Why don't machine learning research agents overfit?
Recorded: Sept. 14, 2026, 5:08 p.m.
| Original | Summarized |
Why machine learning research agents don't overfit — and what compression has to do with it - Amazon Science Close Close Research Research Research areas Automated reasoning Cloud and systems Computer vision Conversational AI Economics Information and knowledge management Machine learning Operations research and optimization Quantum technologies Robotics Search and information retrieval Security, privacy, and abuse prevention Sustainability Our scientific contributions Publications Conferences Research areas Automated reasoning Cloud and systems Computer vision Conversational AI Economics Information and knowledge management Machine learning Operations research and optimization Quantum technologies Robotics Search and information retrieval Security, privacy, and abuse prevention Sustainability Our scientific contributions Publications Conferences News & blog News & blog The latest from Amazon researchers Amazon Science Blog News The latest from Amazon researchers Amazon Science Blog News Collaborations Collaborations Amazon Research Awards Overview Amazon Nova AI Challenge Overview Research collaborations Overview Amazon Research Awards Overview Amazon Nova AI Challenge Overview Research collaborations Overview Resources Resources Code and datasets Amazon Nova Try Amazon’s frontier foundation models. Code and datasets Amazon Nova Try Amazon’s frontier foundation models. Careers Careers Careers Amazon Scholars Postdoctoral Science Program Careers Amazon Scholars Postdoctoral Science Program Social bluesky threads youtube github rss Menu Research Research areas Automated reasoning Cloud and systems Computer vision Conversational AI Economics Information and knowledge management Machine learning Operations research and optimization Quantum technologies Robotics Search and information retrieval Security, privacy, and abuse prevention Sustainability Our scientific contributions Publications Conferences Research areas Automated reasoning Cloud and systems Computer vision Conversational AI Economics Information and knowledge management Machine learning Operations research and optimization Quantum technologies Robotics Search and information retrieval Security, privacy, and abuse prevention Sustainability Our scientific contributions Publications Conferences News & blog The latest from Amazon researchers Amazon Science Blog News The latest from Amazon researchers Amazon Science Blog News Collaborations Amazon Research Awards Overview Amazon Nova AI Challenge Overview Research collaborations Overview Amazon Research Awards Overview Amazon Nova AI Challenge Overview Research collaborations Overview Resources Code and datasets Amazon Nova Try Amazon’s frontier foundation models. Code and datasets Amazon Nova Try Amazon’s frontier foundation models. Careers Careers Amazon Scholars Postdoctoral Science Program Careers Amazon Scholars Postdoctoral Science Program Search Submit Search 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? By Martin Bertran Lopez, Aaron Roth September 10, 2026 11 min read Share Share Copy link X 分享到微信 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 learningReal 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 preciseThe 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 listenerImagine 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 bottleneckThis 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. 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 endThe 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 4xTo 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 becomes12L768 Mu .1 R²Now 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 cheatersA 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'tA 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 Machine learning Tags Generative AI Agentic AI Language models 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 Portable reasoning: Releasing text-bound intelligence into agentic interaction Meiqi Sun Large language models today can solve algebra, pass academic benchmarks, and generate highly structured chain-of-thought explanations. In text-only settings, they often feel startlingly intelligent — methodical, articulate, even strategic. But place those models inside an interactive environment — ask them to click buttons, scroll pages, fill out forms, and submit answers — and their behavior changes. Their careful reasoning falters. They guess where they once deduced. They adhere to templates and produce limited procedural narration: stating what they see and what they will click next, without first forming a structured plan and acting in accordance with plan. It’s as if part of their intelligence has quietly gone offline the moment the cursor appears. Machine learning Optimizing LoRA target module selection for efficient fine tuning Rushil Anirudh, Anjie Fang, Bhoomit Vasani Ablation study clarifies trade-offs between accuracy and efficiency when using low-rank adaptation (LoRA) to fine-tune AI models. Machine learning Promptimus: Improving already good LLM prompts with zero manual engineering Zhengyuan Shen, Yunfei Bai, Sullam Jeoung, Shuai Wang By focusing on specific failure points and suggesting targeted solutions, a new automated prompt-engineering framework improves prompt performance without compromising existing functionality. Work with usSee more jobs Applied Scientist II, HST Health Evaluation, AHS Storefront IN, KA, Bengaluru Amazon Health Services (One Medical) About Us: At Health AI, we're revolutionizing healthcare delivery through innovative AI-enabled solutions. As part of Amazon Health Services and One Medical, we're on a mission to make quality healthcare more accessible while improving patient outcomes. Our work directly impacts millions of lives by empowering patients and enabling healthcare providers to deliver more meaningful care. Role Overview: We're seeking an Applied Scientist to join our dynamic team in building state of the art AI/ML solutions for healthcare. This role offers a unique opportunity to work at the intersection of artificial intelligence and healthcare, developing solutions that will shape the future of medical services delivery. Key job responsibilities • Lead end-to-end development of AI/ML solutions for Amazon Health organization, including Amazon Pharmacy and One Medical • Research, design, and implement state-of-the-art machine learning models, with a focus on Large Language Models (LLMs) and Visual Language Models (VLMs) • Optimize and fine-tune models for production deployment, including model distillation for improved latency • Drive scientific innovation while maintaining a strong focus on practical business outcomes • Collaborate with cross-functional teams to translate complex technical solutions into tangible customer benefits • Contribute to the broader Amazon Health scientific community and help shape our technical roadmap Principal Applied Scientist - AI for Life Sciences, AWS Applied AI Solutions - Life Sciences US, WA, Seattle As part of the AWS Applied AI Solutions organization, we have a vision to provide end user applications, leveraging Amazon's unique experience and expertise, that are used by millions of companies worldwide to manage day-to-day operations. We will accomplish this by accelerating our customers' businesses through delivery of intuitive and differentiated technology solutions that solve enduring business challenges. We blend vision with curiosity and Amazon's real-world experience to build opinionated, turnkey solutions. Where customers prefer to buy over build, we become their trusted partner with solutions that are easy to adopt and easy to use. The Team Join the next science revolution at AWS Life Sciences Applied AI Solutions, where you'll work alongside world-class scientists to build AI that transforms how therapeutics are discovered, developed, and brought to patients. We're out to revolutionize how medicines are discovered, developed, and brought to patients, powered by a new generation of AI. Our team tackles some of the hardest open problems at the intersection of frontier AI and life sciences. We apply biological foundation models, large language models, and agentic reasoning systems to life sciences problems, then put them into the hands of pharma, biotech, and diagnostics customers as applications and managed services they can fine-tune, tailor, and deploy on their own data. The science challenges are deep: how do you design agentic systems that reason correctly over complex biological, regulatory, and clinical logic? How do you enable customers to tailor foundation models to their proprietary data and get better outputs with less effort? How do you adapt models to reason faithfully in high-stakes scientific and regulatory domains? Today we're focused on two areas. In drug design, our products (including Amazon Bio Discovery) accelerate discovery by giving bench scientists AI-guided protein engineering and antibody design capabilities. In clinical trials, we're building AI that automates and optimizes regulatory and clinical development workflows. We combine frontier research with production-scale delivery to put breakthrough science into the hands of customers solving humanity's hardest problems. We value scientific rigor, encourage publication, and support conference participation. If you want to do research that ships, this is the team. The Role We are seeking an exceptional Principal Applied Scientist to set the scientific direction for our life sciences AI portfolio. You will be the scientific leader who defines research agendas, architects novel approaches, and delivers models and methods that give our customers capabilities that did not previously exist. This is a rare role that combines deep expertise in LLM reasoning and agentic AI with applied impact in life sciences. You will innovate on how large language models reason, plan, and act in complex scientific domains, while applying domain knowledge in biology to ensure models produce scientifically valid outputs. The problems span multiple fronts: - How do you build LLM-based agentic systems that correctly reason over clinical protocols, regulatory standards, and complex multi-step scientific workflows? - How do you develop model customization and training methods that let customers get state-of-the-art results from foundation models? - How do you adapt and extend protein and antibody models so customers can fine-tune on proprietary sequence data and get therapeutically relevant outputs? You will work across drug discovery (protein engineering, antibody design) and clinical trial operations (agentic automation, structured reasoning, domain adaptation). You will own end-to-end scientific solutions from research through production, and your work will directly shape the tools that thousands of scientists use daily. Key job responsibilities - Set the scientific vision and research agenda for LLM reasoning, agentic AI, and biological model customization across the portfolio - Innovate on LLM reasoning, planning, and agentic approaches for complex scientific and regulatory workflows - Develop model customization methods (fine-tuning, RLHF, retrieval augmentation, domain adaptation) that enable customers to train better models on their own data with less effort - Advance methods to adapt and extend biological foundation models for customer-specific therapeutic applications - Solve open research problems in faithful reasoning, multi-step planning, and tool use in high-stakes scientific domains - Partner with Life Sciences domain experts and customers to understand their hardest scientific challenges and translate those into tractable research problems - Publish at top-tier venues and build the team's external scientific reputation - Mentor applied scientists across the team while maintaining significant personal research contribution - Collaborate with product and engineering to ensure research translates into shipped products that serve customers at scale - Influence multi-year research roadmaps through deep scientific expertise and customer understanding A day in the life - Push a new reasoning approach into production that measurably improves outputs for a pharma customer's workflow - Design and run experiments to validate a novel fine-tuning method, then ship it as a capability customers can use immediately - Unblock a delivery milestone by diagnosing why a model is failing on a new class of inputs and implementing a fix - Meet with a customer's scientific team to scope what the next model release needs to do for them - Review a teammate's experimental results, sharpen the approach, and help get it over the finish line - Publish results from shipped work at a top venue, closing the loop between research and impact - Prototype a new idea that could become the next major capability in the product Senior Manager, Applied Science US, WA, Seattle We are seeking a Senior Manager, Applied Science to build and lead the science organization across Agentic WorkSpaces. This is a foundational leadership role spanning the full portfolio — Personal, Applications, and Core, and the agentic surfaces (WS4Builders and WorkSpaces for Agents). You will hire, grow, and lead a team of applied scientists who define how we measure and improve the performance of AI agents and human-AI teams. A core part of the role is defining the science agenda itself — identifying which problems are most worth solving and where the highest-leverage bets lie. Directions worth exploring might include Organizational Intelligence (turning institutional knowledge into agent-consumable skills), AI Agent Experience / AiAX (agent observability and autonomous remediation), and contextual, behavioral security that adapts enforcement in real time for human and agent sessions — but these are illustrative examples, not a fixed roadmap, and many other directions are possible. You and your team will define which ones we pursue. The problems your team will solve do not have established industry patterns. You will set the scientific direction and build the team that determines how AI agents and people perceive, reason about, and act reliably within computing environments at enterprise scale. What You Will Do Build and lead the applied science team. Hire, develop, and retain a high-caliber team of applied scientists spanning the Agentic WorkSpaces portfolio. Set the bar for scientific talent, create the growth paths, and build the culture that makes AAWS a destination for the best agent and human-AI researchers. Own the science strategy across the portfolio. Direct the research agenda for how we measure and improve agents and human-AI teams: the benchmarks, task suites, and metrics (accuracy, cost-per-task, task completion, productivity) that turn subjective "it works" judgments into rigorous, reproducible measurement that gates what we ship. Define and drive high-leverage research directions. Work with your team to identify the problems most worth solving and shape the science agenda. Directions worth exploring might include how agents combine deterministic tool use (MCP) with visual reasoning from computer use; Organizational Intelligence and workflow learning (learning from expert recordings, voice annotations, and SOPs); and AI Agent Experience / AiAX (detecting when agents are stuck or degrading productivity and autonomously remediating) — these are illustrative starting points, and your team will weigh them against many other possibilities. Translate science into shipped product. Partner with engineering, product, and program leaders to move models, evaluation, and learning systems from prototype into a decade-old production service operating at massive scale, without compromising the reliability that customers depend on. Represent science in leadership and to customers. Be the scientific voice in org-level planning and roadmap decisions across AAWS, and engage directly with enterprise customers on how agent performance, safety, and human-AI productivity are measured and earned. Key job responsibilities Build and lead the applied science team. Hire, develop, and retain a high-caliber team of applied scientists spanning the Agentic WorkSpaces portfolio. Set the bar for scientific talent, create the growth paths, and build the culture that makes AAWS a destination for the best agent and human-AI researchers. Own the science strategy across the portfolio. Direct the research agenda for how we measure and improve agents and human-AI teams: the benchmarks, task suites, and metrics (accuracy, cost-per-task, task completion, productivity) that turn subjective "it works" judgments into rigorous, reproducible measurement that gates what we ship. Define and drive high-leverage research directions. Work with your team to identify the problems most worth solving and shape the science agenda. Directions worth exploring might include how agents combine deterministic tool use (MCP) with visual reasoning from computer use; Organizational Intelligence and workflow learning (learning from expert recordings, voice annotations, and SOPs); and AI Agent Experience / AiAX (detecting when agents are stuck or degrading productivity and autonomously remediating) — these are illustrative starting points, and your team will weigh them against many other possibilities. Translate science into shipped product. Partner with engineering, product, and program leaders to move models, evaluation, and learning systems from prototype into a decade-old production service operating at massive scale, without compromising the reliability that customers depend on. Represent science in leadership and to customers. Be the scientific voice in org-level planning and roadmap decisions across AAWS, and engage directly with enterprise customers on how agent performance, safety, and human-AI productivity are measured and earned. Set the long-term scientific vision and team strategy: Define what best-in-class agent performance, evaluation, and learning look like across Agentic WorkSpaces — for computer-using agents and human-AI teams alike. Chart a multi-year research roadmap, and build the team and plan to deliver it. Secure buy-in from VP-level leadership. Hire and grow scientific talent: Own recruiting, calibration, development, and retention for the science team. Mentor scientists toward senior and principal scope, and raise the scientific bar across the organization. Direct research on highly ambiguous, novel problems: Guide the team through foundational challenges in agent perception, reasoning, evaluation, reliability, and human-AI collaboration — problems where neither the approach nor the success criteria are pre-defined. Drive cross-organizational alignment: Work across partner teams (AgentCore, Bedrock model teams, Identity, Security, the MCP ecosystem) and across the Applied AI Solutions product portfolio, with product and engineering leadership, to ensure scientific decisions compose into a coherent product. Deliver measurable business impact: Ensure your team's research translates to customer outcomes: higher task accuracy, lower cost-per-action, faster time-to-production, measurable productivity for human-AI teams, and the trust that lets enterprises scale agent workflows. Establish scientific rigor and operational excellence: Set the standard for experimentation, evaluation, and reproducibility, and the mechanisms that keep the science organization productive and accountable. Advance the state of the art: Enable and champion contributions to the external technical community through publications, patents, and open-source work that position AWS as the leader in the science of secure agent-computer interaction and human-AI teamwork. About the team AWS Applied AI Solutions' (AAIS) vision is every business innovating with Amazon AI teammates. Our mission is to build delightful AI solutions that improve human capabilities and business outcomes. The Agentic WorkSpaces organization within AAIS envisions a world where people, teams, and AI collaborate securely from anywhere to create unprecedented value for every organization. We build lovable products that empower every business to unlock the full potential of human-AI teamwork, driving smarter decisions, greater creativity, more value, and faster innovation with confidence. Amazon Agentic WorkSpaces (AAWS) is building the world's most lovable, secure, and trusted always-on workspace where AI agents and humans work as partners behind enterprise-grade security. Our portfolio spans persistent desktops (Personal), application streaming (Applications), and Core, and is evolving into the governed operating environment for the hybrid workforce: humans get AI-native desktops for their role, and agents get governed desktops scoped to their task, with administrators managing both as one. This surface includes WS4Builders (an AI-native environment for builders) and WorkSpaces for Agents (W4A) — enabling AI agents to work the way humans do, with access to real applications, real interfaces, and real computing environments. Enterprises want to use AI agents for critical business workloads that touch legacy desktop applications and mainframes, yet 75% of organizations run legacy applications that lack modern APIs, and 90% of corporate data remains locked in systems never designed for agents. Agentic WorkSpaces solves this: it gives enterprises a secure, governed environment where agents and humans operate both legacy and modern applications directly, just as an employee would, without costly migrations. Applied Scientist II, Amazon Shipping IN, HR, Gurugram Work on ML teams building large-scale forecasting and optimization systems that power Amazon’s global transportation network and directly impact customer experience and cost. As an Applied Scientist II, you will set scientific direction, mentor applied scientists, and partner with engineering and product leaders to deliver production-grade ML solutions at massive scale. Key job responsibilities 1. Lead and grow a high-performing team of Applied Scientists, providing technical guidance, mentorship, and career development. 2. Define and own the scientific vision and roadmap for ML solutions powering large-scale transportation planning and execution. 3. Guide model and system design across a range of techniques, including tree-based models, deep learning (LSTMs, transformers), LLMs, and reinforcement learning. 4. Ensure models are production-ready, scalable, and robust through close partnership with stakeholders. Partner with Product, Operations, and Engineering leaders to enable proactive decision-making and corrective actions. 5. Own end-to-end business metrics, directly influencing customer experience, cost optimization, and network reliability. 6. Help contribute to the broader ML community through publications, conference submissions, and internal knowledge sharing. A day in the life Your day includes reviewing model performance and business metrics, guiding technical design and experimentation, mentoring scientists, and driving roadmap execution. You’ll balance near-term delivery with long-term innovation while ensuring solutions are robust, interpretable, and scalable. Ultimately, your work helps improve delivery reliability, reduce costs, and enhance the customer experience at massive scale. Applied Scientist, Amazon Selection and Catalog Systems (ASCS) US, NY, New York At Amazon Selection and Catalog Systems (ASCS), our mission is to power the online buying experience for customers worldwide so they can find, discover, and buy any product they want. We innovate on behalf of our customers to infer relationships between products in Amazon Catalog to drive the selection gateway for the search and browse experiences on the website. We're solving a fundamental AI challenge: establishing product identity and relationships at unprecedented scale. Using Generative AI, Visual Language Models (VLMs), and multimodal reasoning, we determine what makes each product unique and how products relate to one another across Amazon's catalog. The scale is staggering: billions of products, petabytes of multimodal data, millions of sellers, dozens of languages, and infinite product diversity—from electronics to groceries to digital content. The research challenges are immense. GenAI and VLMs hold transformative promise for catalog understanding, but we operate where traditional methods fail: ambiguous problem spaces, incomplete and noisy data, inherent uncertainty, reasoning across both images and textual data, and explaining decisions at scale. Establishing product identities and groupings requires sophisticated models that reason across text, images, and structured data—while maintaining accuracy and trust for high-stakes business decisions affecting millions of customers daily. Amazon's Item and Relationship Platform group is looking for an innovative and customer-focused applied scientist to help us make the world's best product catalog even better. In this role, you will partner with technology and business leaders to build new state-of-the-art algorithms, models, and services to infer product-to-product relationships that matter to our customers. You will pioneer advanced GenAI solutions that power next-generation agentic shopping experiences, working in a collaborative environment where you can experiment with massive data from the world's largest product catalog, tackle problems at the frontier of AI research, rapidly implement and deploy your algorithmic ideas at scale, across millions of customers. Key job responsibilities * Formulate novel research problems at the intersection of GenAI, multimodal learning, and large-scale information retrieval—translating ambiguous business challenges into tractable scientific frameworks * Design and implement leading models leveraging VLMs, foundation models, and agentic architectures to solve product identity, relationship inference, and catalog understanding at billion-product scale * Pioneer explainable AI methodologies that balance model performance with scalability requirements for production systems impacting millions of daily customer decisions * Own end-to-end ML pipelines from research ideation to production deployment—processing petabytes of multimodal data with rigorous evaluation frameworks * Define research roadmaps aligned with business priorities, balancing foundational research with incremental product improvements * Mentor peer scientists and engineers on advanced ML techniques, experimental design, and scientific rigor—building organizational capability in GenAI and multimodal AI * Represent the team in the broader science community—publishing findings, delivering tech talks, and staying at the forefront of GenAI, VLM, and agentic system research Applied Science Manager, AWS Startups US, WA, Seattle Trusted by more startups around the world, AWS makes the power of cloud computing accessible for all by giving founders everywhere access to the same technology that powers the world's largest companies. With nearly two decades of experience supporting hundreds of thousands of startups, including 80% of unicorns, we democratize cloud computing to help founders bring their innovative ideas to life. We support founders at every stage of their journey, from initial onboarding and credit programs to AI-powered guidance and scale solutions. Data is central to how we do this: it helps us identify high-potential startups early, personalize the guidance we deliver, and prioritize where we can create the most value for founders and for AWS. We are seeking an Applied Science Manager to lead a team of applied scientists and analysts building the data and machine learning capabilities behind AWS Startups. You will own the science roadmap end-to-end, from the data foundation that unifies signals about founders, startups, and their products, through a portfolio of machine learning models, to the surfaces that put insights in the hands of the teams and products that serve startups. You will balance hands-on technical leadership with people management, setting the technical bar for your team while developing their careers. Key job responsibilities · Lead, coach, and grow a team of applied scientists, business intelligence engineers, and business analysts; hire and develop talent and set a high technical bar. · Own and prioritize the team's science roadmap and set technical direction for its machine learning models and data assets, balancing rapid experimentation with production quality, cost, and reliability. · Scope scientific projects, design and evaluate experiments, and productionize models that deliver measurable impact, establishing measurement, evaluation, and operational-excellence standards so quality and impact are quantified and defensible. · Drive the science behind recommendation systems, startup segmentation and targeting, and fraud detection, delivering models that surface relevant opportunities, group and prioritize startups by need and fit, and protect the business from fraud and abuse. · Partner with product, engineering, design, and go-to-market teams to translate science into scalable products, and communicate strategy, results, and trade-offs clearly to technical and non-technical leaders. · Foster a culture of scientific rigor and rapid experimentation, and proactively identify and escalate risks with clear mitigation plans. About the team The AWS Startups team builds innovative products and platforms that support startup customers throughout their journey, from initial onboarding and credit programs to AI-powered guidance and scale solutions. Our portfolio serves hundreds of thousands of startup customers globally, and we partner with business development, field marketing, and solutions architecture teams worldwide. We are building the next generation of AI-native products that make world-class cloud expertise accessible to every founder. Data Scientist, Traffic Quality IN, KA, Bengaluru Amazon Ads is a multi-billion dollar global business that delivers advertising experiences across Amazon's owned-and-operated properties (including Prime Video, Twitch, Fire TV, and Amazon.com), third-party publisher networks, and emerging channels like generative AI-powered shopping experiences. As one of the fastest-growing segments of Amazon, we operate at unprecedented scale across desktop, mobile, connected TV, and emerging surfaces. Within Amazon Ads, Traffic Quality is a critical pillar of advertiser trust and marketplace integrity. Our mission is to build advanced capabilities that work at petabyte scale to detect sophisticated invalid traffic (IVT) which includes sophisticated non-human traffic, bot networks, and fraudulent engagement patterns across programmatic advertising. We are on a journey to establish Amazon Ads as an industry leader in traffic quality standards and transparency. Our research agenda focuses on staying ahead of adversarial actors through continuous innovation in detection methodologies, leveraging state-of-the-art techniques in deep learning and generative modeling, user behavior and multi-modal representation learning, anomaly detection, time-series analysis, and sparse labeling methods. We process billions of ad events daily, developing novel algorithms that balance precision and recall while operating under strict latency constraints. Our work directly protects hundreds of millions of dollars in advertiser spend annually while maintaining a seamless user experience. Key job responsibilities As a Data Scientist II in Traffic Quality, you will solve inherently hard problems in advertising fraud detection by applying advanced statistical techniques and machine learning. You'll work on systems that process billions of ad impressions and clicks per day, using Amazon's cloud services including EC2, S3, EMR, Sagemaker, and RedShift. - Define and frame new research problems in fraud detection where neither problem nor solution is well-defined. - Apply new machine learning approaches, models, and algorithms to detect sophisticated invalid traffic. - Apply domain knowledge to perform broad data analysis as a precursor to modeling and build business insights. - Work with unstructured and massive datasets to deliver results. - Produce research reports meeting top-tier external publication standards. - Mentor and develop junior scientists on the team. About the team Here are a few papers published by the team: 1/ [Scaling Generative Pre-training for User Ad Activity Sequences. AdKDD 2023.](https://assets.amazon.science/b7/42/03be071743d5a57cb1656e6caa34/scaling-generative-pre-training-for-user-ad-activity-sequences.pdf) 2/ [SLIDR: Real-time Robot Detection On Online Ads, IAAI 2023, Deployed Highly Innovative Applications of AI Track (AAAI 2023)](https://assets.amazon.science/75/2f/3b7106b143f38f7f4d2806388ace/real-time-detection-of-robotic-traffic-in-online-advertising.pdf) 3/ [Self-supervised Representation Learning Across Sequential and Tabular Features Using Transformers, NeurIPS 2022, First Table Representation Learning Workshop](https://openreview.net/forum?id=wIIJlmr1Dsk) 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, 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 I, Ads Trust IN, KA, Bengaluru The Ads Trust Science team, based in Bangalore, is responsible for ensuring that ads are relevant and is of good quality, leading to higher conversion for the sellers and providing a great experience for the customers. We deal with one of the world’s largest product catalog, handle billions of requests a day with plans to grow it by order of magnitude and use automated systems to validate tens of millions of offers submitted by thousands of merchants in multiple countries and languages. In this role, you will build and develop ML models to address content understanding problems in Ads. These models will rely on a variety of visual and textual features requiring expertise in both domains. These models need to scale to multiple languages and countries. You will collaborate with engineers and other scientists to build, train and deploy these models. As part of these activities, you will develop production level code that enables moderation of millions of ads submitted each day. See more jobs About Research areas Blog Publications Conferences Code and datasets Academia Amazon News Amazon Developer Amazon Web Services Awards and recognitions Newsletter Careers FAQs Get more from Amazon Science Amazon.com | Conditions of Use | Privacy | © 1996-2026 Amazon.com, Inc. or its affiliates Social bluesky threads youtube github rss |
Machine learning fundamentally concerns generalization rather than mere memorization; a model should aim to perform well on unseen data, and failure to do so indicates overfitting. While standard methodology suggests repeatedly evaluating against a held-out set should lead to overfitting, research in machine learning agents often contradicts this textbook prediction, particularly in the context of iterative improvement loops against benchmarks. This contradiction arises because the learning process, when conducted by autonomous agents, adheres to patterns that possess an underlying structure that is highly compressible, providing an explanation for why successful strategies generalize effectively. The core argument is formalized through the principle of Occam's razor, which suggests that among hypotheses explaining the data equally well, the simpler one—that is, the one describable in fewer bits—is the most likely to generalize to new examples. This intuition is mathematically supported by the idea that if a hypothesis is sufficiently compressed, it cannot retain enough information to encode the training data via memorization. If a compact description performs well on training data, it must reflect genuine structural knowledge rather than specific memorized instances. To empirically test this, the researchers designed an experiment involving an autonomous machine learning research agent, termed the explorer, which iteratively optimizes its strategy by engaging in a hill-climbing loop using a validation set. This process, mirroring human research, is expected to cause overfitting. The experiment then introduces a compression mechanism to test the learnability of the resulting strategy. A second agent, the compressor, attempts to distill the explorer's successful strategy into a severely limited information bottleneck, such as a few tokens. This compressed representation is then passed to a third agent, the reproducer, which must recreate the original strategy from scratch using only this terse prompt, with no access to the original validation set or the explorer's intermediate work. The critical finding is that if the reproducer successfully matches the explorer’s performance, it implies that all the necessary validation-dependent information to specify the strategy was successfully compressed through the narrow channel. This demonstrates that successful machine learning strategies are highly compressible and reflect real structural knowledge rather than overfitting to specific data points. This method serves as a diagnostic tool: strategies that genuinely overfit encode idiosyncrasies of the validation data, and these specific gains vanish when passed through the compression bottleneck. This separation allows compression to serve as a method for catching overfitting by flagging strategies that cannot be adequately summarized. Large language models are essential in this framework because their vast world knowledge enables them to act as powerful compression decoders. They can reconstruct complex machine learning pipelines from terse, expert-shorthand prompts, effectively acting as excellent compression decoders. This capability suggests that the successful strategies discovered through extensive experimentation are inherently simple, often consisting of a short list of familiar choices regarding architecture, optimizers, and regularization schemes. Furthermore, the experiment also explored compressing the input rather than the output, finding that even providing only a single bit of feedback per query allowed the explorer to find strategies comparable to those obtained with full numerical scores, reinforcing the observation that the channel between validation data and the final strategy is inherently narrow. Ultimately, the research suggests that the most effective learning recipes that guide autonomous agents are inherently concise, supporting the hypothesis that success is achieved through generalization, not by rote memorization. |