeping consolidation efficient. Our approach combines calibrated multi-teacher on-policy distillation (C-MOPD), which matches each selected teacher on a local Top-$K$ support with teacher-wise loss normalization and clipping, with an execution path that uses post-capacity bounded-width dispatch, final-route replay, and asynchronous rollout--training overlap. On Turing-20B-A2B, the selected model achieves a four-domain normalized integration score of 0.9988 and retains at least 90.16\% of the measured specialist gain in every domain. At matched $K=32$, the full calibration objective scores 0.9962, compared with 0.9724 without calibration and 0.9834 without clipping. At a matched 8K response cap, the model exceeds both shared SFT and Qwen3-8B on all four reasoning-enabled mathematics and code benchmarks (unweighted mean: 85.92, versus 80.33 for Qwen3-8B and 74.21 for SFT), and improves over SFT on ten of twelve broader evaluation metrics. In a PPU-based profile, the combined execution configuration reduces consolidation step time from 398.7 to 139.3 seconds, an observed 2.86$\times$ speedup over serial execution. These results demonstrate strong specialist integration alongside more efficient consolidation, while reflecting the complete training and execution framework rather than any single component.
A streaming estimator must compute from the evidence that fits in memory. We determine when storing an accurate mean is easier than computing it. For Gaussian mean estimation over the Euclidean unit ball in $\mathbb{R}^d$, we establish the minimax rate under squared error with $K$ persistent states, a known horizon, and a public time index: $\min\{1, K^{-2/d}+\frac{d}{n} (1+\frac{\log(ed)}{1+(\log K)/d})\}$. Unrestricted computation followed by $K$-level encoding has rate $\min\{1,d/n+K^{-2/d}\}$. At a fixed sufficiently small error, statistically optimal sample size therefore requires order $d\log d$ working bits, while order $d$ bits suffice to store the answer. The lower bound tracks repeated state compression through an entropy potential obtained by Gaussian smoothing. A matching construction combines a multiscale modular counter with workspace reused across coordinate groups. In one dimension, it attains the terminal rate $n^{-1}+K^{-2}$. Any fixed confidence requirement that the realized loss not exceed the zero baseline preserves these rates. Increasing confidence creates a further memory requirement: at fixed nontrivial risk in every fixed dimension, failure probability $e^{-γn}$ requires and permits order $\log n$ working states for sufficiently small fixed $γ>0$. In the scalar problem, a fixed terminal alphabet suffices under the same protection. These bounds separate the memory needed to represent an answer from the memory needed to accumulate evidence and validate a decision.
Serverless platforms increasingly host machine learning (ML) inference workflows, modeled as directed acyclic graphs (DAGs) of interdependent functions. Beyond container startup, each ML node must load libraries, model weights, and CUDA kernels before it can run. These costs accumulate along dependent DAG paths and place initialization on the end-to-end critical path. Existing systems optimize one half of this problem. DAG-agnostic pre-loading prepares nodes the request may never reach, wasting host and GPU memory. DAG-aware serving routes and scales whole functions, leaving the loading stages inside a node untouched. This paper presents GearUp, a serverless DAG inference serving system that removes ML-node loading from the end-to-end critical path at low resource cost. GearUp profiles the loading stages of each node and schedules them inside the slack time the DAG exposes, hiding initialization behind upstream execution. It reuses initialized instances within and across requests to avoid redundant cold starts. A probabilistic, traffic-adaptive keep-alive policy sets how long idle instances are retained, balancing memory cost against cold-start risk under bursty load. On the AMBER Alert and FINRA-ML workflows with NVIDIA L40S GPUs, GearUp reduces median end-to-end latency by 64.5% and 75.9% against SMIless, for median speedups of 3.1x and 10.0x over cold execution. Integrated GPU and host memory allocation fall by up to 83.9% and 94.6% against eager pre-loading in InstaInfer. GearUp achieves 4.2x strong scaling on four workers.
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.
Zehao Wang, Hanzhu Chen, Runquan Gui, Zhe Wang, Lin Yang
Reinforcement Learning with Verifiable Rewards (RLVR) derives its learning signal from binary rewards via on-policy exploration and becomes a powerful paradigm for improving the reasoning of large language models. Yet it faces a learning cliff on problems beyond model's current ability: when every rollout fails, the samples reveal nothing about a correct solution, and group-relative estimators (e.g., GRPO, RLOO) collapse the update to zero advantage. Expert trajectories offer a remedy, but they are incompatible with the current policy and often lie in the student's low-probability regions, so the surviving updates demand large parameter shifts and decay into surface-level imitation rather than transferable reasoning. We argue that a useful learning signal on hard problems must be both verifiably correct and sufficiently compatible with the current policy. We propose Policy-Compatible Reinforcement Learning (PICO), which performs distribution-constrained joint decoding: expert data guides decoding at the token level while preserving the student's relative token probabilities, and a dynamic top-k mechanism confines the sampling distribution under a finite divergence budget, yielding trajectories that are verified and policy-compatible by construction. These trajectories are optimized jointly with on-policy rollouts under a mixed-behavior objective that mitigates their distribution mismatch via importance sampling. Experiments across reasoning diverse datasets and agentic benchmarks demonstrate that PICO consistently outperforms standard RLVR and prior off-policy baselines.
Yuhao Wang, Zelin Zang, Yuxuan Liu, Zhen Lei, Stan Z. Li
Single-cell transcriptomes are sparse, unordered measurements of coordinated biological activity, yet many self-supervised models define their learning target at the level of individual genes or reconstructed expression values. Here we introduce GB-JEPA, a joint-embedding predictive architecture in which graph-connected gene blocks form the latent prediction units. A student encoder observes the genes outside each target block, whereas an exponential-moving-average teacher represents the full observed gene set. Teacher states are aggregated within a block before the prediction error is evaluated, converting thousands of sparse gene-level targets into a small collection of intermediate-scale latent targets. The graph therefore serves not only as prior information, but also as a mechanism for controlling the structure, scale and predictability of self-supervision. We evaluate the resulting frozen cell representations through objective diagnostics, cell-state annotation, continuous-expression and pathway probes, perturbation-response prediction, computational measurements and direct biological interrogation. This study presents a graph-structured route for adapting JEPA learning to molecular observations and a general framework for designing prediction units from relationships native to a biological modality.
Zhoufu Wang, Baoshen Guo, Zhiqing Hong, Junyi Li, Kailai Sun, Heye Huang, Alok Prakash, Shenhao Wang 等 9 位作者
Human mobility generation aims to synthesize realistic point-of-interest (POI) visitation trajectories and has become an important tool for travel behavior modeling, transportation management, and urban planning. Existing diffusion-based methods achieve high fidelity but require per-city generation, given the inherent heterogeneity of geospatial locations and POI categories, while large language model-based methods generalize across cities but remain too costly at scale, especially for long-horizon trajectory generation. To address this, we propose SeMoFlow, a Semantic human Mobility generation framework based on latent Flow matching. We first encode heterogeneous POIs from different cities into a shared cross-city representation space via hierarchical Semantic IDs, where shared prefixes capture transferable semantics, and successive codes progressively refine the representation toward individual POIs. Building on the semantic IDs, SeMoFlow follows a plan-to-synthesis hierarchical generation paradigm, in which an autoregressive planner generates coarse-grained semantic and recurrence patterns, and a flow matching realizer synthesizes fine-grained suffix latents. The generated latents are subsequently decoded and grounded to concrete POIs. Extensive experiments on large-scale multi-city datasets show that SeMoFlow achieves higher trajectory fidelity than existing baselines, preserves city-specific mobility motifs, and supports both joint multi-city generation and effective cross-city transfer. The code is available at \href{https://github.com/urban-mobility-generation/SeMoFlow}{https://github.com/urban-mobility-generation/SeMoFlow}.
Radio maps provide essential wireless propagation knowledge for network optimization and autonomous platforms. Acquiring dense measurements, however, is prohibitively costly, requiring active sampling to prioritize informative measurements. Existing active sampling methods typically rely on variances from prescribed spatial models or learned pointwise uncertainty, which may not fully capture the spatial correlation of different measurements. This paper proposes GraphBARC, a graph-structured active radio map construction framework that integrates learned spatial modeling with analytical Bayesian inference. A graph network learns correlated node and edge factors using only acquired measurements, inducing a joint full-map belief that supports analytical observation updates without retraining. Rather than immediate one-step acquisition, it adopts a nonmyopic sensing strategy that evaluates joint map-wide Bayes risk reduction over feasible multi-step routes, accounting for information redundancy among planned measurements. We reuse candidate responses and derive residual-based bounds directly on route-value differences, retaining shared response-error structure to avoid uniformly high-accuracy evaluation. An error-aware continuation bound screens unpromising prefixes, after which only responses needed for unresolved comparisons are refined. We further establish a computable decision-gap bound and show that certified exclusions remain valid during refinement. Experiments show that GraphBARC reduces mean root-mean-square error (RMSE) by over 16% relative to the evaluated baselines and core planning time by 56% compared with the fixed-accuracy, unpruned implementation.
Boyuan Meng, Peihua Bao, Hong Liu, Xiaowei Zhu, Chao Wang, Gen Li, Zhenxuan Pan
Agentic reinforcement learning (RL) generates irregular rollout trees, but trajectory-by-trajectory training recomputes their shared interaction histories. Existing training-time prefix-reuse systems do not jointly address state recovery, dense execution, and distributed training for hybrid-attention models. We present HARTS (Hybrid-Attention RL over Tree Structures), to our knowledge the first system to demonstrate training acceleration from arbitrary rollout-tree prefix sharing on real hybrid-attention models. A two-level planner constructs compression-oriented microbatches, then jointly refines their membership, cross-microbatch sharing, placement, and execution order. Within each sharing window, gradients from dependent microbatches accumulate before one backward pass through the shared prefix computation. Differentiable boundary-state handoff and bounded replay preserve the per-trajectory chunkwise numerical path, while a linear-time planner attains the global minimum number of sequential linear-attention calls under our packed chunkwise-call model. The runtime uses a common compact layout across attention and dense or mixture-of-experts (MoE) layers, with support for distributed parallelism and activation recomputation. Recovering log-probabilities for each original trajectory position preserves RL objectives; for deterministic top-k routing without token dropping, occurrence counts restore MoE objective weights and expert-load statistics. On SWE-bench, HARTS delivers 4.77-10.45x the update throughput of trajectory-expanded training across two sampling schemes on 7.9B- and 124B-parameter hybrid-attention MoE models using 8-64 GPUs. On the 7.9B model, full-model logit differences are comparable to baseline self-rerun variation. The 124B model shows stable online learning trends through step 150 on SWE-bench and $τ^3$-Bench.