machine learning platform engineers

Abstract: Predictive models trained at well-resourced data-rich institutions often fail when deployed in new populations. Motivated by the goal of building clinical diagnosis and decision-making algorithms for data-scarce rural populations using the general knowledge learned from data-rich hospitals, we formulate the fundamental question of \emph{domain generalization with adversarially chosen domains}, which asks a learner to perform on any unseen target domain nearly as well as the worst hypothesis consistent with the evidence from the observed source domains. Assuming only that the hypothesis class contains a "universal" predictor with small error on all $k$ source domains, we characterize the sample complexity of this objective up to logarithmic factors. Letting $d$ denote the VC dimension of the hypothesis class, we find, somewhat surprisingly, that the sample complexity reduces significantly from $Ω(dk)$ to $O(d)$ under a small relaxation of the learning goal in which, roughly speaking, we allow the learned model to compete with hypotheses that perform well on 99% instead of 100% of the source domains. By applying this insight to the meta-distributional setting, in which source domains are selected according to a distribution on distributions, we obtain a corresponding improvement over known sample complexity upper bounds. We also study the effect of data access to the target domain and show that a small labeled sample from the target domain does not improve the worst-case sample complexity on the source domains.

Abstract: Agentic reinforcement learning (RL) often requires repeated tool calls in task-specific environments. A training job may create thousands of these environments in bursts while sharing a cluster with other jobs. We present Sandman, a sandbox service built and operated at Amazon for agent tool execution and evaluation. It serves multiple production agentic RL workloads on shared Kubernetes. A training framework supplies a task image and uses commands and file operations in the returned sandbox. Files and processes persist across calls until the framework releases the environment. Sandman manages creation, readiness and cleanup. Each sandbox is a Kubernetes pod with a small daemon mounted into the unmodified image. In a scale test, Sandman held 11,858 concurrent tiny sandboxes. Four production SWE-bench Verified training campaigns created 3.1 million sandboxes in October 2026. We report the design trade-offs and operational lessons from handling creation bursts, preparing task environments, reclaiming abandoned sandboxes, and running evaluation under controlled conditions. We will open-source Sandman's daemon, provider and pool.

Abstract: We study conformal prediction under covariate shift with a known covariate density ratio, seeking prediction sets with minimum expected length under the target distribution. For every calibration sample size, we prove that the rule combining the inverse conditional-density score with randomized weighted conformal calibration minimizes this expected length among all rules valid over the full weighted-exchangeability class. We approximate the oracle using quantile and kernel histograms, including forest-weighted estimators, and select estimators and tuning parameters by resampling to estimate calibrated target length. On a bounded response domain, we bound excess length linearly in $L^1$ density error at fixed calibration size, without requiring a uniform positive lower bound on the true density. We also bound selection-pool and resampling errors. Under explicit growth conditions, the selected procedure approaches the same oracle while the calibration size remains fixed. Experiments on 36 known-density configurations and four public regression datasets compare coverage, prediction-set length, and runtime with existing methods. In targeted low-coverage and strong-shift comparisons, inverse-density selection yields shorter mean sets in most cases and lower mean runtime than the evaluated weighted baselines, with some differences in achieved coverage.

Abstract: Multilingual BabyLM systems usually fix language sampling proportions even though the weakest language can change during training. We present ARMOR-DP, a modular framework that diagnoses imbalance arising from representation, exposure, capacity, and objective pressure, and uses per-language validation loss to update sampling. Controlled experiments show that equal character budgets create a severe representation artifact for Chinese; byte-level encoding removes much of it, while static Chinese protection can subsequently make English the weakest language. The dynamic controller tracks these transitions and achieves a better score--balance tradeoff than fixed protection. On official evaluation, curriculum-dynamic checkpoints outperform a matched static Chinese-protective baseline while using fewer Chinese epochs; scaling capacity yields the strongest internal ARMOR-DP system. These results support a simple conclusion: multilingual weakness is a training-time state rather than a fixed property of a language. ARMOR-DP therefore separates diagnosis from intervention, allowing protection to follow the current bottleneck instead of a predetermined language label.

Abstract: Large language models acquire broad knowledge from web-scale corpora during pretraining, including knowledge that could be hazardous if misused. Removing such knowledge from the model itself without degrading general capability is desirable but difficult, as it is deeply entangled with general capability within the same parameters. Existing safety alignment and unlearning methods typically adjust these shared parameters, tending to suppress the expression of hazardous knowledge rather than remove it. As a result, the knowledge remains in the model and is readily recoverable by subsequent fine-tuning. To address this limitation, we propose REX (Relocating into Removable EXperts), a knowledge relocation method that separates hazardous knowledge from general capability at the architectural level. Exploiting the modular structure of Mixture-of-Experts models, REX expands the model with a set of new, physically removable experts designated to hold hazardous knowledge. Through expert expansion, selective gradient masking and routing differentiation, it relocates hazardous knowledge out of the original experts and concentrates it in the new ones. Excising these new experts yields a model suitable for public release. This removal holds consistently across Intern-S2, Qwen3.5-35B-A3B and Gemma-4-26B-A4B; on Qwen3.5-35B-A3B, for instance, excision lowers WMDP-Bio accuracy by 48.4 points while MMLU changes by less than 3 points. The excised model matches the strongest of nine unlearning baselines in removing hazardous knowledge, and it is the only one that preserves this removal under subsequent fine-tuning. After 250M tokens of standard fine-tuning on Qwen3.5, the best baseline rebounds on WMDP-Bio from 44.1% to 78.3%, whereas REX rises only from 32.4% to 42.8%.

Abstract: We show that a video can be represented with just a few keyframes without sacrificing a given model's belief about the video. Our method is the first of its kind that extracts these keyframes conditioned on the model's prediction, which is massively helpful for building explainable video models. We then fairly distribute the importance of the video to the model among the keyframes with game theoretically fair Shapley values. Through qualitative and quantitative experiments, we show that keyframe attribution is superior to the traditional frame attribution. Further, we conduct preliminary exploration on highlighting the important changes among the frames by looking at the differences between adjacent keyframes.

Abstract: Long-context training of large language and multimodal models couples structured attention masks (document, prefix-LM, share-question, sliding-window, block-sparse) with multi-node context parallelism, making the mask a dominant bottleneck. FlashMask's column-wise sparse mask compresses them from $O(n^2)$ to $O(n)$ via four query-axis boundary vectors, covering broad structured patterns. Implemented for Ampere GPUs, it neither exploits FlashAttention-3's TMA/WGMMA warp-specialized pipeline on Hopper nor supports distribution. The column-wise sparse mask is indexed by key positions but stores query-axis interval boundaries. Under any multi-chunk context-parallel assignment, it reduces per-rank mask localization to a provably zero-communication \emph{clip--shift--max} on the boundary vectors, compatible with the single-GPU kernel. FlashMask-3 introduces a mask-aware Hopper kernel with three-role warp specialization atop FlashAttention-3's TMA/WGMMA pipeline, plus a mask-aware distributed extension: zero-communication mask localization, a sparsity-aware load balancer (HCC-IPO), and a mask-aware compute-communication overlap with mask-driven communication pruning, whose forward all-gather uses topology-aware hierarchical routing. On a single H100 across 12 masks at 4K--128K context, it improves achieved TFLOPs over FlashMask, FlexAttention, and MagiAttention by 45--141%, 34--71%, and 2--44%. At context-parallel degrees 4--32 with 8K sequence per device, it attains 31--75% compute-communication overlap efficiency and improves throughput over Megatron-Core CP ring, Megatron-Core CP all-gather, and MagiAttention by $1.80\times$--$3.71\times$, $1.71\times$--$2.92\times$, and $1.09\times$--$1.46\times$. In end-to-end training of a 118B-A14B MoE at 128K on 128 H100 GPUs, it raises throughput by 7.3--11.0% over a Dual Chunk all-gather baseline, with attention forward+backward 24.5--44.9% faster.

Abstract: Language models are often evaluated alone but increasingly deployed together, routed between, made to debate, ensembled, or collaboratively trained. For a model inside a system, what matters most is not how well it performs on its own, but how much the system gains from its presence. We call this a model's compositional strength and formalize it using the Shapley value of a cooperative game over coalitions of a model pool, and use a model's leave-one-out contribution to the full coalition as a tractable proxy, requiring n+1 rather than 2^n system evaluations. Evaluating individual vs. compositional strengths across five model pools, twelve collaboration methods, and twelve datasets, we find them largely uncorrelated: in 12 of 15 (pool, method-family) settings, individual strength does not translate into compositional advantages, and the coupling is weakest among general-purpose models that practitioners are most likely to deploy. Standalone benchmarks are therefore often poor proxies for models' value in collaborative systems. We then propose compositional reinforcement learning (C-RL), which trains a policy inside a live multi-model system and rewards the outcome produced by the system as a whole, rather than the policy's correctness in isolation. The leave-one-out baseline in Shapley values cancels under a group policy gradient, so this needs no additional rollout cost beyond RL on the collaborative system's outcome and never requires running a coalition without the policy. Across three collaboration protocols and eight tasks spanning reasoning, coding, and scientific discovery, C-RL produces the strongest collaborator in 21 of 24 settings, lifting collaborative performance by +4.0 to +5.4 pts on average (~3x the gain of baselines), up to +13.1 pts.

Abstract: State space models provide efficient linear-complexity processing for long sequences, but their performance frequently degrades as context length grows. We propose SuPER-Mamba, a training-free framework that uses Surprisal as a model-intrinsic signal for long-range state routing in pretrained Mamba models. SuPER-Mamba computes position-wise Surprisal from next-token predictions, aggregates neighboring responses through Local Pooling, and uses the resulting scores to prioritize Global-Channel updates without query-dependent scoring or additional training. We evaluate SuPER-Mamba on three state-space models across the LongBench-E benchmark. The results show that SuPER-Mamba improves the average score from 8.06 to 15.69, corresponding to a 94.7% relative gain over the Base Models and a 13.1% improvement over the state-of-the-art prior baseline. Ablation and length-wise analyses further support the effectiveness of the proposed routing framework.

Abstract: Disaggregated LLM inference executes the prefill and decode phases on separate resource pools, reducing interference between the two phases. However, this architecture introduces a coupled scheduling problem because a request that completes prefill retains its prompt KV cache in prefill memory until it is admitted to the decode pool. Prefill and decode admission decisions therefore jointly determine memory occupancy and request completion times. To the best of our knowledge, this problem lacks a formal formulation, and no approximation guarantees are known for request scheduling in disaggregated LLM inference. We formalize this setting as the dual-memory request scheduling (DMRS) problem, whose objective is to minimize the sum of request completion times, and prove that it is strongly NP-hard. We develop two deterministic polynomial-time approximation algorithms. The first applies to instances satisfying $α<1$ and $β<1$ and achieves an approximation ratio of $8+\frac{4}{1-α}+\frac{8}{1-β}$, where $α$ and $β$ are the maximum normalized prefill and peak decode memory demands among all requests, respectively. In particular, it yields a $32$-approximation when $α,β\leq 1/2$. The second algorithm applies to all feasible instances and achieves a uniform $36$-approximation. Finally, evaluations on two H100 GPUs using trace-derived offline workloads show that the proposed algorithms reduce the sum of request completion times relative to multiple vLLM deployment baselines.

Abstract: Large-scale data-intensive workloads powered by iterative graph or machine learning processing have an insatiable demand for memory due to the continuous generation of new intermediate data. Although existing caching strategies strive to keep performance-critical data in memory, they often fail in practice because they are limited to node-local memory management and assume high node availability. In this paper, we introduce ShoreKeeper, a framework that maximizes execution efficiency by treating cluster-wide memory as a unified, elastic pool. ShoreKeeper employs a global optimization model that maps distributed dataflows into a single graph abstraction to minimize the recovery cost of data evicted from cache across the cluster. To handle memory working sets and cluster capacities that fluctuate over time, ShoreKeeper employs a graph-pruning technique that permits the fast refinement of the caching plan. These optimizations allow ShoreKeeper to work seamlessly whether deployed on highly available VMs or cost-effective preemptible instances in the cloud. Our experiments show that ShoreKeeper consistently outperforms state-of-the-art systems.

Abstract: Although reinforcement learning (RL) post-training repairs the localized segmental errors of zero-shot text-to-speech (TTS), arriving at a working recipe still relies on tedious manual tuning, and whether LLM agents can take over this research pipeline is unclear. We investigate this question with AgenticTTS-Forge, a collaborative workflow that structures human guidance and agentic execution around a shared workspace, applied to CosyVoice2-0.5B. To measure what the agent automates, we audit its trajectory stage by stage against the published recipe. To measure what it exploits, we score its policies with held-out observers hidden from the agent. Our results show that the agent recovers an underspecified recipe, improves it, and, when gains stall, surveys the literature unprompted and pivots from the LM carrier to the flow carrier, halving Bad cases. However, its autonomy exposes three traps across the data, proxy, and algorithm axes: the held-out set leaks through a channel the contract never reads, a self-shaped reward inflates the proxy where it is scored, and separately tuned policies do not compose additively. These findings show that the binding constraint is measurement rather than reasoning, and can inform the design of harnesses whose contracts read every channel the agent does.

Abstract: On-policy distillation (OPD) combines student-generated rollouts with dense token-level supervision from a teacher, but providing such supervision for every rollout requires substantial teacher computation. We introduce Success-Referenced On-Policy Distillation (SR-OPD), which reduces this cost by selecting which prompts and rollouts receive teacher supervision. When the student produces both successful and failed rollouts for the same prompt, a successful rollout can serve as a natural reference for selecting failed rollouts. SR-OPD therefore focuses on such prompts and prioritizes failed rollouts whose hidden-state trajectories show sustained divergence from a successful reference, while accounting for estimated teacher-input cost. Across three teacher-student pairs and six mathematical reasoning benchmarks, SR-OPD uses only 3.46-5.02% of the teacher-input tokens required by Vanilla OPD in the one-pass setting while maintaining comparable reasoning performance. Under a controlled setting matched to 5% of Vanilla OPD's teacher-input budget, further experiments support both key design choices: focusing supervision on prompts with both successful and failed rollouts, and using successful rollouts to guide failure selection. These results indicate that a student's own successful behavior can serve as a useful reference for allocating teacher supervision under a fixed teacher-input budget.

Abstract: On-policy distillation (OPD) trains a student model on its self-generated trajectories with dense token-level teacher feedback. However, naive OPD may suffer from teacher unanchoring, where the student's reasoning trajectory drifts far from the teacher, causing the teacher to be queried on states it would hardly visit and thus provide unreliable supervision. We propose Interactive-Policy Distillation (IPD), which applies adaptive teacher intervention to the student rollout. Under a bidirectional propose-and-verify state machine, the student and teacher alternately exchange their roles as proposer and verifier, and collaboratively generate mixed-source trajectories. Then different supervisions are applied according to the source of each token. This bidirectional propose-and-verify mechanism and the source-split loss make IPD not only a more performant distillation method, but also a unified bridge between on-policy and off-policy paradigms. To make the interleaved dual-model rollouts more efficient, we also design a dedicated fused inference engine that co-hosts both models in one serving instance with separate KV caches and instantiates the state machine model to distribute, collect, and process requests. On math reasoning tasks and across multiple teacher-student model pairs, student models trained with IPD not only outperform those trained with OPD, but also demonstrate higher data efficiency. Specifically, when distilling Qwen3-30B-A3B into Qwen3-1.7B-Base, IPD brings a +3.28 mean@8 and a +3.28 best@8 benchmark-averaged accuracy improvement compared with OPD. Besides, IPD only consumes about 1/4 of the training examples and steps to outperform OPD trained on the whole training dataset for one epoch. We also investigate the impact of different loss variants and takeover / handback configurations, and demonstrate the robustness of IPD on different training data.

Abstract: Agentic reinforcement learning (RL) differs from conventional RL in that every task executes inside a complex environment, e.g., a virtual machine or a container. We find that agentic RL pays a heavy environment tax: a large share of the iteration time goes to the environment rather than to learning. The root cause is the lack of a full-lifecycle solution to environment management. We present WeEnv, which manages environments across packaging, initialization, and provisioning. WeEnv packages components as independently published layer groups and composes them at initialization, so that updating a component republishes one small group rather than every artifact containing it. To speed up environment initialization, WeEnv launches environments instantly and fetches contents on demand. During task execution, WeEnv provisions CPU and memory elastically, adjusting each environment's quota from its observed usage to fit the varying demands. WeEnv reduces the initialization by 5.6-14.2x over E2B, Docker, and AgentENV, cutting its share of the iteration time from up to 53.4% to 9.1%. WeEnv is deployed for agentic RL at WeChat.

Abstract: Large language models are increasingly trained on GPUs distributed across multiple regions, but geo-distributed training is challenging in practice. Real clusters often contain GPUs with different speeds and memory capacities, and they communicate over slow wide-area networks. Our analysis shows that this creates serious problems: existing synchronous methods preserve stable updates, but fast GPUs wait up to 20.9% of their runtime for slower ones, and all workers spend, on average, 65.8% of their runtime on synchronization. Recent asynchronous methods reduce waiting time but worsen the model accuracy due to stale updates. To address the problems, we present GeoMesh, a synchronous geo-distributed training framework for heterogeneous GPUs. GeoMesh balances per-worker workloads by assigning each GPU a suitable batch size and number of inner steps, so faster GPUs do more useful work instead of waiting. It also reduces communication volume by nearly 32x by exchanging compressed sign-based pseudo-gradients with lightweight magnitude and token count. Across heterogeneous GPUs and Azure-derived WAN, GeoMesh reduces time-to-target perplexity by up to 70.2% over representative baselines and lowers straggler- and WAN-induced GPU idle by up to 8.0x and 5.6x, respectively, while preserving comparable zero-shot accuracy.

Abstract: Knowledge distillation offers an efficient route to transfer a task-adapted vision-language teacher to a compact student. The training target in current vision-language distillation methods is typically constructed from the teacher prediction and applied uniformly to all training samples, making it unreliable under class and domain shifts. In this paper, we argue that distillation target construction should be treated as a dynamic training decision rather than a fixed recipe. To this end, we propose OnPoKD, an on-policy distillation framework for vision-language model adaptation. To the best of our knowledge, OnPoKD is the first framework that applies on-policy distillation to vision-language model adaptation by learning target construction as a policy decision. OnPoKD learns a lightweight controller that constructs sample-wise adaptive targets using reliability and disagreement cues from the teacher model, student model, and zero-shot prior. Instead of relying on a fixed teacher prediction, the controller dynamically balances teacher supervision, zero-shot prior guidance, and hard-label anchoring through bounded policy actions, allowing the distillation target to adapt to varying sample reliability and training stages. The policy controller is updated with validation feedback, encouraging target construction to optimize transferability rather than merely fitting the training distribution. Since the controller is only used during training, OnPoKD can be seamlessly integrated into existing vision-language distillation pipelines while preserving the original inference architecture and test-time cost. Extensive experiments on Base-to-novel generalization and Cross-dataset transfer benchmarks show that OnPoKD consistently improves over strong vision-language distillation baselines.