Social
Automated reasoning

A gentle introduction to automated reasoning

Meet Amazon Science’s newest research area.

By Byron Cook
December 1, 2021
18 min read
Share

This week, Amazon Science added automated reasoning to its list of research areas. We made this change because of the impact that automated reasoning is having here at Amazon. For example, Amazon Web Services’ customers now have direct access to automated-reasoning-based features such as IAM Access Analyzer, S3 Block Public Access, or VPC Reachability Analyzer. We also see Amazon development teams integrating automated-reasoning tools into their development processes, raising the bar on the security, durability, availability, and quality of our products.

The goal of this article is to provide a gentle introduction to automated reasoning for the industry professional who knows nothing about the area but is curious to learn more. All you will need to make sense of this article is to be able to read a few small C and Python code fragments. I will refer to a few specialist concepts along the way, but only with the goal of introducing them in an informal manner. I close with links to some of our favorite publicly available tools, videos, books, and articles for those looking to go more in-depth.

Let’s start with a simple example. Consider the following C function:

bool f(unsigned int x, unsigned int y) {
   return (x+y == y+x);
}



        
Take a few moments to answer the question “Could f ever return false?” This is not a trick question: I’ve purposefully used a simple example to make a point.
To check the answer with exhaustive testing, we could try executing the following doubly nested test loop, which calls f on all possible pairs of values of the type unsigned int:

            

    
#include<stdio.h>
#include<stdbool.h>
#include<limits.h>

bool f(unsigned int x, unsigned int y) {
   return (x+y == y+x);
}

void main() {
   for (unsigned int x=0;1;x++) {
      for (unsigned int y=0;1;y++) {
         if (!f(x,y)) printf("Error!\n");
         if (y==UINT_MAX) break;
      }
      if (x==UINT_MAX) break;
   }
}



        
Unfortunately, even on modern hardware, this doubly nested loop will run for a very long time. I compiled it and ran it on a 2.6 GHz Intel processor for over 48 hours before giving up.
Why does testing take so long? Because UINT_MAX is typically 4,294,967,295, there are 18,446,744,065,119,617,025 separate f calls to consider. On my 2.6 GHz machine, the compiled test loop called f approximately 430 million times a second. But to test all 18 quintillion cases at this performance, we would need over 1,360 years.
When we show the above code to industry professionals, they almost immediately work out that f can't return false as long as the underlying compiler/interpreter and hardware are correct. How do they do that? They reason about it. They remember from their school days that x + y can be rewritten as y + x and conclude that f always returns true.

            

    
        






    
    

        
            
Re:Invent 2021 keynote address by Peter DeSantis, senior vice president for utility computing at Amazon Web Services

        
        
            
Skip to 15:49 for a discussion of Amazon Web Services' work on automated reasoning.

        
    



        
An automated reasoning tool does this work for us: it attempts to answer questions about a program (or a logic formula) by using known techniques from mathematics. In this case, the tool would use algebra to deduce that x + y == y + x can be replaced with the simple expression true.
Automated-reasoning tools can be incredibly fast, even when the domains are infinite (e.g., unbounded mathematical integers rather than finite C ints). Unfortunately, the tools may answer “Don’t know” in some instances. We'll see a famous example of that below.
The science of automated reasoning is essentially focused on driving the frequency of these “Don’t know” answers down as far as possible: the less often the tools report "Don't know" (or time out while trying), the more useful they are.
Today’s tools are able to give answers for programs and queries where yesterday’s tools could not. Tomorrow’s tools will be even more powerful. We are seeing rapid progress in this field, which is why at Amazon, we are increasingly getting so much value from it. In fact, we see automated reasoning forming its own Amazon-style virtuous cycle, where more input problems to our tools drive improvements to the tools, which encourages more use of the tools.
A slightly more complex example. Now that we know the rough outlines of what automated reasoning is, the next small example gives a slightly more realistic taste of the sort of complexity that the tools are managing for us.

            

    
void g(int x, int y) {
   if (y > 0)
      while (x > y)
         x = x - y;
}



        
Or, alternatively, consider a similar Python program over unbounded integers:

            

    
def g(x, y):
   assert isinstance(x, int) and isinstance(y, int)
   if y > 0:
      while x > y:
         x = x - y



        
Try to answer this question: “Does g always eventually return control back to its caller?”
When we show this program to industry professionals, they usually figure out the right answer quickly. A few, especially those who are aware of results in theoretical computer science, sometimes mistakenly think that we can't answer this question, with the rationale “This is an example of the halting problem, which has been proved insoluble”. In fact, we can reason about the halting behavior for specific programs, including this one. We’ll talk more about that later.
Here’s the reasoning that most industry professionals use when looking at this problem:
  1. In the case where y is not positive, execution jumps to the end of the function g. That’s the easy case.
  2. If, in every iteration of the loop, the value of the variable x decreases, then eventually, the loop condition x > y will fail, and the end of g will be reached.
  3. The value of x always decreases only if y is always positive, because only then does the update to x (i.e., x = x - y) decrease x. But y’s positivity is established by the conditional expression, so x always decreases.
The experienced programmer will usually worry about underflow in the x = x - y command of the C program but will then notice that x > y before the update to x and thus cannot underflow.
If you carried out the three steps above yourself, you now have a very intuitive view of the type of thinking an automated-reasoning tool is performing on our behalf when reasoning about a computer program. There are many nitty-gritty details that the tools have to face (e.g., heaps, stacks, strings, pointer arithmetic, recursion, concurrency, callbacks, etc.), but there’s also decades of research papers on techniques for handling these and other topics, along with various practical tools that put these ideas to work.

            

    

        
    
        Policy-code.gif
    


        

        
    


    
        
Automated reasoning can be applied to both policies (top) and code (bottom). In both cases, an essential step is reasoning about what's always true.

    


        
The main takeaway is that automated-reasoning tools are usually working through the three steps above on our behalf: Item 1 is reasoning about the program’s control structure. Item 2 is reasoning about what is eventually true within the program. Item 3 is reasoning about what is always true in the program.
Note that configuration artifacts such as AWS resource policies, VPC network descriptions, or even makefiles can be thought of as code. This viewpoint allows us to use the same techniques we use to reason about C or Python code to answer questions about the interpretation of configurations. It’s this insight that gives us tools like IAM Access Analyzer or VPC Reachability Analyzer.
An end to testing?
As we saw above when looking at f and g, automated reasoning can be dramatically faster than exhaustive testing. With tools available today, we can show properties of f or g in milliseconds, rather than waiting lifetimes with exhaustive testing.
Can we throw away our testing tools now and just move to automated reasoning? Not quite. Yes, we can dramatically reduce our dependency on testing, but we will not be completely eliminating it any time soon, if ever. Consider our first example:

            

    
bool f(unsigned int x, unsigned int y) {
   return (x + y == y + x);
}



        
Recall the worry that a buggy compiler or microprocessor could in fact cause an executable program constructed from this source code to return false. We might also need to worry about the language runtime. For example, the C math library or the Python garbage collector might have bugs that cause a program to misbehave.
What’s interesting about testing, and something we often forget, is that it’s doing much more than just telling us about the C or Python source code. It’s also testing the compiler, the runtime, the interpreter, the microprocessor, etc. A test failure could be rooted in any of those tools in the stack.
Automated reasoning, in contrast, is usually applied to just one layer of that stack — the source code itself, or sometimes the compiler or the microprocessor. What we find so valuable about reasoning is it allows us to clearly define both what we do know and what we do not know about the layer under inspection.
Furthermore, the models of the surrounding environment (e.g., the compiler or the procedure calling our procedure) used by the automated-reasoning tool make our assumptions very precise. Separating the layers of the computational stack helps make better use of our time, energy, and money and the capabilities of the tools today and tomorrow.
Unfortunately, we will almost always need to make assumptions about something when using automated reasoning — for example, the principles of physics that govern our silicon chips. Thus, testing will never be fully replaced. We will want to perform end-to-end testing to try and validate our assumptions as best we can.
An impossible program
I previously mentioned that automated-reasoning tools sometimes return “Don’t know” rather than “yes” or “no”. They also sometimes run forever (or time out), thus never returning an answer. Let’s look at the famous "halting problem" program, in which we know tools cannot return “yes” or “no”.
Imagine that we have an automated-reasoning API, called terminates, that returns “yes” if a C function always terminates or “no” when the function could execute forever. As an example, we could build such an API using the tool described here (shameless self-promotion of author’s previous work). To get the idea of what a termination tool can do for us, consider two basic C functions, g (from above),

            

    
void g(int x, int y) {
   if (y > 0)
      while (x > y)
         x = x - y;
}



        
and g2:

            

    
void g2(int x, int y) {
   while (x > y)
      x = x - y;
}



        
For the reasons we have already discussed, the function g always returns control back to its caller, so terminates(g) should return true. Meanwhile, terminates(g2) should return false because, for example, g2(5, 0) will never terminate.
Now comes the difficult function. Consider h:

            

    
void h() {
   if terminates(h) while(1){}
}



        
Notice that it's recursive. What’s the right answer for terminates(h)? The answer cannot be "yes". It also cannot be "no". Why?
Imagine that terminates(h) were to return "yes". If you read the code of h, you’ll see that in this case, the function does not terminate because of the conditional statement in the code of h that will execute the infinite loop while(1){}. Thus, in this case, the terminates(h) answer would be wrong, because h is defined recursively, calling terminates on itself.
Similarly, if terminates(h) were to return "no", then h would in fact terminate and return control to its caller, because the if case of the conditional statement is not met, and there is no else branch. Again, the answer would be wrong. This is why the “Don’t know” answer is actually unavoidable in this case.
The program h is a variation of examples given in Turing’s famous 1936 paper on decidability and Gödel’s incompleteness theorems from 1931. These papers tell us that problems like the halting problem cannot be “solved”, if by“solved” we mean that the solution procedure itself always terminates and answers either “yes” or “no” but never “Don’t know”. But that is not the definition of “solved” that many of us have in mind. For many of us, a tool that sometimes times out or occasionally returns “Don’t know” but, when it gives an answer, always gives the right answer is good enough.
This problem is analogous to airline travel: we know it’s not 100% safe, because crashes have happened in the past, and we are sure that they will happen in the future. But when you land safely, you know it worked that time. The goal of the airline industry is to reduce failure as much as possible, even though it’s in principle unavoidable.
To put that in the context of automated reasoning: for some programs, like h, we can never improve the tool enough to replace the "Don't know" answer. But there are many other cases where today's tools answer "Don't know", but future tools may be able to answer "yes" or "no". The modern scientific challenge for automated-reasoning subject-matter experts is to get the practical tools to return “yes” or “no” as often as possible. As an example of current work, check out CMU professor and Amazon Scholar Marijn Heule and his quest to solve the Collatz termination problem.
Another thing to keep in mind is that automated-reasoning tools are regularly trying to solve “intractable” problems, e.g., problems in the NP complexity class. Here, the same thinking applies that we saw in the case of the halting problem: automated-reasoning tools have powerful heuristics that often work around the intractability problem for specific cases, but those heuristics can (and sometimes do) fail, resulting in “Don’t know” answers or impractically long execution time. The science is to improve the heuristics to minimize that problem.
Nomenclature
A host of names are used in the scientific literature to describe interrelated topics, of which automated reasoning is just one. Here’s a quick glossary:
  • A logic is a formal and mechanical system for defining what is true and untrue. Examples: propositional logic or first-order logic.
  • A theorem is a true statement in logic. Example: the four-color theorem.
  • A proof is a valid argument in logic of a theorem. Example: Gonthier's proof of the four-color theorem. 
  • A mechanical theorem prover is a semi-automated-reasoning tool that checks a machine-readable expression of a proof often written down by a human. These tools often require human guidance. Example: HOL-light, from Amazon researcher John Harrison. 
  • Formal verification is the use of theorem proving when applied to models of computer systems to prove desired properties of the systems. Example: the CompCert verified C compiler. 
  • Formal methods is the broadest term, meaning simply the use of logic to reason formally about models of systems. 
  • Automated reasoning focuses on the automation of formal methods. 
  • A semi-automated-reasoning tool is one that requires hints from the user but still finds valid proofs in logic. 
As you can see, we have a choice of monikers when working in this space. At Amazon, we’ve chosen to use automated reasoning, as we think it best captures our ambition for automation and scale. In practice, some of our internal teams use both automated and semi-automated reasoning tools, because the scientists we've hired can often get semi-automated reasoning tools to succeed where the heuristics in fully automated reasoning might fail. For our externally facing customer features, we currently use only fully automated approaches.
Next steps
In this essay, I’ve introduced the idea of automated reasoning, with the smallest of toy programs. I haven’t described how to handle realistic programs, with heap or concurrency. In fact, there are a wide variety of automated-reasoning tools and techniques, solving problems in all kinds of different domains, some of them quite narrow. To describe them all and the many branches and sub-disciplines of the field (e.g. “SMT solving”, “higher-order logic theorem proving”, “separation logic”) would take thousands of blogs posts and books.
Automated reasoning goes back to the early inventors of computers. And logic itself (which automated reasoning attempts to solve) is thousands of years old. In order to keep this post brief, I’ll stop here and suggest further reading. Note that it’s very easy to get lost in the weeds reading depth-first into this area, and you could emerge more confused than when you started. I encourage you to use a bounded depth-first search approach, looking sequentially at a wide variety of tools and techniques in only some detail and then moving on, rather than learning only one aspect deeply.
Suggested books:
International conferences/workshops:
Tool competitions:
Some tools:
Interviews of Amazon staff about their use of automated reasoning:
AWS Lectures aimed at customers and industry:
AWS talks aimed at the automated-reasoning science community:
AWS blog posts and informational videos:
Some course notes by Amazon Scholars who are also university professors:
A fun deep track:
Some algorithms found in the automated theorem provers we use today date as far back as 1959, when Hao Wang used automated reasoning to prove the theorems from Principia Mathematica.

    



                            
                        

                    
                


                
                    

                        

    

    

        Research areas
    


    
        
    



                    

                

                
                    

                        

    

    

        Tags
    


    
        
    



                    

                

                

                
                    
About the Author

                    

                        
                            
    

        

            
                

                    Byron Cook
                

            
            
            
                

                    Amazon vice president and distinguished scientist Byron Cook is a leader in the field of formal verification, known for his contributions to SAT, SMT, and symbolic model checking, with applications to biological systems, computer operating systems, programming languages, and security. Byron’s work on automated reasoning at Amazon has led to higher levels of assurance in the cloud as well as new customer features.
                

            
        

    



                        
                    

                
            

        


        
    

    
    



    

        
Related content

        
    




    
        
    

    



    
    



    

        
Work with us
See more jobs
        
            See more jobs
        
    




    
        

            

                
                    

                        

    

    

        
            

                 Applied Scientist, Fauna Robotics
            

        

        

            
                
                    

                
            

            

            

            

            
                
US, CA, Sunnyvale

            
        


        
    
Robots are about to step out of the movies and into our homes. If you're excited about this robotics revolution, come join Amazon's Fauna Robotics. As an Applied Scientist in this org, you'll push the state-of-the-art techniques in Robotics. You'll drive technical excellence in areas such as perception, manipulation, sim2real transfer, reinforcement learning and multi-task learning, designing novel algorithms that bridge the gap between research and real-world deployment. In this role, you will integrate hands-on technical expertise with scientific leadership, ensuring your team delivers robust solutions for dynamic real-world environments. While staying updated with the latest advancements in robotics research, you will leverage your experience and expertise to guide the team towards the most promising direction. Collaborating with a talented group of software engineers and scientists, you will bring your ideas to fruition. Deploying robots in home environments requires you to anticipate challenges not encountered in structured settings like warehouses. Additionally, consumer-grade sensors and actuators necessitate working around their limitations, requiring out-of-the-box innovative ideas. As such, you are someone who truly enjoys thinking about problems using first principles at a holistic and systems level to address the aforementioned challenges. Key job responsibilities - Lead technical initiatives in areas such as robotics foundation models, reinforcement learning and manipulation - - Design experiments to identify the limitations of current state-of-the-art models and developing new models or techniques that can surpass these models - Design and implement novel deep learning architectures that push the boundaries of what robots can understand and accomplish - Mentor fellow scientists while maintaining strong individual technical contributions - Collaborate with engineering teams to optimize and scale models for real-world applications - Influence technical decisions and implementation strategies within your area of focus - Experienced in communicating complex technical work to a non-technical audience - The ability to work with minimal guidance, be proactive and to handle ambiguity and the challenge of quickly evolving goals About the team Fauna Robotics, an Amazon company, is building capable, safe, and genuinely delightful robots for everyday life. Our goal is simple: make robots people actually want to live and interact with in everyday human spaces. We believe that future won’t arrive until building for robotics becomes far more accessible. Today, too much effort is spent reinventing the fundamentals. We’re changing that by developing tightly integrated hardware and software systems that make it faster, safer, and more intuitive to create real-world robotic products. Our work spans the full stack: mechanical design, control systems, dynamic modeling, and intelligent software. The focus is not just functionality, but experience. We’re building robots that feel responsive, expressive, and genuinely useful. At Fauna, you’ll work at the frontier of this space, helping define how robots move, manipulate, and interact with people in natural environments. It’s an opportunity to solve hard problems across hardware and software with a team focused on making robotics accessible and joyful to build. If you care about making robotics real for everyone and building systems that are as delightful as they are capable, we’re interested in hearing from you.




        

        

        

        
    


    



                    

                
                    

                        

    

    

        
            

                Applied Scientist II, Sponsored Products and Brands
            

        

        

            
                
                    

                
            

            

            

            

            
                
US, NY, New York

            
        


        
    
The Sponsored Products and Brands team at Amazon Ads is re-imagining the advertising landscape through industry leading generative AI technologies, revolutionizing how millions of customers discover products and engage with brands across Amazon.com and beyond. We are at the forefront of re-inventing advertising experiences, bridging human creativity with artificial intelligence to transform every aspect of the advertising lifecycle from ad creation and optimization to performance analysis and customer insights. We are a passionate group of innovators dedicated to developing responsible and intelligent AI technologies that balance the needs of advertisers, enhance the shopping experience, and strengthen the marketplace. If you're energized by solving complex challenges and pushing the boundaries of what's possible with AI, join us in shaping the future of advertising. As an Applied Scientist within the Brand Intelligence team, you will be responsible for improving the quality of underlying agentic solutions and their impact within the ads delivery stack. We are looking for a passionate candidate with technical expertise in agentic systems, continual learning, evaluation, information retrieval, Natural Language Processing (NLP), and Large Language Models (LLM). In addition to having hands-on experience in building ML-based solutions, an ideal candidate should be able to create and articulate a customer-centric science vision, show willingness to continuously learn about new scientific approaches, and enjoy operating in startup-like environment.




        

        

        

        
    


    



                    

                
                    

                        

    

    

        
            

                Principal Applied Scientist, Fauna
            

        

        

            
                
                    

                
            

            

            

            

            
                
US, NY, New York

            
        


        
    
Fauna Robotics is building capable, safe, and delightful robots for everyday life, and understanding where it is, and what's around it, is fundamental to everything a robot does. A robot that lives alongside people must build and maintain a map of its environment, plan paths through cluttered and changing spaces, and do so reliably whether it's navigating a hallway for the hundredth time or entering a room it has never seen. We are hiring a Principal Applied Scientist to serve as a technical authority on how our robots understand and move through the world. You will design, prototype, and ship the core algorithms that give the robot spatial intelligence - from real-time SLAM and scene understanding to motion planning and autonomous navigation. You will set the mapping and navigation architecture that other engineers build on, shape sensor and compute decisions across robot generations, mentor the engineers and scientists working on spatial perception, and be the person teams come to when the robot is lost. Key job responsibilities - Set the long-range science roadmap and technical architecture for the robot's mapping, localization, and navigation stack, and serve as Fauna's primary technical authority on spatial intelligence - Design and implement real-time visual-inertial and/or lidar-based SLAM systems that build, update, and maintain persistent maps of home and indoor environments - Develop robust localization algorithms that allow the robot to determine its precise pose in previously mapped spaces, handling environmental changes such as moved furniture, lighting variation, and dynamic occupants - Build navigation and motion planning systems that produce safe, efficient, and natural-looking trajectories through cluttered, dynamic human spaces - Architect semantic mapping capabilities that go beyond geometry — enabling the robot to understand rooms, objects, surfaces, and traversable space at a level that supports higher-level task planning - Drive sensor selection and placement decisions for depth cameras, lidar, IMUs, and encoders in collaboration with mechanical, electrical, and industrial design teams, backed by your own benchmarking and analysis - Design the robot-specific data collection, benchmarking, and evaluation methods to validate mapping accuracy, localization robustness, and navigation safety across diverse real-world environments - Develop strategies for graceful degradation and recovery — ensuring the robot behaves safely when localization is uncertain, maps are incomplete, or sensors are occluded - Present mapping and navigation science and its tradeoffs to senior leadership and partner teams - Mentor scientists and engineers, raising the scientific bar for spatial perception across the organization through design reviews, code reviews, and hiring About the team Fauna Robotics, an Amazon company, is building capable, safe, and genuinely delightful robots for everyday life. Our goal is simple: make robots people actually want to live and interact with in everyday human spaces. We believe that future won't arrive until building for robotics becomes far more accessible. Today, too much effort is spent reinventing the fundamentals. We're changing that by developing tightly integrated hardware and software systems that make it faster, safer, and more intuitive to create real-world robotic products. Our work spans the full stack: mechanical design, control systems, dynamic modeling, and intelligent software. The focus is not just functionality, but experience. We're building robots that feel responsive, expressive, and genuinely useful. At Fauna, you'll work at the frontier of this space, helping define how robots move, navigate, and interact with people in natural environments. It's an opportunity to solve hard problems across hardware and software with a team focused on making robotics accessible and joyful to build. If you care about making robotics real for everyone and building systems that are as delightful as they are capable, we're interested in hearing from you.




        

        

        

        
    


    



                    

                
                    

                        

    

    

        
            

                Research Scientist II, Global Risk Intelligence and Prevention
            

        

        

            
                
                    

                
            

            

            

            

            
                
US, WA, Seattle

            
        


        
    
We are seeking an exceptional Applied Scientist, Seller Abuse Prevention, to lead the development and implementation of advanced AI solutions that will transform how we prevent bad actors from operating in our store and enable Selling Partners to start and grow their business without fear of disruption, so that customers and Selling Partners across the globe trust us and have confidence in the integrity of Amazon's store. This role will focus on building risk detection models leveraging state-of-the-art AI, including small language models, to detect and prevent abuse of Amazon's catalog worldwide. You will design, develop, and deploy scalable AI solutions to proactively detect and prevent marketplace abuse throughout the seller lifecycle. You will work with massive-scale, multi-modal datasets spanning behavioral patterns, transactional histories, and behavioral data to build detection systems that are ahead of evolving adversarial tactics. Key job responsibilities - Design and build predictive risk detection models using advanced AI techniques, including Natural Language Processing including LLMs and agents to proactively identify bad actors and prevent marketplace abuse at scale - Own the end-to-end scientific solution from risk quantification through decision optimization, determining the appropriate actions to take across varying risk levels - Develop interpretability and reasoning pipelines that provide transparent, actionable explanations for model decisions to support enforcement and seller experience - Work with risk programs across the seller lifecycle to define detection strategies, translate operational investigation patterns into automated systems, and prioritize high-impact risk areas - Partner with engineering teams to deploy models into production, define evaluation frameworks, and collaborate with operations and verification teams to measure and improve detection effectiveness A day in the life Day-to-day you can expect to: - Explore datasets to understand predictors and patterns of abuse - Work with product, program, and engineering stakeholders to build solutions into production that will last - Identify new and emerging abuse vectors as abusers get more sophisticated - Use search, graph, computer vision, NLP, and anomaly detection methodologies to automatically detect abusive actions. About the team Our team is driven by the belief that scientific innovation unlocks transformative customer experiences. We tackle hard, ambiguous problems by combining rigorous research methods with practical engineering, and we value collaboration across disciplines. We are building an inclusive environment where scientists at every stage of their career can grow, publish impactful work, and see their research come to life at Amazon scale. If you want to do science that matters, this is the place.




        

        

        

        
    


    



                    

                
                    

                        

    

    

        
            

                AI Engineer - RL Manipulation, RIVR
            

        

        

            
                
                    

                
            

            

            

            

            
                
CH, Zurich

            
        


        
    
Are you passionate about pushing the boundaries of what robots can do? Amazon is looking for a creative and driven Applied Scientist to help build intelligent robotic manipulation systems powered by reinforcement learning. In this role, you will research, design, and deploy machine learning solutions that enable robots to interact with the physical world in increasingly capable ways. You will tackle complex, ambiguous problems where textbook answers don't exist, working backwards from real-world needs to invent new approaches. Your work will directly influence how automation systems learn, adapt, and perform at Amazon's scale — and you will see your research come to life on real hardware. Key job responsibilities - Research and develop reinforcement learning methods for robotic manipulation, advancing the state of the art in sim-to-real transfer, contact-rich control, and adaptive grasping strategies. - Design, implement, and evaluate machine learning models and algorithms, leveraging deep learning frameworks such as PyTorch or TensorFlow to train policies that generalize across varied manipulation tasks. - Build and maintain production-quality code that integrates scientific prototypes into larger robotics systems, making thoughtful tradeoffs between complexity, performance, and maintainability. - Collaborate with cross-functional engineering and science teams to define simulation requirements, validate model performance, and iterate on solutions informed by real-world hardware feedback. - Contribute to the team's scientific roadmap by identifying new research directions, mentoring fellow scientists, and publishing results at peer-reviewed venues when appropriate. A day in the life You might start your morning reviewing overnight training runs, analyzing reward curves and failure modes from your latest reinforcement learning experiments. After a quick sync with robotics engineers, you shift to refining a simulation environment to close a gap you spotted between sim and real performance. In the afternoon, you pair with a teammate to design a new evaluation framework, then wrap up by drafting a short write-up of your findings to share with the broader team. About the team Our team sits at the intersection of machine learning research and real-world robotics, focused on teaching robots to manipulate objects with precision and adaptability. We combine reinforcement learning, physics-based simulation, and hands-on hardware validation to deliver solutions that work beyond the lab. We value intellectual curiosity, collaborative problem-solving, and a shared commitment to turning ambitious research ideas into tangible impact. If you are excited about building intelligent systems that learn from experience and operate at scale, this is the team for you.




        

        

        

        
    


    



                    

                
                    

                        

    

    

        
            

                Applied Scientist, AGI
            

        

        

            
                
                    

                
            

            

            

            

            
                
US, CA, Sunnyvale

            
        


        
    
Amazon is looking for an Applied Scientist with a strong deep learning background to help build the training environments that teach large language models to reason, use tools, and complete complex, multi-step tasks. Key job responsibilities You will design, build, and evaluate reinforcement learning environments (RL gyms) for LLM training: realistic, verifiable task environments where models learn through interaction and feedback. This includes defining tasks and reward signals, building the infrastructure to run environments at scale, measuring how well environments transfer into model capability, and developing rigorous methods to verify environment quality and correctness. You will work closely with model training teams to understand where environments are most valuable and iterate based on what the models learn. Your work will directly shape products and services that use generative AI. About the team The Foundational AI (FAI) team has a mission to push the envelope in GenAI with LLMs, in order to provide the best-possible experience for our customers.




        

        

        

        
    


    



                    

                
                    

                        

    

    

        
            

                Economist, Entertainment Economics
            

        

        

            
                
                    

                
            

            

            

            

            
                
US, WA, Seattle

            
        


        
    
The charter of the Entertainment Economics team includes working on high-impact projects that require economics expertise and a cross-business perspective. Our team supports Amazon’s diverse entertainment businesses – especially Prime Video, Amazon Music, Amazon MGM Studios, and Gaming (Luna) – through complex causal measurements (observational and experimental) and structural modeling. Some of our team’s areas of science expertise include marketing science, long-term value measurement, and experimentation. As a central economics team not embedded in any particular business, we strive to provide objective analyses that take a customer-obsessed, pan-Amazon perspective. This role partners with the Prime Science team to evaluate and improve the way Amazon allocates Prime subscription revenue across its benefit teams, including entertainment businesses. Prime is Amazon's premiere membership program, with over 200mm members world-wide, bringing in billions in Prime subscription revenue annually. Prime is at the center of many major company decisions that impact Amazon's customers. These decisions span a variety of industries, each reflecting the diversity of Prime benefits: fast and free shipping, entertainment & digital content, reading, healthcare, and grocery. Each of these Prime benefits is supported by a different organization, and each relies on Prime subscription revenue to make critical investment decisions. Our team partners with Prime Science to ensure that these allocations are credible, actionable, and incentive compatible for every benefit. To support this work, this role will create science models, validate them, and productionize them. In addition to stellar economics expertise, this role requires systems thinking (anticipating how complex interconnected models will respond to changes in data or modeling), strategic thinking (supporting decisions that affect 200 million+ customers and set Amazon’s direction for years to come), and scientifically novel thinking (using AI and the latest science advancements to achieve move). The ideal candidate is expert in causal inference with experience working in large, complex structural modeling environments. They are a creative problem-solver, able to identify new approaches that solve multiple problems at once while minimizing churn. They think big, not limiting themselves to what’s been done before, while tactfully accounting for the costs of major changes. They can add structure to incredibly ambiguous business contexts, making defensible assumptions and delivering measurements that drive impactful business decisions. They have effective communication and the skill to translate data-driven findings into actionable insights for senior leaders. Finally, they are empathetic and constructive. Our healthy team culture is supportive and fast-paced, and we prioritize learning, growth, and helping each other to continuously raise the bar. Impact and Career Growth In today’s entertainment landscape, critical decisions are made with data and economic models. You’ll help leaders ask the right questions, and then deliver data-driven answers, helping to create the future of Amazon. You’ll help define a long-term science vision in this space and translate it into an actionable roadmap. This role combines science leadership, organizational ability, technical strength, product focus, and business understanding – a perfect recipe for career growth as an economist in tech. Key job responsibilities • Design and build econometric models, especially causal models, to measure the value of the business and its many features • Develop science products from concept to prototype to production, incorporating feedback from scientists and business partners • Independently identify and pursue new opportunities to leverage economic insights across entertainment and Prime • Write business and technical documents communicating business context, methods, and results to business leadership and other scientists • Serve as a technical reviewer for our team and related teams, including document and code reviews




        

        

        

        
    


    



                    

                
                    

                        

    

    

        
            

                Applied Scientist, AWS Science of Security, AWS Provable Security
            

        

        

            
                
                    

                
            

            

            

            

            
                
US, NY, New York

            
        


        
    
Application deadline: Nov 8, 2026 AWS Science of Security is dedicated to making AWS the best computing service in the world for customers who require advanced and rigorous solutions for security, privacy, and sovereignty. The team sits within Amazon Security and operates in a fast-growing, startup-like environment within AWS. You will participate in the design, development, evaluation, and deployment of automated reasoning solutions for verifying the correctness, security, and reliability of software systems. You’ll be applying techniques such as constraint solving, model checking, static analysis, theorem proving, and program analysis to unique applications in business and engineering. How you build and deploy automated reasoning tools and verification frameworks, plus research and implement novel approaches, will add distinct value to the business and our customers. This is work on open, hard problems: the kind where the tools you build get used to secure systems that millions of customers depend on. You will dive deep alongside our world-renowned automated reasoning scientist community, bring fresh ideas from your research experience, and see your work ship rather than sit on a shelf. This is an exceptional opportunity to join a fast moving segment of AWS and help shape the technology, product and business, with notable customer and business impact. Key job responsibilities - Solve significantly complex problems that require deep knowledge and understanding of your domain and scientific innovation - Own the design, implementation, and delivery of solutions that have a long-term quantifiable impact - Provide cross-organizational technical influence, driving effective outcomes by sharing your deep knowledge and experience - Develop fundamentally new solutions for uniquely challenging business problems A day in the life This is a unique opportunity to get in early on a fast-growing segment of AWS and help shape the technology, product and the business. You will have a chance to utilize your deep technical experience within a fast moving, start-up environment and make a meaningful business and customer impact. About the team Amazon Security values diverse experiences. Even if you do not meet all of the qualifications and skills listed in the job description, we encourage you to apply. If your career is just starting, hasn’t followed a traditional path, or includes alternative experiences, don’t let it stop you from exploring what's next.




        

        

        

        
    


    



                    

                
                    

                        

    

    

        
            

                Principal Applied Scientist, Items and Relationships Platform
            

        

        

            
                
                    

                
            

            

            

            

            
                
US, WA, Seattle

            
        


        
    
Interested in setting the scientific vision for how every product at Amazon is understood, identified, and connected, with the opportunity to impact every Amazon customer and every selling partner? The Items and Relationships Platform is the foundation that establishes what every product in Amazon’s catalog is and how products relate to one another. We are rebuilding the platform to establish identity and relationships from meaning and context, using multimodal large language models and generative AI. This is a foundational problem: understanding products deeply enough to identify them, relate them, and keep the catalog trustworthy as it scales to the full breadth of Amazon’s global contribution traffic. A Principal Applied Scientist on the Items and Relationships Platform owns the science direction for product identity and product relationships. It spans how contributions are resolved into a committed product identity, how products are connected into structural relationships (variations, multipacks, newer versions, lines of products, bundles), how confident those decisions are, and how their quality is measured. You will own the strategy for multimodal understanding, representation learning, relationship inference, calibration, and evaluation, so that science capability compounds across identity and relationships rather than being built one use case at a time. This role lives at the intersection of research and production. Product identity and relationships are fundamentally an entity resolution problem: deciding what is the same product, and how different products relate, across diverse sources, languages, and formats. You will make the hard modeling calls on where sameness ends and relatedness begins, how models express and calibrate confidence, and how the platform measures quality when there is no external ground truth. Successful candidates are passionate about translating ambiguous business goals into a coherent science roadmap that enables many businesses at once. You are comfortable being the final scientific arbiter on hard cross-team decisions, and working alongside top-notch applied scientists and software engineers to turn advanced research into production systems. Key job responsibilities As a Principal Applied Scientist on the Items and Relationships Platform, you will: - Own the science direction for product identity: define how products are semantically understood, resolved, and committed as coherent identities, navigating the precision/recall and friction/selection trade-offs inherent in every decision - Build representations that decide both sameness and typed relatedness, and drive relationship inference from semantic identity rather than one pipeline per relationship type - Own structural relationship science across variations, multipacks and base products, newer versions, lines of products, and bundles, and discover new relationship types from the gap between defined policy and what identity reveals - Own calibration and evaluation science: define how models emit calibrated confidence, and build the LLM judge methodology and Catalog Identity Quality metric that measure identity and relationship quality at catalog scale - Partner with Principal Engineers on cost-aware modeling (cascades, distillation, routing) and close the loop from production signals and corrections back into training - Set science direction partnering with ACCS Applied Science teams, mentoring senior and principal-track scientists, and driving the documentation and publication of scientific innovation A day in the life On a typical day, you move between setting science direction and going deep on hard problems: reviewing a multimodal identity model or a relationship inference experiment with scientists, deciding how a borderline case (a multipack, a newer version, a near-identical variant) should be modeled, and pressure-testing whether a quality gain is real by examining the judge and its calibration. You also spend time zoomed out, shaping the longer-term science roadmap, aligning with Principal Engineers and partner teams on the identity and relationship signals they consume, and mentoring scientists as they take on the ambiguous, high-impact problems that define the platform’s future. About the team The Items and Relationships Platform advances state-of-the-art GenAI to deeply understand, uniquely identify, and meaningfully connect every product at Amazon scale. We build the canonical product identity, matching, and relationship systems that power Amazon’s universal catalog: eliminating duplicates, preventing abuse, keeping product families coherent, and integrating the world’s product data to power agentic shopping and trusted discovery for customers worldwide. We solve foundational challenges including schema-agnostic multimodal understanding, relationship inference, evaluation without ground truth, and efficient inference at reduced cost.




        

        

        

        
    


    



                    

                
                    

                        

    

    

        
            

                Applied Scientist, Mobile Manipulation Robotics (I/O)
            

        

        

            
                
                    

                
            

            

            

            

            
                
US, MA, North Reading

            
        


        
    
Amazon Robotics is transforming warehouse automation through edge AI and machine learning applied to real-world robotics challenges. We're seeking a Applied Scientist to advance our mobile manipulation capabilities by developing learning-based approaches that enable robots to navigate and manipulate objects in dynamic fulfillment environments. This role offers the opportunity to apply state-of-the-art research to production systems operating at Amazon's unprecedented scale. Key job responsibilities - Model Development and Training: Designing and implementing the model architectures, training and fine tuning the models using various datasets, and optimize the model performance through iterative experiments - Data Management: Process and prepare training data, including data governance, provenance tracking, data quality checks and creating reusable data pipelines. - Experimentation and Validation: Design and execute experiments to test model capabilities on the simulator and on the embodiment, validate performance across different scenarios, create a baseline and iteratively improve model performance. - Code Development: Write clean, maintainable, well commented and documented code, contribute to training infrastructure, create tools for model evaluation and testing, and implement necessary APIs - Research: Stay current with latest developments in foundation models and robotics, assist in literature reviews and research documentation, prepare technical reports and presentations, and contribute to research discussions and brainstorming sessions. - Collaboration: Work closely with senior scientists, engineers, and leaders across multiple teams, participate in knowledge sharing, support integration efforts with robotics hardware teams, and help document best practices and methodologies. 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: 1. Medical, Dental, and Vision Coverage 2. Maternity and Parental Leave Options 3. Paid Time Off (PTO) 4. 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 Are you inspired by invention? Is problem solving through teamwork in your DNA? Do you like the idea of seeing how your work impacts the bigger picture? Answer yes to any of these and you’ll fit right in here at Amazon Robotics. We are a smart, collaborative team of enthusiastic doers that work passionately to apply innovative advances in robotics and software to solve real-world challenges that will transform our customers’ experiences in ways we can’t even image yet. We invent new improvements every day. We are Amazon Robotics and we will give you the tools and support you need to invent with us in ways that are rewarding, fulfilling and fun!




        

        

        

        
    


    



                    

                
            

        

    

    

        See more jobs
    






        




    

        
            
        
        

            

                
                    

                        

                    

                
            

            
                

                    

                        

  

      
          

              
                  
    
        View from space of a connected network around planet Earth representing the Internet of Things.
    

              
          

      

      

        

          
              
Get more from Amazon Science

          

          
    
Subscribe to our newsletter



        

        
          
        
      

  



                    

                

            
        

    


    

        

        
            
Amazon.com | Conditions of Use | Privacy | © 1996-2026 Amazon.com, Inc. or its affiliates

        
        
            

    
Social