Pathway 如何在 Amazon SageMaker HyperPod 上开发受大脑启发的 BDH 架构

随着 AI 系统承担越来越复杂的任务，行业的大部分进展来自扩大模型规模、训练数据、上下文长度和推理时计算量。不同于将推理工作外化为思维链（即按顺序生成额外 token，再将其反馈到后续步骤），Pathway 受大脑启发的 BDH（Dragon Hatchling）在潜在空间中执行推理。它从示例中学习并不断优化解决方案，而不生成中间文本轨迹。BDH 超越了 Transformer 范式，提供了一种受大脑启发的架构：最初被定义为由神经元构成的图，神经元通过稀疏、局部的交互进行通信，并在类似突触的连接中维持状态。模型状态可在上下文中自适应调整，无需在测试时更新权重；其推理范围既不受固定大小上下文窗口的限制，也不依赖大量低效的思维链 token。

大语言模型（LLM）改变了 AI，重塑了我们处理从代码生成到创意写作等任务的方式。然而，关于通用智能及其能否在长时间跨度内进行连贯推理的根本问题仍然存在。其架构在过去约 10 年间基本保持静态，在训练和推理两方面仍存在显著低效。LLM 往往会在长时间交互中遗忘，无法在不同会话之间保留知识，并且需要重新训练才能获得新知识。Pathway 的 BDH-CQ 会在推理期间通过在循环潜在状态内执行迭代计算来更新模型内部记忆，同时只解码其候选答案。模型可以解决新问题，而无需生成冗长、语言化的推理轨迹，也不需要微调或重新训练。

Pathway 的 BDH 可与 PyTorch 等知名框架集成，并使用 Amazon SageMaker HyperPod 扩展其训练规模。Amazon SageMaker HyperPod 帮助其应用 AI 科学家以具备韧性、可扩展且具成本效益的方式共享计算资源。

Transformer：架构与根本性局限

尽管 Transformer 架构已被广泛用于自然语言处理（NLP），但它在训练和推理工作负载中都面临显著限制。

在训练期间，Transformer 难以实现超越训练数据的系统性泛化，尤其是在长思维链推理任务中；为弥补这一限制，需要投入海量数据和计算资源。Transformer 架构的密集计算模式和完整反向传播要求，导致计算成本随着模型规模增长而急剧上升。微观尺度的动态（即神经元激活）与宏观行为（即模型为何给出某个答案）之间缺乏明确联系。它也难以在不进行完整重新训练或微调的情况下高效更新知识，因为 Transformer 常常遭遇灾难性遗忘：学习新事物会导致遗忘其他内容。

在推理阶段，Transformer 面临其他低效问题：其注意力机制的结构限制了可扩展性；即使采用 Mixture of Experts 技术，密集激活模式仍会造成过多的计算和内存带宽消耗。固定上下文窗口和不断增长的 KV-cache 对序列长度处理施加了限制。此外，Transformer 状态的封闭性使得解释或监控模型推理过程几乎不可能，这引发了人们对生产环境和高度监管行业中可靠性与安全性的担忧。

这些挑战从根本上与该架构的设计选择有关，而不仅仅是实现细节，因此有必要探索更符合计算效率和自然智能原则的替代方法。

“如今的 AI 为推理付出了高昂的 token 成本，但这一成本是由架构施加的，而非任何智能定律所决定。目前，每一步推理都会消耗上下文、增加延迟并耗费计算资源。我们证明，不同的架构能够改变游戏规则，并开启一个全新的空间：每一美元能够获得多少智能。一个基于 Pathway BDH 架构构建的 1.5 亿参数模型，在潜在空间中进行循环推理，并在 ARC-AGI-1 上创下成本效率的新纪录。瓶颈从来不是智能，而是设计。”

— Zuzanna Stamirowska，Pathway CEO 兼联合创始人

BDH：统一人工智能与自然智能的架构

Pathway 的愿景是从根本上改变模型的思考方式。其 BDH 架构是一种可持续学习、演化和推理的后 Transformer 模型。BDH 架构提供了一种构建语言模型的新方式，将序列建模重新定义为相互作用的神经元粒子网络上的局部图动态。该模型采用赫布学习（Hebbian learning），即“共同激发的神经元会形成连接”的原则，来实现注意力机制。这种方法在单一神经元维度（n）上扩展，降低了部署 Transformer 时跨计算资源分布的复杂性。

与大脑一样，BDH 的交互被定义为稀疏且局部的，神经元之间的连接编码了记忆和推理功能。这种稀疏激活特征（通常在任意时刻只有 5% 的神经元处于活跃状态）支持高效计算。更小的活跃状态意味着每个推理步骤所需的计算更少，从而提升生产环境中的性能。

BDH 通过一种线性机制实现注意力，该机制在固定的高维状态上运行，不会产生 Transformer 模型在长上下文场景下特有的复杂度增长。这使模型能够更高效地处理更长序列，而不受 Transformer 架构上下文长度限制。模型状态直接映射到神经元对之间的突触连接，从而为推理过程提供可见性，使模型的决策更易解释。

最近，Pathway 构建了基于 BDH 的推理系统 BDH-CQ。它通过上下文学习和潜在空间迭代推理，扩展了 BDH 的视觉问题求解能力。

如下文所示，BDH-CQ 擅长上下文学习。它对潜在状态进行循环计算，从而支持高效的并行假设探索；其中，神经元群体可以表示当前问题的不同候选解。该模型实现了一种推理效率，能够以固定内存成本处理任意数量的示例。

“客户正越来越多地探索如何将高级推理从实验阶段推向生产环境，而在生产环境中，性能、效率和可扩展性都至关重要。Pathway 在 Amazon SageMaker HyperPod 上训练 BDH-CQ 的工作，展示了一条颇具前景的路径：以更高的成本效益大规模部署高性能系统。”

— Nicolas Tarducci，Amazon Web Services EMEA 初创企业解决方案架构负责人

在 Amazon SageMaker HyperPod 上开发 BDH 架构

Pathway 使用 Amazon SageMaker HyperPod 开发其 BDH 架构。Amazon SageMaker HyperPod 是专为训练需要跨数百或数千块 GPU 进行分布式训练或推理的 LLM 和基础模型（FM）而构建的基础设施解决方案。它提供全托管、高性能的机器学习（ML）训练环境，具备自动化集群配置、优化的网络架构和可定制的软件栈。对于模型生产者，Amazon SageMaker HyperPod 带来三项核心优势：通过消除复杂的基础设施搭建与管理缩短上市时间；通过自动扩缩容提高成本效率；通过专用网络架构实现 GPU 集群近乎线性的扩展能力，从而提升模型性能。这有助于 ML 团队专注于模型开发而非基础设施难题，并相较传统基础设施方法实现更快训练速度和更低成本。

对于 Pathway 而言，要开发超越 Transformer 的 AI 架构，对其训练基础设施进行全面可观测至关重要。Amazon SageMaker HyperPod 与 Amazon Managed Service for Prometheus 等高级监控工具的集成，提供了这类工作负载所需的深度洞察。通过 Amazon Managed Grafana 控制面板提供的可视化、指标收集和主动优化能力相结合，Pathway 可以可视化复杂的分布式训练模式，监控 GPU 利用率和内存模式，并跟踪其并行训练方法的节点间通信效率。

这套全面的可观测性栈有助于缩短迭代新架构方法时的开发周期，优化资源密集型训练的成本性能比，并确保不同训练运行之间结果的可靠性和可复现性。有关此集成的更多信息，以及快速部署此可观测性栈的方法，请参阅 awsome-distributed-ai repository 或 Amazon SageMaker HyperPod 文档。

Amazon Elastic Fabric Adapter（EFA）可原生集成 NVIDIA CUDA 平台，以支持 GPU 加速计算，并可集成 NVIDIA Collective Communications Library（NCCL）以实现 GPU 间通信。它能够以可靠且可扩展的方式分发数据、权重、激活值等内容。Pathway 使用了 Amazon Elastic Compute Cloud（Amazon EC2）p5en.48xlarge 实例；每个实例可提供最高 3200 Gbps 的网络性能，并配备由 NVIDIA H200 GPU 驱动的 NVIDIA 加速 AI 基础设施。所有实例均通过 EFA 互联，并运行于 Amazon EC2 UltraCluster 上，以缩短 GPU 之间的网络距离并降低延迟。

BDH 架构的验证：BDH-CQ 在 ARC-AGI 上实现 29.2% pass@2

BDH-CQ 在 ARC-AGI 基准上实现了 29.2% 的 pass@2，单任务成本为 0.0007 美元。

得益于 Pathway 的新型推理方法、上下文内任务获取能力和迭代式潜在计算，截至 2026 年 8 月，BDH-CQ 改写了 ARC-AGI-1 基准上的成本—准确率帕累托前沿。

ARC-AGI-1 会向 AI 系统提供少量“前后对比”示例，以展示一条未知的视觉规则。系统随后必须推断该规则，并将其应用到一个新的网格上；这种能力通常与类人智能相关联。BDH-CQ 在循环潜在状态中执行迭代计算，并且只解码其候选答案。与将工作外化为思维链、因而增加延迟和推理成本的 Transformer 相比，BDH-CQ 会在处理示例的过程中更新模型内部记忆。它能够解决新问题，而无需生成冗长、语言化的推理轨迹。

ARC-AGI 所探索的模式与真实世界挑战相似：系统必须在信息和约束不断变化时进行可靠推理，例如调查网络安全事件、协调交通运输网络、响应实时工业运营，以及在长时间运行的工作流中操作自主 Agent。BDH-CQ 的新方法表明，它能够让这些应用成本更低、响应更快且更可靠。

结论

BDH 是一种受无标度生物网络启发的新型大语言模型架构。它借鉴了生物学已经验证有效的原则，即局部交互、稀疏活动、持久状态和持续调整，并将这些原则应用于现代序列模型。它提供了对 GPU 友好的实现。BDH 是一种围绕循环记忆和局部计算构建的后 Transformer 架构。BDH-CQ 则将其扩展为一种推理系统：它能够从上下文中学习，并在连续潜在工作区中进行迭代计算，从而支持超越逐 token 生成的高效推理。

如需开始探索和实现 BDH 架构，可以访问详细的技术文档，并通过 Pathway 的代码库获取示例实现。该实现支持标准 PyTorch 工作流，并且可以方便地集成到现有 ML 管道中。

推动 AI 模型开发边界的团队，可以利用 Amazon SageMaker HyperPod 在大规模场景中保持效率和可靠性。Amazon SageMaker HyperPod 训练为大型模型训练中的常见挑战提供了强大的解决方案。要了解更多有关 Amazon SageMaker HyperPod 的信息，请访问 AI on SageMaker HyperPod，获取研讨会、代码示例和故障排除指南。或者阅读 Amazon SageMaker HyperPod 上的无检查点训练，了解其韧性功能。

关于作者

Paulo Aragão

Paulo 是一名首席全球专家解决方案架构师，专注于帮助客户在 AWS 上构建其前沿 AI 战略。他在高性能计算和 AIML 项目领域拥有超过 20 年经验，热衷于从客户挑战出发，帮助客户克服困难。要了解更多有关其项目的信息，请访问他的 GitHub 页面。

Rodrigo Merino

As AI systems take on more complex tasks, much of the industry’s progress has come from increasing model scale, training data, context length, and inference-time computation. Instead of externalizing reasoning work as a chain-of-thought (generating extra tokens sequentially and feeding them back into later steps), Pathway’s brain-inspired BDH (Dragon Hatchling) performs reasoning in latent space. It learns from examples and refines a solution without generating an intermediate text trace. BDH moves beyond the transformer paradigm by offering a brain-inspired architecture, originally formulated as a graph of neurons that communicate through sparse, local interactions and maintain state in synapse-like connections. The model states adapt in context without test-time weight updates, and the reasoning horizon isn’t limited by a fixed-size context window or tied to a flood of inefficient chain-of-thought tokens.

Large language models (LLMs) have transformed AI, changing how we approach tasks from code generation to creative writing. However, fundamental questions remain about general intelligence and their ability to reason over long time periods in a coherent way. Their architecture, kept roughly static for the past 10 years, still presents important inefficiencies both in training and inference. LLMs tend to forget during long interactions, they don’t keep knowledge between sessions, and they need to be retrained to acquire new knowledge. Pathway’s BDH-CQ updates the model’s internal memory during inference by performing iterative computation inside a recurrent latent state and decoding only its candidate answers. Models can work through a new problem without generating long, verbalized reasoning traces, and without requiring fine-tuning or retraining.

Pathway’s BDH integrates with well-known frameworks such as PyTorch and uses Amazon SageMaker HyperPod to scale out their training. Amazon SageMaker HyperPod helps their applied AI scientists share compute resources in a resilient, scalable, and cost-effective way.

Transformers: architecture and fundamental limitations

The transformer architecture, while widely adopted for natural language processing (NLP), faces significant limitations in both training and inference workloads.

During training, transformers struggle with systematic generalization beyond their training data, particularly for long chain-of-thought reasoning tasks, and require massive amounts of data and computational effort to compensate for this limitation. The transformer architecture’s dense computation patterns and full back-propagation requirements lead to substantial computational costs that scale exponentially with model size. There’s a lack of clear connections between the dynamics at the micro-scale (that is, neuron activations) and macro-behavior (that is, why the model answers what it does). It’s difficult to efficiently update its knowledge without full retraining or fine-tuning, because transformers often struggle with catastrophic forgetting, where learning something new leads to forgetting something else.

At inference, transformers face other inefficiencies: the structure of their attention mechanism limits scalability, while their dense activation patterns, even with Mixture of Experts techniques, result in excessive computation and memory bandwidth usage. The fixed context window and growing KV-cache enforce limitations on sequence length processing. Furthermore, the closed-source nature of the transformer state makes it nearly impossible to interpret or monitor the model’s reasoning process, raising concerns about reliability and safety in production environments and heavily regulated industries.

These challenges are fundamentally tied to the architecture’s design choices rather than only implementation details, suggesting the need for alternative approaches that better align with both computational efficiency and natural intelligence principles.

“Today’s AI pays a steep token cost for reasoning, but that cost is imposed by architecture, not by any law of intelligence. Currently, every reasoning step consumes context, adds latency, and burns compute. We show that a different architecture changes the game and opens up a whole new space in terms of how much intelligence per dollar. A 150M-parameter model, built on Pathway’s BDH architecture, reasons recurrently in latent space, and sets a new state of the art in cost efficiency on ARC-AGI-1. The bottleneck was never intelligence. It was design.”

— Zuzanna Stamirowska, CEO and co-founder, Pathway

BDH: A unified architecture for artificial and natural intelligence

Pathway’s vision is to fundamentally change the way models think. Their BDH architecture is a post-transformer model that continually learns, evolves, and reasons. BDH’s architecture represents a new way to build language models that reformulates sequence modeling as local graph dynamics on a network of interacting neuron particles. The model employs Hebbian learning, the principle that “neurons that fire together, wire together”, to implement attention mechanisms. This approach scales in a single neuron dimension (n), which reduces the complexity of distributing across compute resources that we face when deploying transformers.

As in the brain, the interactions of BDH are defined to be sparse and local, and the connections between the neurons encode the memory and reasoning functions. This sparse activation profile (only 5 percent of neurons are typically active at a given time) supports efficient computation. The reduced active state means less computation is required per inference step, leading to performance improvements in production environments.

BDH implements attention through a linear mechanism that operates on fixed, high-dimensional states without incurring the increased complexity characteristic of transformer models for long contexts. This allows the model to process longer sequences more efficiently, without the context length limitations of transformer architectures. The model’s states are directly mapped to synaptic connections between neuron pairs, providing visibility into the reasoning process and making the model’s decision-making easier to interpret.

Recently, Pathway built BDH-CQ, a reasoning system built on top of BDH. It extends BDH with in-context learning and latent iterative reasoning for visual problem-solving.

As the following sections show, BDH-CQ excels at in-context learning. Its recurrent computations over latent states support efficient parallel hypothesis exploration, where communities of neurons can represent different candidate solutions for a problem at hand. The model achieves a reasoning efficiency which can process arbitrary numbers of demonstrations at fixed memory cost.

“Customers are increasingly exploring how to move advanced reasoning from experimentation into production, where performance, efficiency, and scalability all matter. Pathway’s work training BDH-CQ on Amazon SageMaker HyperPod points to a promising path toward deploying high-performing systems more cost-effectively at scale.”

— Nicolas Tarducci, Head of Solution Architecture for Startups EMEA, Amazon Web Services

Developing BDH architecture on Amazon SageMaker HyperPod

Pathway uses Amazon SageMaker HyperPod for developing its BDH architecture. Amazon SageMaker HyperPod is a purpose-built infrastructure solution for training LLMs and foundation models (FMs) that require distributed training or inference across hundreds or thousands of GPUs. It provides a fully managed, high-performance machine learning (ML) training environment with automated cluster provisioning, optimized networking fabric, and customizable software stacks. For model producers, Amazon SageMaker HyperPod delivers three key benefits: reduced time-to-market by removing complex infrastructure setup and management, improved cost efficiency through automatic scaling, and enhanced model performance through specialized networking architecture that achieves near-linear scaling across GPU clusters. This helps ML teams focus on model development rather than infrastructure challenges while achieving faster training times and lower costs compared to traditional infrastructure approaches.

For Pathway to develop AI architectures beyond transformers, comprehensive observability of their training infrastructure is critical. Amazon SageMaker HyperPod integration with advanced monitoring tools, such as Amazon Managed Service for Prometheus, provides the deep insights needed for these workloads. The combination of visualization, metrics collection, and proactive optimization capabilities using Amazon Managed Grafana dashboards, allows Pathway to visualize complex distributed training patterns, monitor GPU utilization and memory patterns, and track inter-node communication efficiency for their parallel training approaches.

This comprehensive observability stack helps reduce development cycles when iterating on new architectural approaches, optimizes cost-performance ratio for resource-intensive training, and ensures reliability and reproducibility of results across training runs. For more information about this integration, and quick ways to deploy this observability stack, see the awsome-distributed-ai repository or the Amazon SageMaker HyperPod documentation.

Amazon Elastic Fabric Adapter (EFA) integrates natively with the NVIDIA CUDA platform for GPU-accelerated computation and the NVIDIA Collective Communications Library (NCCL) for communication across GPUs. It allows distributing data, weights, activations, and more, in a reliable and scalable fashion. Pathway used Amazon Elastic Compute Cloud (Amazon EC2) p5en.48xlarge instances, which have up to 3200 Gbps of network performance per instance and NVIDIA’s accelerated AI infrastructure powered by NVIDIA H200 GPUs. All instances were interconnected using EFA and operated on an Amazon EC2 UltraCluster to reduce the network distance between GPUs and lower latency.

Proof of BDH architecture: BDH-CQ achieving 29.2% pass@2 on ARC-AGI

BDH-CQ achieved 29.2 percent pass@2 on the ARC-AGI benchmark at a cost of US $0.0007 per task.

Thanks to Pathway’s new approach to reasoning, in-context task acquisition and iterative latent computation, BDH-CQ changed the cost-accuracy Pareto frontier on the ARC-AGI-1 benchmark, as of August 2026.

ARC-AGI-1 presents an AI system with a small number of before-and-after examples that illustrate an unknown visual rule. The system must then infer that rule and apply it to a new grid, a capability often associated with human-like intelligence. BDH-CQ performs iterative computation inside a recurrent latent state and decodes only its candidate answers. Compared to transformers, which externalize their work as a chain-of-thought and increase latency and inference cost, BDH-CQ updates its model’s internal memory while examples are processed. It works through new problems without generating a long, verbalized reasoning trace.

The pattern that ARC-AGI explores resembles real-world challenges where systems must reason reliably as information and constraints change, such as investigating cyber security incidents, coordinating transportation networks, responding to real-time industrial operations, and operating autonomous agents across long-running workflows. BDH-CQ’s new approach shows it can make these applications less costly, more responsive, and more reliable.

Conclusion

BDH is a new large language model architecture inspired by scale-free biological networks. It draws on principles biology got right, namely local interaction, sparse activity, persistent state, and continual adjustment, and applies them to a modern sequence model. It offers a GPU-friendly implementation. BDH provides a post-transformer architecture built around recurrent memory and local computation. BDH-CQ extends it into a reasoning system that learns from context and performs iterative computation in a continuous latent workspace, which supports efficient reasoning beyond token-by-token generation.

To begin exploring and implementing the BDH architecture, you can access detailed technical documentation and a sample implementation through Pathway’s repositories. The implementation supports standard PyTorch workflows and can be readily integrated into existing ML pipelines.

Teams that push the boundaries of AI model development can use Amazon SageMaker HyperPod to maintain efficiency and reliability at scale. Amazon SageMaker HyperPod training offers a robust solution to common challenges in large model training. To learn more about Amazon SageMaker HyperPod, go to the AI on SageMaker HyperPod and find workshops, code examples, and troubleshooting guides. Or read about Checkpointless training on Amazon SageMaker HyperPod to learn about its resilience features.

About the authors

Paulo Aragão

Paulo is a Principal WW Specialist Solutions Architect focused on helping customers build their Frontier AI strategy on AWS. With over 20 years of experience dealing with High Performance Computing and AIML projects, he is passionate about working backwards from customer’s challenges and helping overcome them. To know more about his projects, go to his GitHub page.

Rodrigo Merino

Rodrigo is a Generative AI Solutions Architect Manager at AWS. With over a decade of experience deploying emerging technologies, from IoT to GenAI, Rodrigo guides customers to accelerate their AI/ML and generative AI journeys. He specializes in helping organizations train and build models on AWS, as well as operationalize end-to-end ML solutions. Rodrigo’s expertise lies in bridging the gap between cutting-edge technology and practical business applications, enabling companies to harness the full potential of AI.

Ludovic Arnould

Ludovic is a Frontier AI Model Solutions Architect at Pathway, specializing in the design and deployment of advanced AI solutions for enterprise customers. He holds a PhD in Machine Learning from Sorbonne Université and has published research at ICML, ICLR, and AISTATS. His experience spans deep learning, multimodal AI, large language and vision models, with a track record of taking research prototypes through to production-ready systems.