Social
CompressionModels-03-16x9.png
The more your listener already knows, the shorter the message you need to send. An expert ML engineer needs only a few sentences; a newcomer needs the whole manual.
Machine learning

Why don’t machine learning research agents overfit?

New research indicates that AI agents learn compressible models of data, which don’t have enough space to enable memorization.

By Martin Bertran Lopez, Aaron Roth
September 10, 2026
11 min read
Share
Key takeaways
  • ML models don't overfit benchmarks, even after many rounds of iterative improvement. This contradicts textbook predictions that repeatedly evaluating against the same held-out data should lead to overfitting.
  • Experiments with ML research agents indicate that successful strategies are highly compressible. When a successful agent's strategy is squeezed through an information bottleneck (as few as 16 tokens), a fresh agent with no memory can reproduce the original agent's performance, meaning the strategy captured real structure, not memorized data.
  • Compression provides both an explanation and a diagnostic tool. Strategies that genuinely overfit fail the compression test: their validation-specific gains vanish when passed through the bottleneck.
  • LLMs are powerful compression decoders. Because they carry vast world knowledge, they can reconstruct full ML pipelines from terse, expert-shorthand prompts, which is a concrete way of understanding why they're so capable.
Was this answer helpful?

Machine learning, at its core, is about generalization, not memorization. You hand your learning algorithm a pile of training examples and use them to fit a model. But the goal is not to perform well on the training examples — that's easy, you could just memorize the answers. The goal is to perform well on new examples that you have never before seen. If a model does well on the data it was trained on but poorly on fresh data, it hasn’t actually learned anything; you have only fooled yourself into thinking it has. This failure mode has a name: overfitting.

Anyone who has taken an introductory statistics or machine learning class knows the standard defense. You hold out some of your data and refuse to train on it. In practice, this held-out data plays two roles. A validation set is one you consult repeatedly while building the model — to compare candidates, tune hyperparameters, and decide what to try next. A final test set (or holdout) is meant to be touched only once, at the very end: because the training procedure never saw it, strong performance there is a correct proxy for the new examples you will encounter in the wild.

Machine learning, at its core, is about generalization, not memorization

The “holdout” condition is crucial, though. The correct-proxy guarantee holds if the held-out set stays genuinely unseen. If you check your performance on it, tweak your training procedure in response, recheck, and iterate, chasing better and better numbers, that set is no longer unseen; it has become part of your training procedure. Do this enough times, and you can overfit it just as you might have overfit the training set, and you have lost your proxy for unseen data. This is true of any held-out set you reuse this way, including a validation set, which is reused by design.

A puzzle at the heart of machine learning

Real machine learning research looks exactly like the iterative improvement loop we just described. Everyone gauges performance using a handful of benchmark datasets that go unrevised for years. The research community repeats an enormous, distributed loop: evaluate a model on the benchmark, revise the training procedure, re-evaluate, publish, and let the next group eke out a little more improvement.

This is precisely the kind of hill-climbing against a held-out set that, by the textbook account, ought to produce rampant overfitting. By now, the leaderboards should be saturated with models that look great on the benchmark and mediocre everywhere else.

And yet that is not what happens. Studies that build entirely fresh test sets for old, heavily reused benchmarks have found that improvements largely transfer: on the new data, models demonstrate the same gains they did on the old benchmark. Benchmark-driven machine learning, against the textbook's prediction, has produced rapid and largely real progress. Why?

There is no shortage of hypotheses, but they have been hard to test empirically, because the "subject" of the experiment is the entire human research community. You cannot reset a field, wipe its memory, and rerun the last decade under controlled conditions.

But we can do something similar. We now have capable, LLM-based research agents that can autonomously run the same machine-learning optimization loops that human communities run. They engage in the same benchmark hill-climbing — and, intriguingly, they too seem not to overfit. The difference is that an agent, unlike a research community, is something you can reset. You can clear its memory, control exactly what information it sees, and run the experiment again. In a recent paper, "What fits (into few tokens) doesn't overfit: Compression and generalization in ML research agents", we do exactly that — and in the process offer a concrete explanation for the long-standing mystery.

Occam's razor, made precise

The explanation begins with a very old idea. Occam's razor says that among hypotheses that explain the data equally well, the simpler one is more likely to be correct. It turns out this intuition has a precise mathematical form, and it is what underlies the whole story.

Suppose you can describe your hypothesis — your model, your strategy — in a small number of bits, far fewer than it would take to memorize the training data. If that compact hypothesis performs very well on the training data, it must also perform well on new data.

Occam's razor, formalized: among hypotheses that explain the data equally well, the simpler one — describable in fewer bits — is more likely to generalize to new examples.

The reasoning runs through a counting argument. There simply are not very many short descriptions, because there are not very many short strings. The fewer candidate hypotheses there are, the less likely it is that any one of them fooled you on the training set by luck — even though you used the training set to guide your search.

Another way to get the intuition: if your compressed description is too small to secretly record the training data, then when it performs well on the training data, it cannot be because it memorized the answers — it didn't have space to do that. It must be because it captured something true about the data's structure. Short descriptions cannot cheat because there isn't room.

Here is an attractive hypothesis: successful machine learning strategies are highly compressible. A researcher might stare at thousands of benchmark scores over the course of a project, but the strategy that ultimately survives is usually a short list of familiar choices — an architecture family, an optimizer, a learning-rate schedule, a data-handling recipe, a regularization scheme. If that final recipe can be communicated in just a few bits, then the model's true dependence on the benchmark is far smaller than the long, winding transcript of experiments would suggest. The hill-climbing was extensive, but the thing that came out the other end was — or could have been — tiny.

Compression, intelligence, and the power of a knowledgeable listener

Imagine trying to explain a specific machine learning pipeline to a bright high-school student, in enough detail that they could actually reproduce it. It would be a long, laborious conversation. You would have to explain what gradient descent is, what a neural network is, what PyTorch or JAX or TensorFlow does, what a learning rate is, and on and on. Almost none of that is specific to your problem; it is general background about how machine learning works.

Now imagine explaining the same pipeline to an expert ML engineer. The conversation now collapses to a few sentences. You skip everything that counts as common knowledge and communicate only what is genuinely specific to this problem: the architecture choice, the batch size, the optimizer, a couple of hyperparameters. The more your listener already knows about the world, the shorter the message you need to send — and the more aggressively you can compress. None of this "world knowledge" counts against you in the Occam's-razor argument, because you could have written all of that down without having looked at the training set.

This is where large language models enter the picture. Modern LLMs carry an enormous amount of world knowledge. They know how ML tooling works; they know the standard optimization algorithms; they know the conventional hyperparameter choices and the common defaults. If a detail is left unspecified, they can fill in a plausible value. That makes them extraordinarily good compression decoders: hand an LLM a terse, expert-to-expert message, and it can unpack it into a full, working procedure. If you think about it, this is exactly why they are so powerful.

The experiment: Squeezing a strategy through a bottleneck

This suggests a clean experiment. Have an ML research agent — the explorer — try to solve a new machine learning problem. Give it full access to a validation set and let it experiment and iterate freely, chasing better validation performance over hundreds of rounds. Here the validation set plays the role of the benchmark: a reusable holdout the agent queries again and again. This is the hill-climbing loop that ought to overfit.

Then test how compressible the solution is. A second agent, the compressor, reads the entire transcript of the explorer's work and tries to distill the winning strategy into a very short prompt — just a handful of tokens. That prompt is handed to a third agent, the reproducer, which must implement the strategy from scratch using only the prompt and the training data. Critically, the reproducer has no access to the validation set, the explorer's code, or its transcript. The short prompt is the only channel through which anything learned from the validation set can reach it. (In the study we report in our paper, the compressor and reproducer are both Claude models.)

If the reproducer — starting cold, armed only with a few tokens — matches the explorer's performance, then all the validation-dependent information needed to specify the strategy fit through that tiny channel. The strategy was compressible. We call this a certificate of output compression.

The setup has a very useful property that human research communities lack: the reproducer can be reset over and over. The compressor can try many different compressions and see how well each is decoded, because every attempt lands on a fresh reproducer with no memory of the last one. It is a little like the film Memento — you are leaving a terse note for a version of yourself whose memory will be wiped before reading it. You learn to write notes that a knowledgeable but amnesiac copy of you can act on; those notes can be very short because the receiver will fill in anything you leave unsaid exactly as you would have.

CompressionModels-04-1x1.png
In the researchers' experiments, an explorer agent's strategy is squeezed through a narrow information bottleneck. Whatever survives compression must reflect real structure, not memorized data.

What comes out the other end

The compressions turn out to be remarkably small. Across eight datasets — spanning tabular classification, image classification, language modeling, diffusion modeling, and reward modeling — 32-token prompts were enough for a fresh reproducer to match the explorer's adaptively optimized models on the large majority of problems. One language-modeling strategy survived compression down to just 16 tokens with no loss in held-out performance.

What do these prompts actually look like? The most revealing examples are right at the border of conciseness where the compression almost breaks. In one language-modeling experiment, the explorer discovered a custom GPT-style training recipe. Under a 16-token budget, this was still enough for fresh reproducers to match the uncompressed explorer:

QKn 12L768 Mu .1 R² b2M 4x

To a human reader this looks cryptic, but to another ML agent it says something concrete: QKn means “QK normalization”, 12L768 means a 12-layer, 768-dimensional transformer, Mu .1 means the Muon optimizer with learning rate 0.1, R² means squared-ReLU activations, b2M means a two-million-token batch, and 4x means a fourfold feed-forward block. Cut the budget to eight tokens, however, and the prompt becomes

12L768 Mu .1 R²

Now the reproducer no longer matches the explorer. The missing pieces specified real training choices that were made as a function of the data and differ from the most obvious defaults. This boundary shows the limits of compressibility and is important. It shows that the reproducer is not succeeding from prior knowledge alone. A few compressed tokens are carrying genuine information learned from the data itself, and when those tokens disappear, so does the performance.

We also ran a set of experiments that imposed an information bottleneck from the other direction. Instead of compressing the explorer's output, we compressed its input: rather than telling the explorer each model's numerical validation score, we returned only a single bit — did this model beat the running best, or not? Even reduced to one bit of feedback per query, the explorer found strategies as good as those it found with full numerical scores. The channel between the validation set and the final strategy is narrow in both directions, and the one-bit version even comes with a rigorous mathematical guarantee on generalization.

Across eight datasets, strategies that emerged from hundreds of iterative experiments could be compressed into prompts as short as 16 to 32 tokens — small enough for a fresh agent with no memory to reproduce the original results.

Catching cheaters

A good empirical theory should be falsifiable — and this one is. If low overfitting is really explained by compressibility, then models that genuinely overfit should fail to be compressible via this pipeline.

To check, we deliberately pushed agents into overfitting by handing them direct validation-set access and prompting them to maximize validation performance at any cost. The agents took the bait: in 38 of 102 experimental runs, validation accuracy ran more than 10% ahead of true held-out accuracy.

The theory predicts that these gains should not survive the compression bottleneck, because they encode idiosyncrasies of specific validation examples, not transferable structure. Sure enough, when squeezed through a short prompt to a fresh reproducer, the validation-specific advantages vanished. Compression separated the legitimate strategies from the overfitting ones with very high accuracy.

So compression does not merely explain why autonomous research agents tend not to overfit but offers a tool for catching overfitting when it does occur, by flagging the cases where no short description can reproduce the result.

What this tells us — and what it doesn't

A few caveats are in order. The whole framework assumes that the only path from the validation data to the final model runs through the prompt we feed the reproducer. Of course, if a model had memorized the validation data during pretraining, it would have a side channel that bypasses the information bottleneck we are trying to impose. We don't think that is what is happening in our experiments: agents improve gradually through real search rather than starting at their best, and performance degrades at very short token budgets. But fully resolving this question will likely require experimenting with fresh datasets collected after a model's training cutoff, which we haven’t done.

Most importantly, our results are about LLM agents, because that is where the experiment is possible — where you can reset the subject, control its inputs, and count their length. But the picture they paint is strongly suggestive about human research communities too. When a field spends years climbing a fixed benchmark, and the gains keep transferring to fresh data, it may be for the same reason the agents' strategies survive a 32-token prompt: the recipes that actually work are simple. Or in other words, "What fits (into few tokens) doesn't overfit."

Acknowledgments: Steven Wu

Research areas
Tags
About the Author
Martin Bertran Lopez
Martin Bertran Lopez is an applied scientist in Amazon Web Services' Privacy organization.
Aaron Roth
Aaron Roth is the Henry Salvatori Professor of Computer and Cognitive Science at the University of Pennsylvania and an Amazon Scholar. His research focuses on the algorithmic foundations of data privacy, algorithmic fairness, game theory, learning theory, and machine learning. Together with Cynthia Dwork, he is the author of the book The Algorithmic Foundations of Differential Privacy; together with Michael Kearns, he is the author of The Ethical Algorithm: The Science of Socially Aware Algorithm Design.

Related content

Work with us

See more jobs See more jobs
Applied Scientist, Pricing Science
US, WA, Seattle
Pricing is one of the most consequential decisions Amazon makes — and the science behind it needs to be causally rigorous, not just predictive. The P2 Optimization Science (P2OS) team builds the machine learning systems that power Amazon's pricing decisions at scale: demand lift models, customer lifetime value frameworks, and the experimentation infrastructure that validates whether our pricing changes actually work. We're hiring an Applied Scientist to own causal inference at the intersection of ML and pricing experimentation. This role exists because our team has identified a real gap: the methodological bridge between econometric analysis (owned by our economists) and production-scale ML pipelines (owned by our engineers) needs a practitioner who lives in both worlds. You'll build CATE estimation models, design analysis workflows for pricing weblabs, and develop the reusable causal ML infrastructure that the broader team — including non-ML scientists — can rely on. This is not a research role. The bias here is toward shipping production-quality causal pipelines with real downstream business impact. You'll measure success by what changes in LTV estimates, what pricing errors your models help avoid, and whether the economists on your team can actually use what you build. If you're a scientist who wants to work on hard causal identification problems in a high-stakes production environment — and who finds satisfaction in making rigorous methods accessible to a broader team — this role is for you. Key job responsibilities * Build causal ML pipelines for pricing — Design, train, evaluate, and deploy end-to-end causal estimation models for pricing use cases. * Own the science on heterogeneous treatment effects — Be the team SME on causal ML methodology: identification strategies, model selection, evaluation standards, and the tradeoffs between econometric and ML approaches to causal estimation. * Support pricing experiment analysis — Contribute causal analysis methodology to pricing weblab and A/B test post-analysis; build reusable tooling that economists can use without requiring ML expertise * Connect model outputs to business outcomes — Define, before writing code, what business metric each model moves; deliver model evaluation reports framed around pricing errors avoided and LTV estimate changes. * Evaluate and adopt novel techniques — Assess applicability of emerging causal inference methods (synthetic DiD, generalized random forests, causal representation learning) to Amazon's pricing context; write internal methodology proposals for adoption * Write internal documentation and methodology papers — Produce at least one internal write-up per half that connects a causal ML technique to a concrete pricing use case; make pipelines extensible and well-documented so other scientists can build on them. * Collaborate across disciplines — Partner closely with the Sr. Economist on identification strategy and causal assumptions; work with SDE and DE partners on production deployment; align with PMs on experiment design requirements A day in the life As an Applied Scientist on the P2OS team, your work directly shapes the prices customers see on hundreds of millions of Amazon products. In a given workweek, you might: * Investigate an optimization anomaly in simulation and trace it back to a model input gap or an unmodeled market dynamic * Design an offline evaluation framework to benchmark competing optimization approaches before committing to online testing * Collaborate with Sr. Economists on the identification strategy for the model you're building for a pricing lab * Present a science proposal for incorporating a new competitiveness or inventory signal into an optimization system * Work cross-team with the experimentation platform team on randomization design. * Develop and write up a novel scientific finding — preparing a paper or technical report for submission to a top-tier venue such as KDD, NeurIPS, or the ACM Conference on Economics and Computation
Applied Scientist, Fauna
US, CA, Sunnyvale
We are seeking an Applied Scientist to focus on Robot Navigation. In this role, you'll research and develop advanced navigation systems that enable robots to move reliably and safely through complex, dynamic environments. You'll work across a broad spectrum of navigation approaches—from classical methods to learning-based techniques and foundation models—to build robust solutions for autonomous robot navigation. Key job responsibilities - Develop and implement robust navigation systems that enable reliable autonomous operation in complex, dynamic indoor environments with static and dynamic obstacles - Build simulation-based and on-device evaluation frameworks with comprehensive benchmarks and metrics for systematic comparison of navigation methods - Conduct sim-to-real transfer experiments, analyzing performance gaps and developing techniques to ensure reliable real-world navigation performance - Collaborate with world model, manipulation, and other teams to ensure seamless integration of navigation capabilities into the full robot system - Stay current with the latest advances in robot navigation, spatial reasoning, and related fields, and apply relevant findings to improve system performance - Mentor fellow scientists and engineers while maintaining strong individual technical contributions 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.
Applied Scientist II, Translation Services
IN, TS, Hyderabad
Have you ever wondered how Amazon launches and maintains a consistent customer experience across hundreds of countries and languages it serves its customers? Are you passionate about data and mathematics, and hope to impact the experience of millions of customers? Are you obsessed with designing simple algorithmic solutions to very challenging problems? If so, we look forward to hearing from you! At Amazon, we strive to be Earth's most customer-centric company, where both internal and external customers can find and discover anything they want in their own language of preference. Our Translations Services (TS) team plays a pivotal role in expanding the reach of our marketplace worldwide and enables thousands of developers and other stakeholders (Product Managers, Program Managers, Linguists) in developing locale specific solutions. Amazon Translations Services (TS) is seeking an Applied Scientist to be based in our Hyderabad office. As a key member of the Science and Engineering team of TS, this person will be responsible for designing algorithmic solutions based on data and mathematics for translating billions of words annually across 130+ and expanding set of locales. The successful applicant will ensure that there is minimal human touch involved in any language translation and accurate translated text is available to our worldwide customers in a streamlined and optimized manner. With access to vast amounts of data, technology, and a diverse community of talented individuals, you will have the opportunity to make a meaningful impact on the way customers and stakeholders engage with Amazon and our platform worldwide. Together, we will drive innovation, solve complex problems, and shape the future of e-commerce. Key job responsibilities * Apply your expertise in LLM models to design, develop, and implement scalable machine learning solutions that address complex language translation-related challenges in the eCommerce space. * Collaborate with cross-functional teams, including software engineers, data scientists, and product managers, to define project requirements, establish success metrics, and deliver high-quality solutions. * Conduct thorough data analysis to gain insights, identify patterns, and drive actionable recommendations that enhance seller performance and customer experiences across various international marketplaces. * Continuously explore and evaluate state-of-the-art modeling techniques and methodologies to improve the accuracy and efficiency of language translation-related systems. * Communicate complex technical concepts effectively to both technical and non-technical stakeholders, providing clear explanations and guidance on proposed solutions and their potential impact. About the team We are a start-up mindset team. As the long-term technical strategy is still taking shape, there is a lot of opportunity for this fresh Science team to innovate by leveraging Gen AI technoligies to build scalable solutions from scratch. Our Vision: Language will not stand in the way of anyone on earth using Amazon products and services. Our Mission: We are the enablers and guardians of translation for Amazon's customers. We do this by offering hands-off-the-wheel service to all Amazon teams, optimizing translation quality and speed at the lowest cost possible.
Applied Scientist, Navigation
US, CA, Sunnyvale
Amazon is on a mission to redefine the future of automation — and we're looking for exceptional talent to help lead the way. We are building the next generation of advanced robotic systems that seamlessly blend cutting-edge AI, sophisticated control systems, and novel mechanical design to create adaptable, intelligent automation solutions capable of operating safely alongside humans in dynamic, real-world environments. At Amazon, we leverage the power of machine learning, artificial intelligence, and advanced robotics to solve some of the most complex operational challenges at a scale unlike anywhere else in the world. Our fleet of robots spans hundreds of facilities globally, working in sophisticated coordination to deliver on our promise of customer excellence — and we're just getting started. As a Sr. Scientist in Robot Navigation, you will be at the forefront of this transformation — architecting and delivering navigation systems that are intelligent, safe, and scalable. You will bring deep expertise in learning-based planning and control, a strong understanding of foundation models and their application to embodied agents, and as well as have in-depth understanding of control-theoretic approaches such as model predictive control (MPC)-based trajectory planning. You will develop navigation solutions that seamlessly blend data-driven intelligence with principled control-theoretic guarantees. Our vision is bold: to build navigation systems that allow robots to move fluidly and safely through dynamic environments — understanding context, anticipating change, and adapting in real time. You will lead research that bridges the gap between cutting-edge academic advances and production grade deployment, collaborating with world-class teams pushing the boundaries of robotic autonomy, manipulation, and human-robot interaction. Join us in building the next generation of intelligent navigation systems that will define the future of autonomous robotics at scale. Key job responsibilities - Design, develop, and deploy perception algorithms for robotics systems, including object detection, segmentation, tracking, depth estimation, and scene understanding - Lead research initiatives in computer vision, sensor fusion and 3D perception - Collaborate with cross-functional teams including robotics engineers, software engineers, and product managers to define and deliver perception capabilities - Drive end-to-end ownership of ML models — from data collection and labeling strategy to training, evaluation, and deployment - Mentor junior scientists and engineers; contribute to a culture of technical excellence - Define and track key metrics to measure perception system performance in real-world environments - Publish research findings in top-tier venues (CVPR, ICCV, ECCV, ICRA, NeurIPS, etc.) and contribute to patents A day in the life - Train ML models for deployment in simulation and real-world robots, identify and document their limitations post-deployment - Drive technical discussions within your team and with key stakeholders to develop innovative solutions to address identified limitations - Actively contribute to brainstorming sessions on adjacent topics, bringing fresh perspectives that help peers grow and succeed — and in doing so, build lasting trust across the team - Mentor team members while maintaining significant hands-on contribution to technical solutions About the team Our team is a group is a diverse group of scientists and engineers passionate about building intelligent machines. We value curiosity, rigor, and a bias for action. We believe in learning from failure and iterating quickly toward solutions that matter.
Applied Scientist II, CMT
IN, KA, Bengaluru
As a member of the CMT team, you'll play a key role in the evolution of our Competitive Monitoring systems to solve significantly complex and interesting technical challenges in machine learning, large language models in production, and recommender systems to name a few. The team's work directly impacts customer experience at a worldwide scale. Key job responsibilities Thought leader on the team and help set team directions Research multiple problem domains, suggest various approaches to try and be as hands-on as needed while providing more junior scientists with critical mentorship Collaborate with engineers to come up with the right LLD and HLD to solve key business problems Strong emphasis on communication via writing, internal and external talks, and being able to align with multiple stakeholders A day in the life As an Applied scientist II, a typical day will involve aligning with key product, engineering and business stakeholders ; advising junior scientists on the work they are doing ; reading current research papers and staying up-to-date on AI research ; diving deep as needed to improve CMT models and addressing stakeholders from the science perspective ; writing python code
Applied Scientist, CMT
IN, KA, Bengaluru
As a member of the CMT team, you'll play a key role in the evolution of our Competitive Monitoring systems to solve significantly complex and interesting technical challenges in machine learning, large language models in production, and recommender systems to name a few. The team's work directly impacts customer experience at a worldwide scale. Key job responsibilities 1. Research the problem domain and come up with various approaches to solve the problem. 2. Be willing to experiment quickly and fail fast. 3. Collaborate with engineers to come up with the right end to end solution to the business problems. 4. Ideate on future roadmap for science in CMT 5. Be willing to roll up your sleeves and learn core topics outside applied science, for example ML engineering A day in the life A typical day might involve (a) working on ideas for improving models around product similarity or price recommendations, (b) working closely with other scientists and our ML engineers to ensure that the best models are in production, (c) writing good maintainable code that can be reused and reproduced, (d) sharing your work across CMT and beyond via technical writings and presentations
Applied Scientist II, Reinforcement Learning
US, MA, N.reading
Amazon is seeking exceptional talent to help develop the next generation of advanced robotics systems that will transform automation at Amazon's scale. We're building revolutionary robotic systems that combine cutting-edge AI, sophisticated control systems, and advanced mechanical design to create adaptable automation solutions capable of working safely alongside humans in dynamic environments. This is a unique opportunity to shape the future of robotics and automation at an unprecedented scale, working with world-class teams pushing the boundaries of what's possible in robotic dexterous manipulation, locomotion, and human-robot interaction. This role presents an opportunity to shape the future of robotics through innovative applications of deep learning and large language models. At Amazon we leverage advanced robotics, machine learning, and artificial intelligence to solve complex operational challenges at an unprecedented scale. Our fleet of robots operates across hundreds of facilities worldwide, working in sophisticated coordination to fulfill our mission of customer excellence. The ideal candidate will contribute to research that bridges the gap between theoretical advancement and practical implementation in robotics. You will be part of a team that's revolutionizing how robots learn, adapt, and interact with their environment. Join us in building the next generation of intelligent robotics systems that will transform the future of automation and human-robot collaboration. Key job responsibilities - Design and implement whole body control methods for balance, locomotion, and dexterous manipulation - Utilize state-of-the-art in methods in learned and model-based control - Create robust and safe behaviors for different terrains and tasks - Implement real-time controllers with stability guarantees - Collaborate effectively with multi-disciplinary teams to co-design hardware and algorithms for loco-manipulation - Mentor junior engineer and scientists
Applied Scientist-II, Amazon Music - Search Science, Amazon Music – Search Science
IN, KA, Bengaluru
Amazon Music is an immersive audio entertainment service that deepens connections between fans, artists, and creators. From personalized music playlists to exclusive podcasts, concert livestreams to artist merch, Amazon Music is innovating at some of the most exciting intersections of music and culture. The Amazon Music Search Science team is seeking an innovative and driven Applied Scientist to join our engineering and science hub in Bangalore. You will work alongside a world-class team of machine learning experts to break new ground in understanding user intent, classifying complex audio and musical forms, and creating next-generation interactive search experiences that help users find the exact music, podcasts, and audio content they are in the mood for. In this role, you will own the design, development, and deployment of end-to-end machine learning systems. You will balance execution on core search and discovery priorities—such as improving retrieval accuracy, latency, and relevance for millions of daily queries—while laying the foundational modeling capabilities for broader semantic understanding and advanced conversational search experiences across mobile, web, and voice-forward devices (like Alexa and Echo). Key job responsibilities - Core Search & Execution: Collaborate with scientists, software engineers, and product managers to define, frame, and solve complex business and ranking problems as machine learning, information retrieval, or optimization tasks. - Advanced AI & Modeling: Design, build, train, and evaluate production-grade ML models using classical machine learning, deep learning, Large Language Models (LLMs), and Agentic AI techniques to scale music discovery and intent resolution. - End-to-End Production Ownership: Take algorithms from research ideation to production deployment. Build scalable data pipelines, efficient model-serving systems, and robust offline/online evaluation frameworks. - Experimentation & Iteration: Design and analyze large-scale A/B experiments across millions of customers to measure impact on search relevance, engagement, and customer satisfaction, refining models for continuous improvement. - Forward-Looking Innovation: Research and implement novel statistical and machine learning approaches, exploring multi-modal understanding, rich content semantics, and advanced retrieval mechanisms that extend beyond traditional search boundaries. - Technical Communication: Communicate findings, architectural decisions, and technical roadmaps clearly to both technical peers and executive stakeholders, authoring robust design documents and contributing to team standards. Basic Qualifications - PhD, or Master’s degree and 4+ years of relevant experience in Computer Science, Computer Engineering, Machine Learning, Statistics, or a related quantitative field. - 3+ years of hands-on experience building machine learning models or algorithms for business applications and deploying them into production. - Strong programming skills in Python, Java, C++, or related languages, with a solid foundation in data structures, algorithms, and object-oriented design. - Experience in one or more of the following areas: Information Retrieval, Natural Language Processing (NLP), Recommender Systems, Deep Learning, or Numerical Optimization. - Demonstrated ability to work effectively with cross-functional teams in a fast-paced environment. Preferred Qualifications - Experience with large-scale distributed computing frameworks and big data systems (e.g., Spark, Hadoop, AWS infrastructure). - Experience building search ranking, query understanding, or semantic retrieval systems for high-scale consumer applications. - Familiarity with modern foundation models, LLMs, fine-tuning techniques, and efficient inference optimization for production services. - Track record of peer-reviewed publications or patents at top-tier machine learning/AI conferences (e.g., NeurIPS, KDD, ACL, SIGIR, ICML). - Experience in designing, executing, and evaluating rigorous online A/B experiments.
Senior Applied Scientist, Amazon Music - Search Science, Amazon Music – Search Science
IN, KA, Bengaluru
Amazon Music is an immersive audio entertainment service that deepens connections between fans, artists, and creators. From personalized music playlists to exclusive podcasts, concert livestreams to artist merch, Amazon Music is innovating at some of the most exciting intersections of music and culture. The Amazon Music Search Science team is looking for an execution-focused Senior Applied Scientist to spearhead core scientific initiatives within our Bangalore hub. In this leadership-by-example role, you will define and execute the applied science roadmap for search and content discovery systems, directly impacting millions of customers worldwide. You will operate at the exciting intersection of large-scale search infrastructure, applied machine learning, and foundation models. You will drive immediate, high-impact business deliverables in music search relevance, personalization, and retrieval performance, while simultaneously architecting the long-term technological vision that expands our search ecosystem toward deeper semantic intelligence, agentic workflows, and cross-domain audio understanding. Key job responsibilities - Strategic Roadmap & Architecture: Define and execute the technical and scientific roadmap for music search systems, making critical architectural decisions that balance short-term feature delivery with long-term scalability, maintainability, and extensibility. - Pioneering Applied Science: Lead the design and implementation of state-of-the-art machine learning solutions leveraging deep learning, LLMs, and agentic workflows to solve complex search, ranking, and intent-matching challenges. - Cross-Functional Leadership: Partner closely with product management, engineering leaders, and peer teams to harmonize technical direction and deliver synchronized customer experiences. - Technical Excellence & Mentorship: Drive engineering and scientific excellence across the team by conducting rigorous design reviews, establishing modeling best practices, setting high bars for artifact delivery, and mentoring junior/mid-level scientists. - Experimentation & Scaling: Establish robust scientific processes for large-scale data analysis, offline model validation, and online A/B experimentation, ensuring high statistical rigor and measurable business impact across millions of active listeners. - Stakeholder Influence & Writing: Author strategic whitepapers, and executive-level documentation. Communicate complex technical options and trade-offs to senior leadership to drive informed decision-making. Basic Qualifications - PhD, or Master’s degree and 6+ years of applied research and industrial machine learning experience in Computer Science, Machine Learning, or a related field. - 3+ years of specialized experience designing, building, and scaling machine learning models for core production business applications (e.g., Search, Recommendation Systems, or Large-Scale Information Retrieval). - Expert programming proficiency in Python, Java, C++, or related languages, combined with deep familiarity with neural network architectures and deep learning frameworks. - Proven track record of owning end-to-end technical deliverables from problem formulation and model architecture to production deployment and performance tuning. - Demonstrated leadership in mentoring technical talent and driving engineering/scientific best practices. Preferred Qualifications - Deep expertise in search retrieval, query understanding, ranking algorithms, and large-scale vector search/embedding systems. - Experience building applied science solutions on top of foundation models, large language models (LLMs), or multi-modal architectures. - Experience with large-scale distributed training and inference optimization on cloud infrastructure (AWS). - A strong publication record or patent portfolio in top-tier peer-reviewed venues (e.g., NeurIPS, SIGIR, KDD, ACL, ICML). - Experience designing and interpreting complex online experimentation frameworks for consumer-facing recommendation or search products.
Principal Applied Scientist, CX Science
US, CA, Santa Clara
We are looking for passionate, talented, and inventive Principal Applied Scientist with a strong machine learning background to help build industry-leading Conversational AI Systems. Our mission is to provide a delightful experience to Amazon’s customers by pushing the envelope in Natural Language Understanding (NLU), Dialog Systems including Generative AI with Large Language Models (LLMs) and Applied Machine Learning (ML). As part of our team, you will work alongside internationally recognized experts to develop novel algorithms and modeling techniques to advance the state-of-the-art in human language technology. Your work will directly impact millions of our customers in the form of products and services that make use language technology. You will gain hands on experience with Amazon’s heterogeneous text, structured data sources, and large-scale computing resources to accelerate advances in language understanding. We are hiring in all areas of human language technology: NLU, Dialog Management, Conversational AI, LLMs and Generative AI. A day in the life The team uses generative AI and foundation models to reimagine the experience of all customers on AWS. We explore new technologies and find creative solutions. Curiosity and an explorative mindset can find a place here to impact the life of engineers around the world. If you are excited about this space and want to enlighten your peers with new capabilities, this is the team for you.
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