Can we change an AI model’s architecture without retraining it from scratch?

“Priming” reuses pretrained Transformer weights to build models with more efficient architectures at less than 1% of the original training cost and with throughputs up to twice as high.

Key takeaways
  • A new AWS paper introduces priming, a method that reduces pretraining costs for hybrid architectures by over 99% by leveraging pretrained Transformer weights, saving trillions of tokens and hundreds of thousands of GPU-hours.
  • A 32-billion-parameter Gated KalmaNet (GKA)-based hybrid model matches its source Transformer on reasoning benchmarks while decoding up to twice as quickly with 128,000 tokens of context.
  • State space models consistently outperform sliding-window attention (SWA) even when SWA receives twice the memory budget, and GKA's variable-iteration mechanism enables 10% faster decoding with only ~1% accuracy loss.
Was this answer helpful?

AI agents are rapidly scaling: longer sessions, deeper reasoning chains, more autonomy. But the Transformer architecture underneath them still pays a memory tax that grows with the number of tokens it processes. As agents become more capable, that cost scales with them.

Now, there’s a proliferation of hybrid architectures that use some combination of methods to achieve a better balance of memory, speed, and quality. Those methods include sparse attention, which selects a subset of past tokens to consider when selecting the next token; sliding-window attention, which preserves only the most recent tokens; and state space models, which compress information from the full history into a fixed-size state.

These hybrid architectures can be understood as points in a sequence-model design space: to differing degrees, they mix information while traversing a sequence and passing through different states. What distinguishes them from each other is how they retain and use the past.

Until now, exploring this design space has required training every new architecture from scratch, using trillions of tokens and hundreds of thousands of GPU-hours. This has made controlled comparisons of hybrid architectures extremely expensive.

In a new paper, “Priming: Hybrid state space models from pre-trained Transformers,” our team at Amazon Web Services addresses this problem. We introduce priming, a method that leverages pretrained Transformer weights as a starting point to reduce the pretraining cost of hybrid architectures by more than 99%.

With priming, model developers can ask questions that were previously too expensive to answer at scale: Which attention layers can be replaced by more-efficient layers? How many attention layers can be replaced without sacrificing quality? Which state space layer works best? And what does the trade-off between speed and accuracy look like?

In our paper, we compare hybrid models that use three different types of state space models (Gated KalmaNet, Gated DeltaNet, and Mamba-2). Some of our findings:

  • Not all state space models are equal. Under controlled conditions, Gated KalmaNet (GKA) performs best, followed by Gated DeltaNet (GDN) and then Mamba-2. GKA uniquely offers variable-length inference computation without retraining and gives the user direct control over its position on the speed-accuracy frontier.
    Clock time vs. accuracy graph.png
    A 32-billion-parameter hybrid model reaches each accuracy threshold faster than an equivalently sized Transformer, converging to similar final accuracy in less wall clock time. The dataset is the AIME 2025 math competition benchmark (hard subset). Accuracy at time t is the fraction of rollouts for which a correct final answer has been produced by time t.
  • A good lossy summary of everything beats a perfect recording of the most recent past. State space models consistently outperform sliding-window attention (SWA), even when SWA gets twice the memory budget.
  • Same quality, twice the throughput. A 32-billion-parameter GKA-based hybrid model matches its source Transformer on reasoning benchmarks while decoding up to twice as quickly with 128,000 tokens of context and reaching correct answers in as little as two-thirds of the wall-clock time.

To foster further research on hybrid architectures, we have released a zoo of primed hybrid models at the eight-billion-parameter and 32-billion-parameter scales for long-context reasoning and instruction following, together with training code, optimized kernels, and a vLLM-serving plugin under the Apache 2.0 License.

The attention mechanism

A Transformer’s attention mechanism — which determines how relevant previously seen tokens are to the token currently being processed — relies on three different vectors, which take their names from database design: the query, the key, and the value. For each token, the query captures what that token is “looking for,” the key captures what it “offers,” and the value carries the actual information it will pass along. Relevance is computed by comparing the current token’s query against every other token’s key; the closer the match, the more that token’s value contributes to the output of attention.

Each of these vectors is produced by multiplying the token by a learned weight projection matrix. The Transformer outputs one token at a time, and each output token becomes a new input, so the context grows one token at a time as well. To save computation, the Transformer thus stores the key and value projections of each new input in a “KV cache”.

Two kinds of memory layers: Eidetic versus fading

The attention mechanism gives Transformers what we call eidetic memory: verbatim recall of every past token. The possibility of recalling over an LLM’s entire context window is what made in-context learning, long-document reasoning, and retrieval-augmented generation possible. But it comes at a cost that scales with the number of input tokens.

However, not all attention layers do the same job. Some perform long-range retrieval across tens of thousands of tokens. Others operate more locally, tracking the current clause, copying nearby tokens, and the like. The local ones don't need a growing cache: a compact, fixed-size summary of the past — what we call fading memory — suffices.

Hybrid models mix both: attention layers (eidetic, growing cache) for the layers that need global recall and state space layers (fading, fixed size) for the rest. The result is a model with a reduced memory footprint that still has access to verbatim recall where it matters.

Priming: Knowledge isn't locked to the architecture

Exploring hybrids’ design space has been impractical because each hybrid configuration would normally require pretraining from scratch. Priming sidesteps this by reusing the stored knowledge in a pretrained Transformer's weights.

Priming is possible because a Transformer is itself a special case of a state space realization. The outputs of both attention and SSM layers can be expressed as the result of multiplying the value sequence by a “mixing” matrix that decides how much each past token contributes to each current token. Attention computes the mixing matrix by multiplying the query and key matrices; an SSM produces a similar output through a recurrent state of fixed dimension that all past information must pass through.

Priming takes advantage of this structural similarity: if an SSM with a fixed-dimensional state can produce a mixing matrix close enough to the attention layer's, the query, key, and value projections from the source Transformer can be reused as the SSM's own input projections, with only the SSM-specific parameters needing to be learned.

Whether the approximation holds depends on whether the attention layer performs complex, long-range retrievals or simple, local operations. We distinguish between the two using a cheap empirical proxy: replace each layer with an SWA — which preserves only a fraction of the context — and measure the degradation.

Layers that tolerate this substitution are natural candidates for SSM conversion: their mixing matrices don’t need to capture long-range relationships, so the state of an SSM can faithfully represent what they do. This proxy connects to a formal quantity from dynamical-systems theory, the Hankel rank of the layer's mixing matrix, which measures how much information must pass through the state bottleneck and therefore the minimum SSM state dimension required to realize the layer exactly.

Priming proceeds in three stages:

  1. Layer selection: identify which layers can be faithfully compressed into a fixed-size state while retaining the rest as attention;
  2. Weight transfer and alignment: initialize SSM layers from the corresponding attention weights, then run a short distillation process aligning each SSM to the behavior of the layer it replaces;
  3. End-to-end adaptation: fine-tune the full hybrid model on next-token prediction, letting all layers coadapt.
Workflow.png
Priming begins with the use of a simple proxy (a sliding-window attention) to identify attention layers that largely operate locally. The weights of those layers are distilled into state-space-model (SSM) layers, which are then aligned with the behaviors of the layers they replace. Finally, the full hybrid model is fine-tuned on next-token prediction.

The entire pipeline consumes less than 0.5% of the source model's pretraining token budget, and it works across Transformer families (Qwen, Llama, Mistral), model classes (dense or MoE), and scales (eight billion to 32 billion and more).

With priming we can now explore the design space, the layer type, the hybrid ratio, and layer placement cheaply and systematically, because every variant starts from the same strong initialization and requires minimal training, compared to pretraining.

A controlled comparison of fading-memory layers

For instance, consider this question: given the same data, the same training budget, and the same ratio of eidetic to fading layers, which SSM layer works best in a hybrid model?

In our paper, we evaluate three families of state space models, each representing a different approach to using a fixed-size state:

  • Mamba-2 applies a uniform scalar decay to the entire state at each step, a running average that forgets everything at the same rate. When a new token arrives, it is written into the state, while all previous entries are uniformly attenuated. Mamba-2 is simple and efficient but indiscriminate.
  • Gated DeltaNet (GDN) adds targeted updates on top of uniform decay. Rather than decaying everything uniformly, it can selectively erase the value previously associated with a specific key before writing the new one (the delta rule). This makes it more expressive than Mamba-2: it can revise a specific association without uniformly decaying the rest.
  • Gated KalmaNet (GKA) goes further. Its erase direction depends not only on the current key but on statistics of all past keys, a history-aware update that accounts for the cumulative structure of everything the model has seen.
Qwen accuracy on four datasets.png
The results of hybrid models using Mamba-2, GDN, and GKA state space models on four tasks. GKA generally outperforms GDN, which always outperforms Mamba-2, in line with theoretical predictions.

The theoretical expressiveness hierarchy predicts the empirical one: GKA > GDN > Mamba-2. The gap widens on long-context reasoning, where compression quality matters most, and narrows on short-context tasks. Furthermore, at the 32-billion-parameter (32B) scale, hybrid GKA is within 1% of its source Transformer trained on the same data, despite requiring half the KV cache at inference.

Fading memory beats verbatim windowed memory

An interesting finding in our study comes from a direct comparison of hybrid models using SSM layers and those using sliding-window attention (SWA), which retain a fixed window of recent tokens verbatim rather than compressing the past into a state. One might expect that verbatim recall, even over a limited window, would outperform lossy compression. After all, no information is lost within the sliding window’s range. And stacking multiple SWA layers increases the “receptive field” of the end-to-end model.

Instead, every SSM-based hybrid we tested outperformed the corresponding SWA hybrid on long-context benchmarks, even when the window gets two to four times as much memory as the state.

Qwen vs. hybrid.png
Hybrid models using lossy context summaries (orange) consistently outperformed hybrid models using sliding context windows (teal).

A runtime knob for speed versus accuracy: Variable computation at inference time

GKA has a property that neither Mamba-2 nor GDN offers: variable test-time computation.

Mamba-2 and GDN have the same computation budget for every input token, regardless of its content. GKA, by contrast, solves a regression problem iteratively, and the iteration count is adjustable at serving time without retraining.

Long context vs. speedup.png
The Gated KalmaNet (GKA) accuracy-speed frontier, at four different context lengths. Geometric shapes indicate different numbers of iterations on a regression problem.

In practice, serving GKA at 10 iterations instead of its usual 30 sacrifices roughly 1% of its accuracy on AIME 2025 while delivering 10% faster decoding. A single trained model can serve multiple points on the speed-accuracy frontier, so model providers can offer different latency tiers without maintaining separate model variants.

This connects to a principle we have discussed in earlier work: intelligence is the ability to allocate effort commensurate with difficulty. GKA's variable-computation mechanism is an architectural embodiment of this principle: a deployment that demands speed can run fewer iterations, while one that demands accuracy can run more.

Faster reasoning, not just faster tokens

Raw per-token throughput tells only part of the story. A model that is cheaper per token can still be slower end to end if its reasoning is weaker and it needs more tokens to compensate. The deployment-relevant question is how much more rapidly, in wall clock terms, the hybrid reaches a target accuracy.

We measured this on AIME 2025, a challenging math competition benchmark, comparing our GKA-Primed-HQwen3-32B-Reasoner against the source Qwen3-32B trained on the same reasoning recipe. Both models generate similar-length thinking traces (a median of about 14,000 tokens on the full set, 28,000 on the hardest problems) — that is, they produce equally in-depth responses — but the hybrid has a much higher throughput.

At realistic concurrency levels, where the Transformer's KV cache saturates GPU memory, the hybrid reaches target accuracy thresholds 1.2 to 1.4 times as quickly on the full benchmark and up to 1.6 times as quickly on the hardest problems. This is a consequence of the SSM layers' fixed-size state, which roughly halves the per-sequence memory footprint, enabling more concurrent operations on the same hardware.

Decode throughput vs context length.png
Decode throughput speedup of GKA-Primed-HQwen3-32B over the source Transformer as a function of context length. Replacing half the attention layers with a fixed-size SSM state roughly halves the size of the KV cache, yielding up to 2.06 times the throughput at a context length of 128,000 tokens. Lowering the number of GKA solver iterations from 30 to 10 at serving time (no retraining necessary) trades a small accuracy margin for additional throughput.

The throughput advantage grows with context length: 1.50× at 16K tokens, scaling to 2.06× at 128K for the 32B model (GKA at 10 iterations). With the 8B parameter models, the gains are even greater: 1.94× at 16K to 2.29× at 128K. At long context, KV cache reads dominate the forward pass through the Transformer, and halving the cache roughly doubles the effective batch size at memory saturation.

This has implications beyond inference serving. In reinforcement learning (RL) post-training pipelines, throughput gates how rapidly the policy generates correct trajectories. The hybrid's advantage translates directly to a 20% to 100% increase in RL training data rate at a fixed GPU budget.

Comparison of state space models.png
Reasoning accuracy of GKA-primed hybrid models versus source Qwen3 Transformer models across five benchmarks at eight-billion-parameter and 32-billion-parameter scale. Primed hybrid models match their source Transformers in quality at half the memory, and they decode faster. At eight billion parameters, the GKA hybrid uniformly outperforms its GDN counterpart, consistent with the theoretical expressiveness hierarchy GKA > GDN.

Beyond the training window: Context extension with state composition

Our primed hybrids natively support 128,000 tokens. But agent sessions, codebases, and long legal documents routinely exceed that. Extending context further would normally require another round of continued pretraining on curated data. Can the hybrid architecture itself generalize to longer contexts without additional training?

Existing training-free methods for expanding the context window fall into two classes: those for pure Transformers (positional-embedding rescaling schemes like YaRN) and those for pure SSMs, such as our PICASO, which we presented last year at the International Conference on Learning Representations (ICLR). Neither addresses the hybrid setting, where KV caches and recurrent states must be reconciled jointly.

In our new arXiv paper, we describe a method of extending hybrid models’ context beyond their native limits without training. The procedure is simple:

  • partition a long input into chunks shorter than the native context length and prepend a shared prefix to each chunk;
  • process all chunks independently in parallel;
  • concatenate the KV caches across chunks while merging the SSM states via the weighted-averaging scheme we introduced in our PICASO paper.
State composition for hybrid models.png
State composition for hybrid models. Long inputs are partitioned into chunks at the native context length and processed independently. KV caches are concatenated, while SSM states are merged via the weighted-averaging scheme we proposed in our PICASO paper.

This approach enables us to extend the context length of our primed hybrid models from 128K to 256K tokens while preserving substantial accuracy on both retrieval (RULER NIAH) and reasoning (BABILong) tasks at lengths the model was never trained on. In our experiments we show that hybrid state composition preserves the SSM expressivity hierarchy (GKA > GDN > Mamba-2).

The hybrid design space is now open

All artifacts discussed in this blog are released under the Apache 2.0 License in our Hybrid Model Factory on GitHub and our model zoo on HuggingFace:

  • A collection of primed hybrids for long-context reasoning and instruction following (8B and 32B);
  • The priming training code;
  • Point-to-point sequence parallelism algorithms for SSM layers (up to 2.3 times as fast as all-to-all-based alternatives at multinode scale);
  • Universal sequence parallelism for prototyping novel layers on long sequences without custom kernels or distributed-training expertise;
  • Optimized Triton kernels for GKA decoding and training;
  • A vLLM serving plugin supporting primed hybrid architectures with continuous batching, chunked prefill, and tensor parallelism for inference deployment at scale.

If you serve long-context workloads, the 32B GKA reasoning model with the vLLM plugin gives you higher throughput with no quality loss on long context reasoning tasks.

If you're scaling RL post-training, the hybrid's throughput advantage means more correct rollouts per unit time at a fixed GPU budget, for faster policy iteration without changing your training recipe.

If you're researching new architectures, priming lets you evaluate any layer type under controlled conditions at a fraction of the pretraining cost while starting from a strong initialization. The universal SP module means you don't need custom distributed kernels to train at long context.

Priming is not tied to a specific Transformer family or model class. Extending it to Llama, Mistral, or MoE models requires no algorithmic changes. We hope these tools enable the community to explore architectures beyond what we've tested and to develop new algorithms and training recipes for building scalable next-generation architectures for AI agents.

For full details, see our paper, "Priming: Hybrid state space models from pre-trained Transformers". Code, models, and serving plugin are available on GitHub, and the models are also on Hugging Face.

Research areas

Related content

US, WA, Seattle
Amazon's eCommerce Foundation (eCF) organization provides the core technologies that drive and power Amazon's Stores, Digital, and Other (SDO) businesses. Millions of customer page views and orders per day are enabled by the systems eCF builds from the ground up. CloudTune, within eCF, empowers growth and business agility needs by automatically and efficiently managing AWS capacity and business processes needed to safely meet Amazon's customer demand. CloudTune serves its primary customers, internal software teams, through forecast-driven automation of cost controllership, capacity management, and scaling. We predict expected load, and drive procurement and allocation of AWS capacity for new product launches and high-velocity events like Prime Day and Cyber Monday. CloudTune is looking for a Data Scientist to join our forecasting and optimization team. The team develops sophisticated algorithms that combine machine learning-based demand forecasting with mathematical optimization to solve large-scale capacity planning problems under uncertainty. We work with massive datasets — spanning thousands of availability zone and instance family combinations across global regions — to determine optimal resource allocation strategies that balance infrastructure holding costs against service availability requirements. These models directly inform multi-million dollar capacity investment decisions and drive automated procurement, retention, and release policies across Amazon's compute infrastructure. As a Data Scientist in CloudTune, you will work with other scientists, software engineers, data engineers, and product managers on a variety of important research problems in the areas of stochastic optimization, time series modeling, and operations research. You will work on statistical problems with a high level of ambiguity. You will analyze and process large amounts of data, develop new algorithms and improve existing approaches based on statistical models, machine learning algorithms and big data solutions to automatically scale Amazon’s compute infrastructure, optimizing the balance between availability risk and cost efficiency for all of Amazon businesses. Key job responsibilities - Process and analyze large data sets, mining additional data sources as needed - Build statistical models and drive scalable solutions for multi-year capacity demand forecasting horizons - Create, enhance, and maintain technical documentation, and present to other scientists and business leaders - Process and analyze large-scale operational data to identify patterns and inform model development - Create, enhance, and maintain technical documentation, and present research findings to scientists, engineering teams, and senior leadership
US, CA, Irvine
Do you want to write the algorithms that control next-generation batteries, and shape the future of energy storage? This role sits where electrochemistry, data/ML, and control theory meet. You will develop the physics-informed and data-driven methods that estimate state of charge, health, power capability, and remaining life, and validate them against lab and fleet data until they are ready to ship. Your algorithms will decide how hard a battery can be pushed, when it needs attention, and how long it will last, so they shape the safety, warranty, and economics of the whole product. You will work side by side with cell, systems, and software engineers, and as an early hire you will help define how we develop and validate battery algorithms. This role can be based in Boston, Irvine, Sunnyvale, or Seattle, and involves occasional domestic and international travel, up to 10% of the time. Key job responsibilities • Develop and own algorithms for state of charge, state of health, power capability, and remaining useful life, combining physics-based models with data-driven and estimation methods. • Deliver validated reference implementations, test vectors, and accuracy targets that define what production code must achieve. • Design the test plans and data needed to develop and validate algorithms, working with the cell and testing teams to turn lab and field data into model improvements. • Analyze algorithm performance across the deployed fleet, find where models drift or fail, and drive improvements. • Own the accuracy and robustness of battery algorithms, and sign off that production implementations match the validated reference. Provide algorithm evidence for safety certification. A day in the life Your week moves between data, models, and experiments. One day you might be working out why a health estimate drifts on cells that have aged in the field, and the next deciding which model structure holds accuracy at cold temperatures. You could be designing a test with the cell team to expose a behavior your model misses, or reviewing an implementation with an applied scientist to confirm that what ships matches what you validated. You will regularly share results with software, systems, and cell leaders, and your findings will shape decisions well beyond the algorithm itself. About the team We are building Amazon's energy storage capability, providing power for advanced technologies while contributing to a stable grid. Our product is highly connected, yet secure and resilient by design. We own it end to end, from software, ML, and controls through system design to cell selection and supply, and on to a safety-certified product in the field. The organization operates like a startup inside Amazon: small, fast-moving, and accountable for our decisions. We work across hardware, software, and science, with suppliers and partners worldwide, and with Amazon infrastructure, procurement, legal, and finance. The organization is at an early stage and growing quickly. You will be among the first algorithm hires, so you will help shape how we design, build, certify, and ship, with a direct line of sight to the product and business decisions your work informs. This is a dynamic role that will regularly present new challenges as the product and organization evolve.
US, WA, Redmond
At Amazon, we’re inventing on behalf of customers, and with Amazon Leo, we’re redefining what global connectivity looks like. Our mission is to deliver fast, affordable broadband to unserved and underserved communities around the world through a constellation of low Earth orbit (LEO) satellites. Every system we build helps connect people to education, healthcare, opportunity, and each other. As a Data Scientist, you will be responsible for developing advanced analytics and machine learning solutions for user terminals. You will develop predictive models to proactively identify possible user terminal failures in the field. You will work in a collaborative environment with a multi-disciplinary team, including constellation, RF, antenna, silicon, algorithm, and software engineers. Key job responsibilities As a Data Scientist, you will develop analytic tools for a team developing current and future user terminals. Your responsibilities include: • Develop statistical and analytical tool to enable the regression decision from on-orbit and lab measurement of user terminals • Publish documents and create compelling visualizations and presentations to communicate insights to stakeholders • Create and manage datasets for continued pre-training and supervised fine-tuning of LLMs • Develop scalable visualizations for analysis of user terminal performance • Work closely with constellation, RF, antenna, silicon, algorithm, and software engineers to root-cause the failures using data as the primary tool • Drive consensus on metrics and analysis approaches to support product development strategy Export Control Requirement: Due to applicable export control laws and regulations, candidates must be a U.S. citizen or national, U.S. permanent resident (i.e., current Green Card holder), or lawfully admitted into the U.S. as a refugee or granted asylum. A day in the life As a Data Scientist in the LEO Customer Terminal Team, you will work daily with satellite constellation, algorithm, RF, antenna, silicon, hardware, and software teams in a collaborative environment. Your focus will be using data as an intelligence source to enable design decisions for the team. About the team The LEO Customer Terminal team is responsible for developing both outdoor and indoor devices that enable customers to access internet service via the LEO satellite network. We own the entire process from early prototypes through mass production, including requirements documentation, architecture definition, hardware development, algorithm development, and all integration and verification testing.
US, CA, San Francisco
The Amazon Center for Quantum Computing (CQC) is seeking an Applied Scientist to research and design cryogenic control hardware for large-scale quantum processors, with an emphasis on superconducting digital circuits. Controlling and reading out orders of magnitude more qubits than today's systems demands rethinking where classical control lives and what technology implements it; you will design and model superconducting digital logic (e.g., single-flux-quantum-family circuits) and related cryogenic electronics, evaluate them against the wiring, power, and thermal budgets of a scaled system, and co-design control approaches with the qubit devices they serve. You will carry designs from concept through simulation, layout, design review, and fabrication handoff, and your evaluations will inform which control-electronics directions the organization invests in. Key job responsibilities • Design and simulate superconducting digital circuits and cryogenic control/readout electronics for scalable qubit control • Develop and validate Josephson-junction circuit models, timing and margin analyses, and verification methodology for superconducting digital designs • Quantify system-level trade-offs — wiring count, power dissipation at each cryogenic stage, latency, footprint — across candidate control architectures, and communicate recommendations • Co-design control interfaces with qubit-device colleagues so control electronics and quantum devices are developed as one system • Deliver design artifacts through the organization's design-request and design-review processes, including tapeout-ready layouts and verification reports • Present results internally and, where appropriate, publish in peer-reviewed venues About the team The Enabling Technologies team develops technologies and relationships facilitating scaling, including alternative quantum processor architectures, superconducting qubit device packaging, and cryogenic control device electronics. We work alongside a co-located quantum processor design team focused on current-generation devices; the two teams share design and review processes, and moving ideas between them is an explicit part of our charter.
US, WA, Redmond
Amazon Leo is building a constellation of thousands of low Earth orbit satellites. It will bring fast, reliable internet to customers beyond the reach of existing networks. On the Guidance, Navigation and Control (GNC) team, you will design the algorithms, simulations and onboard autonomy that keep every satellite precisely pointed and stable. Your work lets the payload serve customers on the ground, and it keeps the fleet flying safely. This is satellite GNC at constellation scale. The constellation is launching and growing today, so you will see your work go from concept to orbit, and then fly on many satellites. In this role, you will: - Take your GNC algorithms from concept to orbit, and stay with them through integration, test and on-orbit operations. - Write the production C++ flight code for your own algorithms. The code that flies is the code you write. - Solve hard problems in precise pointing, payload cueing and fleet autonomy. - Work across design, simulation, test and flight operations, not a narrow slice of the lifecycle. **Export Control Requirement:** Due to applicable export control laws and regulations, candidates must be a U.S. citizen or national, U.S. permanent resident (i.e., current Green Card holder), or lawfully admitted into the U.S. as a refugee or granted asylum. Key job responsibilities - Design and analyze algorithms for state estimation, attitude determination and control, precise pointing and payload cueing against demanding performance requirements. - Implement these algorithms in production C++ flight software. Our GNC engineers own their algorithms all the way into flight, and partner with our embedded software engineers to integrate and test them. - Write concepts of operations for each phase of a satellite's life, from launch and orbit raising to service and deorbit, including failures and recovery. Balance what the payload needs against momentum, power, thermal and ground-contact limits. - Develop momentum management strategies with reaction wheels and magnetic torquers. Help select, size and place the GNC sensors and actuators for future satellites, and design the algorithms that use them. These include star trackers, magnetometers, sun sensors, IMUs and GNSS receivers. - Develop onboard autonomy, including Fault Detection, Isolation and Recovery (FDIR), so that the fleet operates safely at scale with minimal ground intervention. - Build and validate satellite and constellation models and simulations at several levels of fidelity, including 6-DOF simulation and Monte Carlo analysis. Verify closed-loop performance in hardware-in-the-loop (HITL) tests. - Support component environmental tests, functional and performance checkout, and subsystem and satellite integration in our labs. - Analyze telemetry from the whole fleet. Build the metrics, dashboards and tools that track pointing performance and payload operations time as the constellation grows, and use them to improve the fleet. - Operate the GNC system on orbit and take part in the team's on-call rotation. Review fleet data, support maneuvers, and resolve anomalies when on-orbit behavior differs from the ground model. - Review designs and code with the team, and document your algorithms, analyses and results clearly. A day in the life This role covers a wide range of problems across the spacecraft and the network. You will derive and prototype estimation and control algorithms, and then turn them into production flight code. You will prove them in 6-DOF and Monte Carlo simulation and in hardware-in-the-loop tests in the lab. You will support satellite integration. When a satellite does not behave as the ground model predicted, you will dig into fleet telemetry, find the cause and drive the fix. You will own each problem end to end, with the focus to go deep on it. GNC is at the center of safe flight, so you will work across the whole program. The payload, power, thermal and operations teams all depend on GNC, and you will balance their needs against the pointing and momentum budgets. You will partner closely with avionics, flight software and embedded software engineers to make your algorithms run reliably on the spacecraft. You will help the mechanical teams shape the next generation of satellite hardware, from reaction wheels to star trackers. As satellites move from design through manufacturing into orbit, you will work with systems engineering and Assembly, Integration and Test. The problems are hard: tight pointing requirements, a large constellation to keep flying safely, and the chance to invent capabilities that do not exist yet. Constant work with other teams makes this a place where you keep learning and growing. About the team You will join the Satellite Attitude Determination and Control team. We own GNC for our satellites end to end: algorithm design, flight software, lab integration, and flight operations of our own satellites on orbit. Our engineers bring deep experience from many satellite programs and other flight vehicles, with real strength in both our mission and the core GNC disciplines. We design, prototype, test, iterate and learn together. Because GNC is central to safe flight, we drive the concepts of operations and many of the system-level analyses on the program. We also have a lot of freedom. We trust our engineers to choose their own tools and approaches for the problems they face, and to own the results. If you want the autonomy to solve hard problems your way, with a team of strong engineers, you will feel at home here.
US, WA, Redmond
Amazon Leo is building a constellation of thousands of low Earth orbit satellites. It will bring fast, reliable internet to customers beyond the reach of existing networks. On the Guidance, Navigation and Control (GNC) team, you will design the algorithms, simulations and onboard autonomy that keep every satellite precisely pointed and stable. Your work lets the payload serve customers on the ground, and it keeps the fleet flying safely. This is satellite GNC at constellation scale. The constellation is launching and growing today, so you will see your work go from concept to orbit, and then fly on many satellites. In this role, you will: - Take your GNC algorithms from concept to orbit, and stay with them through integration, test and on-orbit operations. - Write the production C++ flight code for your own algorithms. The code that flies is the code you write. - Solve hard problems in precise pointing, payload cueing and fleet autonomy. - Work across design, simulation, test and flight operations, not a narrow slice of the lifecycle. **Export Control Requirement:** Due to applicable export control laws and regulations, candidates must be a U.S. citizen or national, U.S. permanent resident (i.e., current Green Card holder), or lawfully admitted into the U.S. as a refugee or granted asylum. Key job responsibilities - Design and analyze algorithms for state estimation, attitude determination and control, precise pointing and payload cueing against demanding performance requirements. - Implement these algorithms in production C++ flight software. Our GNC engineers own their algorithms all the way into flight, and partner with our embedded software engineers to integrate and test them. - Write concepts of operations for each phase of a satellite's life, from launch and orbit raising to service and deorbit, including failures and recovery. Balance what the payload needs against momentum, power, thermal and ground-contact limits. - Develop momentum management strategies with reaction wheels and magnetic torquers. Help select, size and place the GNC sensors and actuators for future satellites, and design the algorithms that use them. These include star trackers, magnetometers, sun sensors, IMUs and GNSS receivers. - Develop onboard autonomy, including Fault Detection, Isolation and Recovery (FDIR), so that the fleet operates safely at scale with minimal ground intervention. - Build and validate satellite and constellation models and simulations at several levels of fidelity, including 6-DOF simulation and Monte Carlo analysis. Verify closed-loop performance in hardware-in-the-loop (HITL) tests. - Support component environmental tests, functional and performance checkout, and subsystem and satellite integration in our labs. - Analyze telemetry from the whole fleet. Build the metrics, dashboards and tools that track pointing performance and payload operations time as the constellation grows, and use them to improve the fleet. - Operate the GNC system on orbit and take part in the team's on-call rotation. Review fleet data, support maneuvers, and resolve anomalies when on-orbit behavior differs from the ground model. - Review designs and code with the team, and document your algorithms, analyses and results clearly. A day in the life This role covers a wide range of problems across the spacecraft and the network. You will derive and prototype estimation and control algorithms, and then turn them into production flight code. You will prove them in 6-DOF and Monte Carlo simulation and in hardware-in-the-loop tests in the lab. You will support satellite integration. When a satellite does not behave as the ground model predicted, you will dig into fleet telemetry, find the cause and drive the fix. You will own each problem end to end, with the focus to go deep on it. GNC is at the center of safe flight, so you will work across the whole program. The payload, power, thermal and operations teams all depend on GNC, and you will balance their needs against the pointing and momentum budgets. You will partner closely with avionics, flight software and embedded software engineers to make your algorithms run reliably on the spacecraft. You will help the mechanical teams shape the next generation of satellite hardware, from reaction wheels to star trackers. As satellites move from design through manufacturing into orbit, you will work with systems engineering and Assembly, Integration and Test. The problems are hard: tight pointing requirements, a large constellation to keep flying safely, and the chance to invent capabilities that do not exist yet. Constant work with other teams makes this a place where you keep learning and growing. About the team You will join the Satellite Attitude Determination and Control team. We own GNC for our satellites end to end: algorithm design, flight software, lab integration, and flight operations of our own satellites on orbit. Our engineers bring deep experience from many satellite programs and other flight vehicles, with real strength in both our mission and the core GNC disciplines. We design, prototype, test, iterate and learn together. Because GNC is central to safe flight, we drive the concepts of operations and many of the system-level analyses on the program. We also have a lot of freedom. We trust our engineers to choose their own tools and approaches for the problems they face, and to own the results. If you want the autonomy to solve hard problems your way, with a team of strong engineers, you will feel at home here.
US, WA, Seattle
Do you want to make AI provably correct for chip design and verification? We're hiring an Applied Scientist to combine large language models with formal methods such as Lean 4 and turn AI from a helpful assistant into a trusted engineering tool for silicon. Our team builds the formal verification methods behind custom silicon at Amazon scale. As AI agents take on more of chip design and verification, we need people who can make sure their output is checkable, sound and grounded in mathematics. You will research and build systems where LLMs work alongside interactive theorem provers, SMT solvers and model checkers. Engineers write the specifications. AI refines them against RTL, generates proof steps and closes proofs at machine speed. Every result is checked by a formal kernel, not trusted on faith. Work at the frontier of AI reasoning, theorem proving and hardware verification Ship solutions that directly affect real chips running in AWS data centers Publish and help shape how the industry formalizes AI use in hardware engineering Key job responsibilities Design and build LLM-driven systems for formal verification and design tasks, including proof generation, invariant and lemma discovery, spec-to-RTL refinement, and counterexample analysis Develop frameworks in Lean 4 (and related tools such as SMT solvers, model checkers and commercial FV engines) that formally check AI-generated artifacts, so they are correct by construction rather than just plausible Formalize hardware semantics, ISAs and microarchitectural properties in Lean 4 to create machine-checkable foundations for AI agents to reason over Track and evaluate state-of-the-art AI work (LLM agents, autoformalization, neural theorem proving, reinforcement learning from proof feedback) and turn promising research into production tools Build benchmarks and evaluation methods that measure AI soundness, proof success rate and engineering impact on real designs Partner with FV and design engineers to find the highest-leverage problems and deploy solutions into active chip programs Publish research, file patents and represent Amazon in the formal methods and AI communities Mentor engineers and scientists on applying formal AI methods
US, CA, Pasadena
We are seeking Applied Scientists to join Amazon. In this role, you will develop and deploy the science of safe autonomy on highly dynamic robots, advancing the frontier through both foundational research and realization on hardware. You will contribute to achieving safe autonomy across three key areas: control barrier function (CBF) theory for robust and performant safety on hardware; safe reinforcement learning for agile whole-body control; and layered safety filters that interface with learning, perception, and semantic reasoning systems. You will validate your work experimentally on state-of-the-art robotic platforms — removing bottlenecks to deployment and enabling robots to safely operate around humans. You will work with the inventor of CBFs, as well as top scientists and engineers at Amazon developing the next generation of safe autonomy. Key job responsibilities -Advance the science of safe autonomy from formal foundations to the integration with learning and perception to validation on hardware — with particular emphasis on methods that bridge these domains. -Test and validate hardware, with a focus on next generation robotic systems. Including locomotion, manipulation and loco-manipulation. -Leverage the deployments on hardware to validate the underlying science, identifying gaps between theory and practice that drive the next cycle of research. hardware to validate the underlying science, identifying gaps between theory and practice that drive the next cycle of research. -Publish research at top-tier robotics, control, and ML venues, and contribute to Amazon’s scientific reputation in advanced robotics. -Collaborate with SAF Lab and Amazon production teams to move research into robots deployed at Amazon scale. A day in the life Amazon offers a full range of benefits that support you and eligible family members, including domestic partners and their children. Benefits can vary by location, the number of regularly scheduled hours you work, length of employment, and job status such as seasonal or temporary employment. The benefits that generally apply to regular, full-time employees include: -Medical, Dental, and Vision Coverage -Maternity and Parental Leave Options -Paid Time Off (PTO) -401(k) Plan If you are not sure that every qualification on the list above describes you exactly, we’d still love to hear from you! At Amazon, we value people with unique backgrounds, experiences, and skillsets. If you’re passionate about this role and want to make an impact on a global scale, please apply! About the team The Safe Autonomy Frontiers (SAF) Lab is the first industry research lab dedicated to safe autonomy, founded by the inventor of control barrier functions. We are developing a universal safety layer for the next generation of robotic systems: mobile robots, manipulators, quadrupeds, and humanoids. Here you will advance performant safety for highly dynamic robots — CBF theory integrated with perception and learning, evaluated on next-generation platforms — and your work will underpin robots operating alongside people at Amazon’s unprecedented scale.
US, MA, North Reading
Robotics Applied Scientist II Intern/Co-op (2027) Are you passionate about robotics and research? Do you want to solve real customer problems through innovative technology? Do you enjoy working on scalable research and projects in a collaborative team environment? Do you want to see your science solutions directly impact millions of customers worldwide? We’re looking for current PhD students with a passion for robotic research and applications to join us in 2027 as Robotics Applied Scientist II Interns/Co-ops to shape the future of robotics and transform automation at Amazon’s scale. By applying to this role, you will be considered for Robotics Applied Scientist II Intern/Co-op (2027) opportunities across various Robotics teams at Amazon with different robotics research focus. Internship positions are available for multiple locations, durations (3 to 6+ months), and year-round start dates (winter, spring, summer, fall). Join us at the forefront of Applied Science, where your contributions will shape the future of Robotics! Key job responsibilities As a Robotics Applied Scientist II Intern/Co-op, you will work alongside global experts to advance the state of the art in robotics and artificial intelligence, tackling innovative problems that bridge the gap between research and real-world robotics implementation to innovate and build solutions on behalf of our customers. We are particularly interested in students with specialization and expertise in one or more of the following research areas: - Robotics - Artificial Intelligence - Robotics Manipulation (e.g., robot arm, grasping, dexterous manipulation, end of arm tools/end effector) - Autonomous Mobile Robots - Mobile Manipulation - Multi-Robot Coordination - Autonomous Navigation - Locomotion - Motion/Path Planning - Controls - Perception - Sensing - Robot Learning - Machine Learning - Computer Vision - Large Language Models (LLM), Vision-Language-Action Models (VLA), Vision-Language-Models (VLM) - World Models - Human-Robot Interaction - Robotics Simulation - Optimization - And related! Key responsibilities may include: - Develop novel, scalable algorithms and/or modeling techniques that advance the state-of-the-art in science areas such as robotics, artificial intelligence, or related - Tackle challenging, innovative research problems within robotics, artificial intelligence, or related that bridge the gap between theoretical advancement and practical robotics implementation - Collaborate with cross-functional teams to solve complex problems, and to communicate research findings clearly, leveraging your expertise in robotics, artificial intelligence, or related - Demonstrate the ability to work independently, thrive in a fast-paced, ever-changing environment - Leverage AI-powered tools where applicable to accelerate research, experimentation, and prototyping. - Critically review and validate outputs from AI tools and automated systems. About the team Explore Robotics Teams at Amazon At Amazon, robotics is where innovative research meets real-world deployment. There are many different specialized robotics teams, including the few highlighted below. Learn more about Robotics at Amazon here: https://www.amazon.science/research-areas/robotics Fulfillment Technologies and Robotics We build dynamic partnerships between people and intelligent machines. This intricate collaboration helps Amazon fulfill orders with unmatched accuracy. Working in symphony with our robotic technology, you’ll have the opportunity to extend your technical capabilities by working alongside some of the industry’s most advanced technologies. This includes our fleet of autonomous mobile robots, sophisticated control software, and technologies like language perception, machine learning, object recognition, and semantic understanding of commands. These technologies help us deliver an ever-improving customer and employee experience, as well as improve the safety of our facilities. Amazon Robotics Amazon Robotics is at the epicenter of robotics innovation. We design, test, and build mobile robots and mechatronic systems. Our offices, labs, test floors, and factories are close together, which means we can design the hardware, write the code, test, build, and watch it all come to life. Our areas of focus include advancements in autonomous mobility, AI, robotic management software, and much more. Fauna Fauna is focused on building capable, safe, and genuinely delightful robots for everyday life. Fauna’s goal is simple: make robots people actually want to live and interact with everyday. Fauna develops tightly integrated hardware and software systems that make it faster, safer, and more intuitive to create real-world robotic products. Fauna’s work spans the full stack: mechanical design, control systems, dynamic modeling, and intelligent software. The focus is not just functionality, but overall experience. Amazon Fulfillment Technologies Amazon Fulfillment Technologies develops tech solutions that create network optimizations for all Amazon sites worldwide. We own network infrastructure and devices in all buildings and labor planning and scheduling technology for fulfillment centers in North America and Europe.
US, WA, Seattle
Amazon Web Services (AWS) provides a highly reliable, scalable, low-cost infrastructure platform in the cloud that powers hundreds of thousands of businesses in 190 countries around the world. We are hiring an Applied Scientist in our AWS Sales Science & Economics team. The team uses machine learning, econometrics, and data science to optimize AWS’s service catalogue, driving customer engagement and generating insights to guide AWS sales strategy. We use detailed customer behavioral and usage data to predict and understand what customers want from AWS. Key job responsibilities As an Applied Scientist, you will apply the latest research, state of the art machine learning and GenAI to build highly scalable models in support of predictive research and products used across AWS Sales and leadership. You will be responsible for designing and executing science experiments in short and agile cycles. You will use the latest AWS technologies in Machine Learning, Data Science, and AI. You will work closely with business leaders to identify opportunities to drive business growth by applying machine learning and building out predictive models to support analytics and insights. The ideal candidate will be curious and able to learn new technologies quickly. You will have hands-on experience building data science and machine learning pipelines in production working with large scale data sets. You are passionate about sharing your knowledge to enable others. You empathize with the needs of our customers and align your work accordingly, and provide thought leadership and guidance across the organization. A day in the life You will start your day reviewing experimental results and refining model architectures based on new insights. You might spend the morning prototyping a novel approach to a customer-facing problem, then shift to a design review with engineering partners in the afternoon to align on how your solution integrates into production systems. Along the way, you will mentor junior scientists, contribute to peer reviews, and stay connected to the latest research through reading groups and internal science discussions. Your contributions will move from research notebooks to live systems that customers interact with every day. About the team Our team is focused on applying advanced science and machine learning to build intelligent systems that improve the Amazon experience. We value collaboration, intellectual rigor, and a willingness to explore new ideas. You will work alongside scientists and engineers who are committed to turning ambitious research into practical, high-impact solutions. We are investing in the next generation of machine learning capabilities and are looking for team members who are excited to help define that future.