Bernie Wang

Yongjun He, Shuai Zhang, Jiading Gai et al. · amazon-science

As large language models (LLMs) continue to scale and new GPUs are released even more frequently, there is an increasing demand for LLM post-training in heterogeneous environments to fully leverage underutilized mid-range or previous-generation GPUs and alleviate the shortage of homogeneous high-end GPUs within a single availability zone. However, achieving high-performance reinforcement learning (RL) training for LLMs on such computing resources remains challenging because the workflow involves multiple models and tasks with complex computation and data dependencies. In this paper, we present HetRL, a distributed system for efficient RL training in infrastructures with heterogeneous GPUs and networks. HetRL formulates the scheduling of RL training in heterogeneous environments as a constrained joint optimization problem and provides two complementary approaches for addressing this problem: (1) a hybrid scheduling algorithm that efficiently identifies near-optimal solutions, and (2) an integer linear programming (ILP)-based scheduling algorithm that obtains optimal solutions, enabling flexible trade-offs between solution optimality and efficiency. Our extensive evaluation, consuming 20,000 GPU-hours, shows that HetRL achieves up to 9.17x the throughput of state-of-the-art systems, and 3.17x on average, across a wide range of workloads and settings.

Zelin He, Haotian Lin, Boran Han et al.

Agentic reinforcement learning (RL) enables LLM agents to improve continuously from environment rewards, yet the resulting policies do not systematically accumulate reusable strategies that generalize across tasks. Modular skills can provide such reusable strategies, yet existing skill-augmented RL methods decouple skill creation from policy optimization, risking adopting skills that conflict with the evolving policy. Inspired by Anthropic's Skill Creator, we introduce ReSkill, an RL-in-the-loop skill creation framework that reconciles skill evolution with policy learning. ReSkill exploits the group-wise structure of GRPO to naturally embed three mechanisms with only marginal additional overhead: (1) an assertion-driven skill creator that diagnoses failures from past experience and proposes conditional, trigger-based skill revisions; (2) within-group rollout sampling that enables controlled comparison of skill versions, capturing which version best supports the policy's ongoing learning; and (3) Thompson Sampling with adaptive discounting to balance exploration and exploitation in skill version selection as the policy evolves. Across several domains, ReSkill consistently outperforms existing memory and skill-based RL methods, with the largest gains on unseen tasks. Analysis of the skill lifecycle shows skills being automatically created, tested, refined, and pruned as the policy improves, demonstrating reconciled skill-policy co-evolution.

Zelin He, Boran Han, Xiyuan Zhang et al.

Time-series diagnostic reasoning is essential for many applications, yet existing solutions face a persistent gap: general reasoning large language models (GRLMs) possess strong reasoning skills but lack the domain-specific knowledge to understand complex time-series patterns. Conversely, fine-tuned time-series LLMs (TSLMs) understand these patterns but lack the capacity to generalize reasoning for more complicated questions. To bridge this gap, we propose a hybrid knowledge-injection framework that injects TSLM-generated insights directly into GRLM's reasoning trace, thereby achieving strong time-series reasoning with in-domain knowledge. As collecting data for knowledge injection fine-tuning is costly, we further leverage a reinforcement learning-based approach with verifiable rewards (RLVR) to elicit knowledge-rich traces without human supervision, then transfer such an in-domain thinking trace into GRLM for efficient knowledge injection. We further release SenTSR-Bench, a multivariate time-series-based diagnostic reasoning benchmark collected from real-world industrial operations. Across SenTSR-Bench and other public datasets, our method consistently surpasses TSLMs by 9.1%-26.1% and GRLMs by 7.9%-22.4%, delivering robust, context-aware time-series diagnostic insights.

Haoyang Fang, Shuai Zhang, Yifei Ma et al.

Domain-specific finetuning is essential for dense retrievers, yet not all training pairs contribute equally to the learning process. We introduce OPERA, a data pruning framework that exploits this heterogeneity to improve both the effectiveness and efficiency of retrieval model adaptation. We first investigate static pruning (SP), which retains only high-similarity query-document pairs, revealing an intrinsic quality-coverage tradeoff: ranking (NDCG) improves while retrieval (Recall) can degrade due to reduced query diversity. To resolve this tradeoff, we propose a two-stage dynamic pruning (DP) strategy that adaptively modulates sampling probabilities at both query and document levels throughout training, prioritizing high-quality examples while maintaining access to the full training set. Evaluations across eight datasets spanning six domains demonstrate the effectiveness of both approaches: SP improves ranking over standard finetuning (NDCG@10 +0.5\%), while DP achieves the strongest performance on both ranking (NDCG@10 +1.9\%) and retrieval (Recall@20 +0.7\%), with an average rank of 1.38 across all methods. These findings scale to Qwen3-Embedding, an LLM-based dense retriever, confirming architecture-agnostic benefits. Notably, DP reaches comparable performance in less than 50\% of the training time required by standard finetuning.

Jiading Gai, Shuai Zhang, Xiang Song et al.

Modern RL post-training methods such as GRPO and DAPO train on $N$ response sequences of $R$ tokens sampled from a shared prompt of $P$ tokens, but standard FlashAttention replicates all $P$ prompt tokens $N$ times across both forward and backward passes -- duplicating compute and memory on identical hidden states. In large-rollout, long-context RL training ($N{\geq}16$, $P{\geq}8\text{K}$), this redundancy dominates the policy update cost. We observe that in decoder-only models, causal masking makes prompt representations invariant across sequences at every layer, so all per-token operations (norms, projections, MLP) and attention can process the prompt once -- a property not yet exploited at the kernel level for training. We propose \textbf{DualKV}, the first FlashAttention kernel variant that eliminates shared-prompt replication during RL training, via (1)~fused CUDA forward and backward kernels that iterate over two disjoint KV regions -- shared context and per-sequence response -- in a single kernel launch, and (2)~a data-pipeline redesign in veRL that repacks $N(P{+}R)$ tokens into $P{+}NR$ tokens per micro-batch, extending the token reduction from attention to the entire model by a factor $ρ= N(P{+}R)/(P{+}NR)$. DualKV is mathematically equivalent to standard attention and introduces no approximation. On Qwen3-8B GRPO training with 8$\times$H100 GPUs ($N{=}32$, 8K-context), DualKV achieves $1.63$--$2.09\times$ policy-update speedup, enables $2\times$ larger micro-batches, and raises MFU from $36\%$ to $76\%$. Similar gains hold for DAPO ($2.47\times$ speedup, $77\%$ MFU). At 30B MoE scale on 16$\times$H100, DualKV achieves $3.82\times$ policy-update and $3.38\times$ end-to-end step speedup over FlashAttention (which requires 4-way Ulysses sequence parallelism to avoid OOM).

Bernie Wang, Virginia Wu, Bichen Wu et al.

LiDAR (Light Detection And Ranging) is an essential and widely adopted sensor for autonomous vehicles, particularly for those vehicles operating at higher levels (L4-L5) of autonomy. Recent work has demonstrated the promise of deep-learning approaches for LiDAR-based detection. However, deep-learning algorithms are extremely data hungry, requiring large amounts of labeled point-cloud data for training and evaluation. Annotating LiDAR point cloud data is challenging due to the following issues: 1) A LiDAR point cloud is usually sparse and has low resolution, making it difficult for human annotators to recognize objects. 2) Compared to annotation on 2D images, the operation of drawing 3D bounding boxes or even point-wise labels on LiDAR point clouds is more complex and time-consuming. 3) LiDAR data are usually collected in sequences, so consecutive frames are highly correlated, leading to repeated annotations. To tackle these challenges, we propose LATTE, an open-sourced annotation tool for LiDAR point clouds. LATTE features the following innovations: 1) Sensor fusion: We utilize image-based detection algorithms to automatically pre-label a calibrated image, and transfer the labels to the point cloud. 2) One-click annotation: Instead of drawing 3D bounding boxes or point-wise labels, we simplify the annotation to just one click on the target object, and automatically generate the bounding box for the target. 3) Tracking: we integrate tracking into sequence annotation such that we can transfer labels from one frame to subsequent ones and therefore significantly reduce repeated labeling. Experiments show the proposed features accelerate the annotation speed by 6.2x and significantly improve label quality with 23.6% and 2.2% higher instance-level precision and recall, and 2.0% higher bounding box IoU. LATTE is open-sourced at https://github.com/bernwang/latte.

Wenqi Jiang, Shuai Zhang, Boran Han et al. · amazon-science

Retrieval-augmented generation (RAG) can enhance the generation quality of large language models (LLMs) by incorporating external token databases. However, retrievals from large databases can constitute a substantial portion of the overall generation time, particularly when retrievals are periodically performed to align the retrieved content with the latest states of generation. In this paper, we introduce PipeRAG, a novel algorithm-system co-design approach to reduce generation latency and enhance generation quality. PipeRAG integrates (1) pipeline parallelism to enable concurrent retrieval and generation processes, (2) flexible retrieval intervals to maximize the efficiency of pipeline parallelism, and (3) a performance model to automatically balance retrieval quality and latency based on the generation states and underlying hardware. Our evaluation shows that, by combining the three aforementioned methods, PipeRAG achieves up to 2.6$\times$ speedup in end-to-end generation latency while improving generation quality. These promising results showcase the effectiveness of co-designing algorithms with underlying systems, paving the way for the adoption of PipeRAG in future RAG systems.

Andreas Auer, Raghul Parthipan, Pedro Mercado et al.

Pretrained time series models, capable of zero-shot forecasting, have demonstrated significant potential in enhancing both the performance and accessibility of time series forecasting. However, existing pretrained models either do not support covariates or fail to incorporate them effectively. We introduce COSMIC, a zero-shot forecasting model that utilizes covariates via in-context learning. To address the challenge of data scarcity, we propose Informative Covariate Augmentation, which enables the training of COSMIC without requiring any datasets that include covariates. COSMIC achieves state-of-the-art performance in zero-shot forecasting, both with and without covariates. Our quantitative and qualitative analysis demonstrates that COSMIC effectively leverages covariates in zero-shot forecasting.

Haoyang Fang, Boran Han, Nick Erickson et al.

Existing AutoML systems have advanced the automation of machine learning (ML); however, they still require substantial manual configuration and expert input, particularly when handling multimodal data. We introduce MLZero, a novel multi-agent framework powered by Large Language Models (LLMs) that enables end-to-end ML automation across diverse data modalities with minimal human intervention. A cognitive perception module is first employed, transforming raw multimodal inputs into perceptual context that effectively guides the subsequent workflow. To address key limitations of LLMs, such as hallucinated code generation and outdated API knowledge, we enhance the iterative code generation process with semantic and episodic memory. MLZero demonstrates superior performance on MLE-Bench Lite, outperforming all competitors in both success rate and solution quality, securing six gold medals. Additionally, when evaluated on our Multimodal AutoML Agent Benchmark, which includes 25 more challenging tasks spanning diverse data modalities, MLZero outperforms the competing methods by a large margin with a success rate of 0.92 (+263.6\%) and an average rank of 2.28. Our approach maintains its robust effectiveness even with a compact 8B LLM, outperforming full-size systems from existing solutions.

Xiyuan Zhang, Danielle C. Maddix, Junming Yin et al. · amazon-science

Since the seminal work of TabPFN, research on tabular foundation models (TFMs) based on in-context learning (ICL) has challenged long-standing paradigms in machine learning. Without seeing any real-world data, models pretrained on purely synthetic datasets generalize remarkably well across diverse datasets, often using only a moderate number of in-context examples. This shifts the focus in tabular machine learning from model architecture design to the design of synthetic datasets, or, more precisely, to the prior distributions that generate them. Yet the guiding principles for prior design remain poorly understood. This work marks the first attempt to address the gap. We systematically investigate and identify key properties of synthetic priors that allow pretrained TFMs to generalize well. Based on these insights, we introduce Mitra, a TFM trained on a curated mixture of synthetic priors selected for their diversity, distinctiveness, and performance on real-world tabular data. Mitra consistently outperforms state-of-the-art TFMs, such as TabPFNv2 and TabICL, across both classification and regression benchmarks, with better sample efficiency.

Xiyuan Zhang, Boran Han, Haoyang Fang et al. · amazon-science

Recently, there has been growing interest in incorporating textual information into foundation models for time series forecasting. However, it remains unclear whether and under what conditions such multimodal integration consistently yields gains. We systematically investigate these questions across a diverse benchmark of 16 forecasting tasks spanning 7 domains, including health, environment, and economics. We evaluate two popular multimodal forecasting paradigms: aligning-based methods, which align time series and text representations; and prompting-based methods, which directly prompt large language models for forecasting. Our findings reveal that the benefits of multimodality are highly condition-dependent. While we confirm reported gains in some settings, these improvements are not universal across datasets or models. To move beyond empirical observations, we disentangle the effects of model architectural properties and data characteristics, drawing data-agnostic insights that generalize across domains. Our findings highlight that on the modeling side, incorporating text information is most helpful given (1) high-capacity text models, (2) comparatively weaker time series models, and (3) appropriate aligning strategies. On the data side, performance gains are more likely when (4) sufficient training data is available and (5) the text offers complementary predictive signal beyond what is already captured from the time series alone. Our study offers a rigorous, quantitative foundation for understanding when multimodality can be expected to aid forecasting tasks, and reveals that its benefits are neither universal nor always aligned with intuition.

Yixin Chen, Shuai Zhang, Boran Han et al. · amazon-science

High-resolution (HR) images are pivotal for enhancing the recognition and understanding capabilities of multimodal large language models (MLLMs). However, directly increasing image resolution can significantly escalate computational demands. In this study, we propose a method called Chain of Region-of-Interest (CoRoI) for Visual Instruction Tuning, aimed at alleviating the computational burden associated with high-resolution images for MLLMs. Drawing inspiration from the selective nature of the human visual system, we recognize that not all regions within high-resolution images carry equal importance. CoRoI seeks to identify and prioritize the most informative regions, thereby enhancing multimodal visual comprehension and recognition while circumventing the need for processing lengthy HR image tokens. Through extensive experiments on 11 benchmarks, we validate the efficacy of CoRoI across varying sizes, ranging from 7B to 34B in parameters. Our models consistently demonstrate superior performance across diverse multimodal benchmarks and tasks. Notably, our method outperforms LLaVA-NeXT on almost all benchmarks and our finetuned 34B model surpasses proprietary methods like Gemini Pro 1.0 on six benchmarks, as well as outperforming GPT-4V on MMB, SEED-I, and MME.

Mert Kayaalp, Caner Turkmen, Oleksandr Shchur et al.

Is bigger always better for time series foundation models? With the question in mind, we explore an alternative to training a single, large monolithic model: building a portfolio of smaller, pretrained forecasting models. By applying ensembling or model selection over these portfolios, we achieve competitive performance on large-scale benchmarks using much fewer parameters. We explore strategies for designing such portfolios and find that collections of specialist models consistently outperform portfolios of independently trained generalists. Remarkably, we demonstrate that post-training a base model is a compute-effective approach for creating sufficiently diverse specialists, and provide evidences that ensembling and model selection are more compute-efficient than test-time fine-tuning.

Bernie Wang, Simon Xu, Kurt Keutzer et al.

Meta-reinforcement learning typically requires orders of magnitude more samples than single task reinforcement learning methods. This is because meta-training needs to deal with more diverse distributions and train extra components such as context encoders. To address this, we propose a novel self-supervised learning task, which we named Trajectory Contrastive Learning (TCL), to improve meta-training. TCL adopts contrastive learning and trains a context encoder to predict whether two transition windows are sampled from the same trajectory. TCL leverages the natural hierarchical structure of context-based meta-RL and makes minimal assumptions, allowing it to be generally applicable to context-based meta-RL algorithms. It accelerates the training of context encoders and improves meta-training overall. Experiments show that TCL performs better or comparably than a strong meta-RL baseline in most of the environments on both meta-RL MuJoCo (5 of 6) and Meta-World benchmarks (44 out of 50).