Social
Conversational AI

A quick guide to Amazon’s 50-plus papers at EMNLP 2024

Large language models predominate, both as a research subject themselves and as tools for researching topics of particular interest to Amazon, such as speech, recommendations, and information retrieval.

By Staff writer
November 14, 2024
7 min read
Share

Large language models (LLMs) have come to dominate the field of natural-language processing, so it’s no surprise that they also dominate the research that Amazon scientists are presenting at this year’s Conference on Empirical Methods in Natural-Language Processing (EMNLP). LLM training is the topic with the greatest number of Amazon papers, followed closely by strategies for mitigating misinformation in LLMs’ outputs — including but not limited to hallucinations. At the same time, a number of papers apply LLMs to topics of traditional interest at Amazon, such as speech, recommender systems, and information retrieval. (Papers marked with asterisks were accepted to Findings of EMNLP.)

AI agents

MARCO: Multi-agent real-time chat orchestration
Anubhav Shrimal, Shervin Malmasi, Kriti Biswas, Swarnalatha Raghuraman, Anish Nediyanchath, Yi Zhang, Promod Yenigalla

Code generation

CodeFort: Robust training for code generation models
Yuhao Zhang, Shiqi Wang, Haifeng Qian, Zijian Wang, Mingyue Shang, Linbo Liu, Sanjay Krishna Gouda, Baishakhi Ray, Murali Krishna Ramanathan, Xiaofei Ma, Anoop Deoras

Socratic human feedback (SoHF): Expert steering strategies for LLM code generation
Subramanian Chidambaram, Erran Li, Min Bai, Xiaopeng LI, Kaixiang Lin, Xiong Zhou, Alex C. Williams

Structured object language modeling (SoLM): Native structured objects generation conforming to complex schemas with self-supervised denoising
Amir Tavanaei, Kee Kiat Koo, Hayreddin Ceker, Shaobai Jiang, Qi Li, Julien Han, Karim Bouyarmane

Contrastive decoding

Explaining and improving contrastive decoding by extrapolating the probabilities of a huge and hypothetical LM
Haw-Shiuan Chang, Nanyun Peng, Mohit Bansal, Anil Ramakrishna, Tagyoung Chung

Explaining and improving contrastive decoding by extrapolating the probabilities of a huge and hypothetical LM.png
Given a simple question with clues, contrastive decoding could have an “obvious blindness” (e.g., assigning higher probability to an uncommon answer, such as "invertebrate", than to the most obvious answer, "bees"). In contrast, the asymptotic probability decoding proposed in "Explaining and improving contrastive decoding by extrapolating the probabilities of a huge and hypothetical LM" correctly assigns the highest probability to "bees" by leveraging the probabilities from multiple LMs of different sizes.

Data integration

ASTRA: Automatic schema matching using machine translation
Tarang Chugh, Deepak Zambre

Learning from natural language explanations for generalizable entity matching
Somin Wadhwa, Adit Krishnan, Runhui Wang, Byron C. Wallace, Chris (Luyang) Kong

Pretraining and finetuning language models on geospatial networks for accurate address matching
Saket Maheshwary, Arpan Paul, Saurabh Sohoney

Retrieval augmented spelling correction for e-commerce applications
Xuan Guo, Rohit Patki, Dante Everaert, Christopher Potts

Dataset distillation

Textual dataset distillation via language model embedding
Yefan Tao, Chris (Luyang) Kong, Andrey Kan, Laurent Callot

Textual dataset distillation via language model embedding: DaLLME.png
The DaLLME framework proposed in "Textual dataset distillation via language model embedding" begins by using a language model to transform raw textual data into embedding vectors. A set of distilled vectors is then derived in the embedding space, through a process designed to encapsulate maximum informational content. Finally, the vec2text model translates these distilled vectors back into textual form.

Document understanding

DocKD: Knowledge distillation from LLMs for open-world document understanding models
Sungnyun Kim, Haofu Liao, Srikar Appalaraju, Peng Tang, Zhuowen Tu, Ravi Kumar Satzoda, R. Manmatha, Vijay Mahadevan, Stefano Soatto

Information retrieval

Evaluating D-MERIT of partial-annotation on information retrieval
Royi Rassin, Yaron Fairstein, Oren Kalinsky, Guy Kushilevitz, Nachshon Cohen, Alexander Libov, Yoav Goldberg

Identifying high consideration e-commerce search queries
Zhiyu Chen, Jason Choi, Besnik Fetahu, Shervin Malmasi

Learning when to retrieve, what to rewrite, and how to respond in conversational QA*
Nirmal Roy, Leonardo Ribeiro, Rexhina Blloshmi, Kevin Small

Natural-language understanding

Intent detection in the age of LLMs
Gaurav Arora, Shreya Jain, Srujana Merugu

Intent detection in the age of LLMs.png
"Intent detection in the age of LLMs" proposes a methodology for adaptive in-context learning and chain-of-thought-based intent detection using LLMs.

Predicting entity salience in extremely short documents
Ben Bullough, Harrison Lundberg, Chen Hu, Weihang Xiao

LLM evaluation

AXCEL: Automated eXplainable consistency evaluation using LLMs*
P Aditya Sreekar, Sahil Verma, Suransh Chopra, Sarik Ghazarian, Abhishek Persad, Narayanan Sadagopan

Precise model benchmarking with only a few observations
Riccardo Fogliato, Pratik Patil, Nil-Jana Akpinar, Mathew Monfort

LLM fine tuning

AdaZeta: Adaptive zeroth-order tensor-train adaption for memory-efficient large language models fine-tuning
Yifan Yang, Kai Zhen, Ershad Banijamali, Thanasis Mouchtaris, Zheng Zhang

RoseLoRA: Row and column-wise sparse low-rank adaptation of pre-trained language model for knowledge editing and fine-tuning
Haoyu Wang, Tianci Liu, Ruirui Li, Monica Cheng, Tuo Zhao, Jing Gao

RoseLoRA.png
The row- and column-wise sparse low-rank adaptation (RoseLoRA) framework proposed in "RoseLoRA: Row and column-wise sparse low-rank adaptation of pre-trained language model for knowledge editing and fine-tuning".

LLMs for speech

Speechworthy instruction-tuned language models
Hyundong Cho, Nicolaas Jedema, Leonardo Ribeiro, Karishma Sharma, Pedro Szekely, Alessandro Moschitti, Ruben Janssen, Jonathan May

LLM misinformation mitigation

ECON: On the detection and resolution of evidence conflicts
Cheng Jiayang, Chunkit Chan, Qianqian Zhuang, Lin Qiu, Tianhang Zhang, Tengxiao Liu, Yangqiu Song, Yue Zhang, Pengfei Liu, Zheng Zhang

Generative subgraph retrieval for knowledge graph–grounded dialog generation
Jinyoung Park, Minseok Joo, Joo-Kyung Kim, Hyunwoo J. Kim

HalluMeasure: Fine-grained hallucination measurement using chain-of-thought reasoning
Shayan Ali Akbar, Md Mosharaf Hossain, Tess Wood, Si-Chi Chin, Erica Salinas, Victor Alvarez, Erwin Cornejo

Knowledge-centric hallucination detection
Xiangkun Hu, Dongyu Ru, Lin Qiu, Qipeng Guo, Tianhang Zhang, Yang Xu, Yun Luo, Pengfei Liu, Zheng Zhang, Yue Zhang

LLM reasoning

Auto-evolve: Enhancing large language model’s performance via self-reasoning framework*
Krishna Aswani, Alex Lu, Pranav Patankar, Priya Dhalwani, Iris Tan, Jayant Ganeshmohan, Simon Lacasse

LLM self-correction

LLM self-correction with DeCRIM: Decompose, critique, and refine for enhanced following of instructions with multiple constraints
Thomas Palmeira Ferraz, Kartik Mehta, Yu-Hsiang Lin, Haw-Shiuan Chang, Shereen Oraby, Sijia Liu, Vivek Subramanian, Tagyoung Chung, Mohit Bansal, Nanyun Peng

DeCRIM.png
In the DeCRIM pipeline proposed in "LLM self-correction with DeCRIM: Decompose, critique, and refine for enhanced following of instructions with multiple constraints", an LLM first generates a response to a user request. The Decomposer then breaks down the request into granular constraints, and the Critic model gives feedback on whether the response meets those constraints. If it does, the response is output; if not, the LLM uses the feedback to refine the response.

LLM training

Dancing in chains: Reconciling instruction following and faithfulness in language models
Zhengxuan Wu, Yuhao Zhang, Peng Qi, Yumo Xu, Rujun Han, Yian Zhang, Jifan Chen, Bonan Min, Zhiheng Huang

DEM: Distribution edited model for training with mixed data distributions
Dhananjay Ram, Aditya Rawal, Momchil Hardalov, Nikolaos Pappas, Sheng Zha

DEM: Distribution Edited Model for Training with Mixed Data Distributions
The distribution-edited model (ΘD) described in "DEM: Distribution edited model for training with mixed data distributions" results from fine-tuning a pretrained model (Θ) on n individual data distributions (Di) and combining the resulting models with basic element-wise vector operations. Here, the extracted distribution vectors (∆ΘDi ) are multiplied by weight coefficients, and the weighted sum is added to the base model.

Evolutionary contrastive distillation for language model alignment
Julian Katz-Samuels, Zheng Li, Hyokun Yun, Priyanka Nigam, Yi Xu, Vaclav Petricek, Bing Yin, Trishul Chilimbi

Hop, skip, jump to convergence: Dynamics of learning rate transitions for improved training of large language models
Shreyas Subramanian, Vignesh Ganapathiraman, Corey Barrett

Learning from relevant subgoals in successful dialogs using iterative training for task-oriented dialog systems
Magdalena Kaiser, Patrick Ernst, Gyuri Szarvas

Quality matters: Evaluating synthetic data for tool-using LLMs
Shadi Iskander, Nachshon Cohen, Zohar Karnin, Ori Shapira, Sofia Tolmach

Query autocompletion

AmazonQAC: A large-scale, naturalistic query autocomplete dataset
Dante Everaert, Rohit Patki, Tianqi Zheng, Christopher Potts

DiAL: Diversity aware listwise ranking for query auto-complete
Sonali Singh, Sachin Farfade, Prakash Mandayam Comar

Question answering

RAG-QA arena: Evaluating domain robustness for long-form retrieval-augmented question answering
Rujun Han, Yuhao Zhang, Peng Qi, Yumo Xu, Jenyuan Wang, Lan Liu, William Yang Wang, Bonan Min, Vittorio Castelli

Retrieving contextual information for long-form question answering using weak supervision
Philipp Christmann, Svitlana Vakulenko, Ionut Teodor Sorodoc, Bill Byrne, Adrià de Gispert

Recommender systems

Efficient pointwise-pairwise learning-to-rank for news recommendation
Nithish Kannen Senthilkumar, Yao Ma, Gerrit van den Burg, Jean Baptiste Faddoul

Efficient pointwise-pairwise learning-to-rank for news recommendation.png
An illustration of the GLIMPSE framework proposed in "Efficient pointwise-pairwise learning-to-rank for news recommendation". GLIMPSE adopts a multitask approach in which a pretrained language model is fine-tuned on both the relevance prediction task and the pairwise-preference task. During inference, the relevance predictions are used to produce an initial pointwise ranking, which is subsequently improved by one or more right-to-left (RTL) passes using pairwise comparisons.

PEARL: Preference extraction with exemplar augmentation and retrieval with LLM agents
Vijit Malik, Akshay Jagatap, Vinayak Puranik, Anirban Majumder

Sequential LLM framework for fashion recommendation
Han Liu, Xianfeng Tang, Tianlang Chen, Jiapeng Liu, Indu Indu, Henry Peng Zou, Peng Dai, Roberto Fernandez Galan, Mike Porter, Dongmei Jia, Ning Zhang, Lian Xiong

Responsible AI

Attribute controlled fine-tuning for large language models: A case study on detoxification
Tao Meng, Ninareh Mehrabi, Palash Goyal, Anil Ramakrishna, Aram Galstyan, Richard Zemel, Kai-Wei Chang, Rahul Gupta, Charith Peris

FLIRT: Feedback loop in-context red teaming
Ninareh Mehrabi, Palash Goyal, Christophe Dupuy, Qian Hu, Shalini Ghosh, Richard Zemel, Kai-Wei Chang, Aram Galstyan, Rahul Gupta

Order of magnitude speedups for LLM membership inference
Rongting Zhang, Martin Bertran Lopez, Aaron Roth

Synthetic data generation

CorrSynth: A correlated sampling method for diverse dataset generation from LLMs
Suhas Kowshik, Abhishek Divekar, Vijit Malik

A Correlated Sampling Method for Diverse Dataset Generation from LLMs
"CorrSynth: A correlated sampling method for diverse dataset generation from LLMs" introduces a sampling method that uses anti-correlation between examples rather than few-shot generation.

DATA ADVISOR: Dynamic data curation for safety alignment of large language models
Fei Wang, Ninareh Mehrabi, Palash Goyal, Rahul Gupta, Kai-Wei Chang, Aram Galstyan

Evaluating differentially private synthetic data generation in high-stakes domains
Krithika Ramesh, Nupoor Gandhi, Pulkit Madaan, Lisa Bauer, Charith Peris, Anjalie Field

SYNTHESIZRR: Generating diverse datasets with retrieval augmentation
Abhishek Divekar, Greg Durrett

Abstract depiction of the SYNTHESIZRR procedure
Abstract depiction of the procedure proposed in "SYNTHESIZRR: Generating diverse datasets with retrieval augmentation". The content sourcing stage retrieves K unique documents {r1,...,rK} from a large corpus for each in-context covariate xICL. The task-inversion stage uses a parameterized context refinement prompt, Pτ, which takes parameters Rinv (inversion instruction), rk (a retrieved document), and V(yICL) (the verbalized target label). A generalist teacher LLM autoregressively generates a synthetic covariate. Each in-context example thus produces K unique synthetic examples {x̃1,..., x̃K}, which we include in the dataset with target yICL.

Text classification

Distance-aware calibration for pre-trained language models*
Alberto Gasparin, Gianluca Detommaso

Performance-guided LLM knowledge distillation for efficient text classification at scale

Flavio Di Palo, Prateek Singhi, Bilal Fadlallah

Prompt-tuned muti-task taxonomic transformer (PTMTTaxoFormer)
Rajashekar Vasantha, Nhan Nguyen, Yue Zhang

Text summarization

Salient information prompting to steer content in prompt-based abstractive summarization
Lei Xu, Asad Karim, Saket Dingliwal, Aparna Elangovan

Research areas
Tags
About the Author
Staff writer

Related content

Work with us

See more jobs See more jobs
Principal Audio Scientist, Fauna
US, NY, New York
Fauna Robotics is building capable, safe, and delightful robots for everyday life, and voice is one of the most natural ways people will interact with them. Cloud speech and language models are good and getting better, but they can only work with the audio they receive, and a robot is a hard place to listen. Its microphones sit beside motors, fans, and moving joints. It speaks through its own loudspeaker while people talk over it. It moves, turns, and shares a room with several people at once. We are hiring a Principal Audio Scientist to be Fauna's technical authority on how our robots hear. You will design, prototype, and ship the hardest algorithms in the robot's audio system. You will set the audio architecture that other engineers build on, shape hardware decisions across robot generations, mentor the engineers and scientists working on audio, and be the person teams come to when the robot can't hear. Key job responsibilities - Set the long-range science roadmap and technical architecture for the robot's audio system, and serve as Fauna's primary technical authority on robot hearing - Design and implement suppression of the robot's own noise from motors, fans, and moving joints - Design and implement echo cancellation for the robot's own voice, so people can interrupt it naturally - Develop multi-microphone processing that holds up as the robot and the people around it move, including locating who is speaking so the robot can turn toward them - Make on-robot listening decisions robust to internal and external noise sources: wake word, voice activity, and whether speech is directed at the robot - Drive microphone and speaker placement, enclosure acoustics, and vibration isolation decisions with mechanical, electrical, and industrial design, backed by your own measurements - Design the robot-specific data collection and evaluation methods to validate the performance of our audio design - Present audio science and its tradeoffs to senior leadership and partner teams - Mentor scientists and engineers, raising the scientific bar for audio across the organization through design reviews, code reviews, and hiring
Applied Scientist, FinAuto
IN, KA, Bengaluru
Interested to build the next generation Financial systems that can handle billions of dollars in transactions? Interested to build highly scalable next generation systems that could utilize Amazon Cloud? Massive data volume + complex business rules in a highly distributed and service oriented architecture, a world class information collection and delivery challenge. Our challenge is to deliver the software systems which accurately capture, process, and report on the huge volume of financial transactions that are generated each day as millions of customers make purchases, as thousands of Vendors and Partners are paid, as inventory moves in and out of warehouses, as commissions are calculated, and as taxes are collected in hundreds of jurisdictions worldwide. Key job responsibilities • Understand the business and discover actionable insights from large volumes of data through application of machine learning, statistics or causal inference. • Analyse and extract relevant information from large amounts of Amazon’s historical transactions data to help automate and optimize key processes • Research, develop and implement novel machine learning and statistical approaches for anomaly, theft, fraud, abusive and wasteful transactions detection. • Use machine learning and analytical techniques to create scalable solutions for business problems. • Identify new areas where machine learning can be applied for solving business problems. • Partner with developers and business teams to put your models in production. • Mentor other scientists and engineers in the use of ML techniques. A day in the life • Understand the business and discover actionable insights from large volumes of data through application of machine learning, statistics or causal inference. • Analyse and extract relevant information from large amounts of Amazon’s historical transactions data to help automate and optimize key processes • Research, develop and implement novel machine learning and statistical approaches for anomaly, theft, fraud, abusive and wasteful transactions detection. • Use machine learning and analytical techniques to create scalable solutions for business problems. • Identify new areas where machine learning can be applied for solving business problems. • Partner with developers and business teams to put your models in production. • Mentor other scientists and engineers in the use of ML techniques. About the team The FinAuto TFAW(theft, fraud, abuse, waste) team is part of FGBS Org and focuses on building applications utilizing machine learning models to identify and prevent theft, fraud, abusive and wasteful(TFAW) financial transactions across Amazon. Our mission is to prevent every single TFAW transaction. As a Machine Learning Scientist in the team, you will be driving the TFAW Sciences roadmap, conduct research to develop state-of-the-art solutions through a combination of data mining, statistical and machine learning techniques, and coordinate with Engineering team to put these models into production. You will need to collaborate effectively with internal stakeholders, cross-functional teams to solve problems, create operational efficiencies, and deliver successfully against high organizational standards.
Robotics/AI Motor Control Scientist, Fauna
US, NY, New York
We are seeking a Robotics/AI Motor Control Scientist to develop cutting-edge machine learning algorithms for motor control systems in robots. In this role, you will focus on creating and optimizing intelligent motor control strategies to enable robots to perform complex, whole-body tasks. Your contributions will be essential in advancing robotics by enabling fluid, reliable, and safe interactions between robots and their environments. Key job responsibilities - Develop controllers that leverage reinforcement learning, imitation learning, or other advanced AI techniques to achieve natural, robust, and adaptive motor behaviors - Collaborate with multi-disciplinary teams to integrate motor control systems with robotic hardware, ensuring alignment with real-world constraints such as actuator dynamics and energy efficiency - Use simulation and real-world testing to refine and validate control algorithms - Stay updated on advancements in robotics, AI, and control systems to apply advanced techniques to robotic motion challenges - Lead technical projects from conception through production deployment - Mentor junior scientists and engineers - Bridge research initiatives with practical engineering implementation About the team Fauna Robotics, an Amazon company, is building capable, safe, and genuinely delightful robots for everyday life. Our goal is simple: make robots people actually want to live and interact with in everyday human spaces. We believe that future won’t arrive until building for robotics becomes far more accessible. Today, too much effort is spent reinventing the fundamentals. We’re changing that by developing tightly integrated hardware and software systems that make it faster, safer, and more intuitive to create real-world robotic products. Our work spans the full stack: mechanical design, control systems, dynamic modeling, and intelligent software. The focus is not just functionality, but experience. We’re building robots that feel responsive, expressive, and genuinely useful. At Fauna, you’ll work at the frontier of this space, helping define how robots move, manipulate, and interact with people in natural environments. It’s an opportunity to solve hard problems across hardware and software with a team focused on making robotics accessible and joyful to build. If you care about making robotics real for everyone and building systems that are as delightful as they are capable, we’re interested in hearing from you. an opportunity to solve hard problems across hardware and software with a team focused on making robotics accessible and joyful to build. If you care about making robotics real for everyone and building systems that are as delightful as they are capable, we’re interested in hearing from you.
Applied Scientist III, AGI Responsible AI (RAI)
US, CA, Sunnyvale
Amazon's Artificial General Intelligence (AGI) organization is seeking an Applied Scientist III to advance the science of Responsible AI evaluation for large language models and generative AI. In this role, you will lead the design and development of rigorous evaluation methods, benchmarks, and metrics that measure the safety, fairness, robustness, and trustworthiness of frontier models. You will work with large-scale datasets, modern deep learning frameworks, and world-class scientists and engineers to turn research into evaluation systems that shape model launch decisions at Amazon scale. Key job responsibilities - Lead the design and implementation of evaluation frameworks, benchmarks, and metrics for responsible AI, including safety, fairness, robustness, and harmful content. - Build scalable automated evaluation pipelines for large language models, including model-based and human-in-the-loop evaluation. - Partner with pretraining, post-training, and product teams to translate evaluation results into model improvements and launch decisions. - Conduct rigorous experimentation and statistical analysis, and publish research at top venues. - Mentor junior scientists and help raise the scientific bar of the team. - Champion responsible AI practices across the model development lifecycle. About the team The AGI Responsible AI (RAI) team builds the science and systems that make Amazon's large language models safe, fair, and trustworthy. We work on problems spanning safety evaluation, content moderation, watermarking, bias mitigation, and alignment. Our team values scientific rigor, customer obsession, and rapid iteration, and we collaborate closely with pretraining, post-training, and product teams across AGI.
(Senior) AI Engineer - Reinforcement Learning Manipulation, RIVR
CH, Zurich
RIVR, an Amazon company, is building Physical AI by deploying autonomous robots for real-world doorstep delivery. Operating daily in diverse urban environments, RIVR's robots continuously learn from and navigate the millions of scenarios encountered during deliveries. By owning the full stack from software to hardware, RIVR is purpose-built for safety, reliability, and the customer from day one. Reinforcement learning is transforming our robotic intelligence, enabling autonomous behavior without human guidance. We are seeking a Senior AI Engineer with deep expertise in reinforcement learning and deep learning, including supervised and self-supervised learning with a focus on dexterous manipulation. Your role will involve leveraging both simulated and real-world data to address practical challenges in dynamic grasping, contact-rich manipulation, and object interaction. If you are passionate about advancing AI and developing innovative solutions, join us in shaping the future of intelligent robotics. Key job responsibilities Develop cutting-edge reinforcement learning algorithms to enable robust, contact-rich dexterous manipulation, translating vision, depth, tactile, and proprioceptive sensor input into precise end-effector and joint-level motor commands. Design, test, and refine algorithms to solve complex real-world manipulation challenges, such as handling diverse package form factors, dynamic hand-offs, and operating door handles or latches. Collaborate with the foundation model team to innovate methods that leverage both simulated and real-world data.
Applied Scientist, TSI Science
US, WA, Seattle
Do you want to join an innovative team of scientists who use machine learning and statistical techniques to help Amazon provide the best customer experience by preventing eCommerce fraud? Are you excited by the prospect of analyzing and modeling terabytes of data and creating state-of-the-art algorithms to solve real world problems? Do you like to own end-to-end business problems/metrics and directly impact the profitability of the company? Do you enjoy collaborating in a diverse team environment? If yes, then you may be a great fit to join the Amazon Selling Partner Trust & Store Integrity Science Team. We are looking for a talented scientist who is passionate to build advanced machine learning systems that help manage the safety of millions of transactions every day and scale up our operation with automation. Key job responsibilities Innovate with the latest GenAI/LLM/VLM technology to build highly automated solutions for efficient fraud detection, risk evaluation and automated operations Design, develop and deploy end-to-end advance machine learning solutions with vision anf GenAi technologies in the Amazon production environment to create impactful business value Learn, explore and experiment with the latest machine learning advancements to create the best customer experience A day in the life You will be working within a dynamic, diverse, and supportive group of scientists who share your passion for innovation and excellence. You'll be working closely with business partners and engineering teams to create end-to-end scalable machine learning solutions that address real-world problems. You will build scalable, efficient, and automated processes for large-scale data analyses, model development, model validation, and model implementation. You will also be providing clear and compelling reports for your solutions and contributing to the ongoing innovation and knowledge-sharing that are central to the team's success.
Applied Scientist, AGI/AI Security
US, NY, New York
We are seeking an Applied Scientist to contribute to research and development of novel security validation and monitoring techniques for AI systems at scale. You will own and contribute to four critical work-streams: 1. Real-Time Agent Monitoring Design and implement scientific approaches for continuous behavioral analysis of AI agents in production—detecting anomalous actions, prompt injection exploitation, and policy violations in real time. 2. Protection & Automated Remediation Invent and deliver novel protection technologies and automated remediation techniques building on research in security, cryptography, privacy, automated reasoning, and others domains, to enable safe and secure agentic AI models and AI applications. 3. AI Application and Capabilities Validation Invent and deliver scalable methodologies for security testing of AI applications and AI capabilities (e.g. MCP, skills), including adversarial robustness evaluation, safety guardrail bypass detection, tool-use authorization boundaries, and trust boundary verification. 4. AI Asset Discovery & Inventory Research and build scalable techniques to automatically discover, identify, and catalog all AI-enabled applications, services, and capabilities across the company—maintaining a comprehensive, continuously updated database of AI assets. Key job responsibilities Invent • Identify and frame new research challenges in AI security where problems are ill-defined and require novel scientific paradigms at the product level. • Contribute to the team's scientific agenda for agent monitoring, protection, remediation, validation research, and AI asset discovery. • Publish research results at peer-reviewed internal and external venues (e.g., USENIX Security, ACM CCS, IEEE S&P, NeurIPS, ICML security workshops, ICSE, PETS) when appropriate. • Articulate key scientific challenges of current and future AI security threats and deliver novel research to address them. • Design, implementation, and successful delivery of scientifically complex security solutions into production—both brand new systems and evolutions of existing ones. • Write significant portions of critical-path code for detection models, validation engines, protection technologies, and asset discovery / classification systems. • Assess and select appropriate technologies (e.g., data protection, private inference, graph-based anomaly detection, NLP-based service classification, code/traffic analysis for AI fingerprinting) for production systems. • Use best practices in scientific methodology and software engineering across the team; provide insightful peer reviews of code, design, and architecture artifacts. • Deliver solutions that are inventive, maintainable, scalable, and extensible. Influence • Autonomously drive discussions with security engineers, software engineers, product managers, and scientist peers across multiple teams. • Build consensus on larger cross-team security initiatives and factor complex efforts into independent workstreams. • Identify and resolve endemic problems, including areas where current security tooling limits innovation of partner teams. • Contribute to the broader internal and external scientific communities as a subject matter expert in AI security. About the team Diverse Experiences Amazon Security values diverse experiences. Even if you do not meet all of the qualifications and skills listed in the job description, we encourage candidates to apply. If your career is just starting, hasn’t followed a traditional path, or includes alternative experiences, don’t let it stop you from applying. Why Amazon Security? At Amazon, security is central to maintaining customer trust and delivering delightful customer experiences. Our organization is responsible for creating and maintaining a high bar for security across all of Amazon’s products and services. We offer talented security professionals the chance to accelerate their careers with opportunities to build experience in a wide variety of areas including cloud, devices, retail, entertainment, healthcare, operations, and physical stores. Inclusive Team Culture In Amazon Security, it’s in our nature to learn and be curious. Ongoing DEI events and learning experiences inspire us to continue learning and to embrace our uniqueness. Addressing the toughest security challenges requires that we seek out and celebrate a diversity of ideas, perspectives, and voices. Training & Career Growth We’re continuously raising our performance bar as we strive to become Earth’s Best Employer. That’s why you’ll find endless knowledge-sharing, training, and other career-advancing resources here to help you develop into a better-rounded professional. Work/Life Balance We value work-life harmony. Achieving success at work should never come at the expense of sacrifices at home, which is why flexible work hours and arrangements are part of our culture. When we feel supported in the workplace and at home, there’s nothing we can’t achieve.
Applied Scientist II, Last Mile Science
IN, KA, Bangalore
Amazon’s Last Mile Team is looking for a passionate individual with strong optimization and analytical skills to join its Last Mile Science team in the endeavor of designing and improving the most complex planning of delivery network in the world. Last Mile builds global solutions that enable Amazon to attract an elastic supply of drivers, companies, and assets needed to deliver Amazon's and other shippers' volumes at the lowest cost and with the best customer delivery experience. Last Mile Science team owns the core decision models in the space of jurisdiction planning, delivery channel and modes network design, capacity planning for on the road and at delivery stations, routing inputs estimation and optimization. Our research has direct impact on customer experience, driver and station associate experience, Delivery Service Partner (DSP)’s success and the sustainable growth of Amazon. Optimizing the last mile delivery requires deep understanding of transportation, supply chain management, pricing strategies and forecasting. Only through innovative and strategic thinking, we will make the right capital investments in technology, assets and infrastructures that allows for long-term success. Our team members have an opportunity to be on the forefront of supply chain thought leadership by working on some of the most difficult problems in the industry with some of the best product managers, scientists, and software engineers in the industry. Key job responsibilities Candidates will be responsible for developing solutions to better manage and optimize delivery capacity in the last mile network. The successful candidate should have solid research experience in one or more technical areas of Operations Research or Machine Learning. These positions will focus on identifying and analyzing opportunities to improve existing algorithms and also on optimizing the system policies across the management of external delivery service providers and internal planning strategies. They require superior logical thinkers who are able to quickly approach large ambiguous problems, turn high-level business requirements into mathematical models, identify the right solution approach, and contribute to the software development for production systems. To support their proposals, candidates should be able to independently mine and analyze data, and be able to use any necessary programming and statistical analysis software to do so. Successful candidates must thrive in fast-paced environments, which encourage collaborative and creative problem solving, be able to measure and estimate risks, constructively critique peer research, and align research focuses with the Amazon's strategic needs. As a senior scientist, you will also help coach/mentor junior scientists in the team.
Applied Scientist, Fauna
US, NY, New York
We are seeking an Applied Scientist to lead the development of evaluation frameworks and data collection protocols for robotic capabilities. In this role, you will focus on designing how we measure, stress-test, and improve robot behavior across a wide range of real-world tasks. Your work will play a critical role in shaping how policies are validated and how high-quality datasets are generated to accelerate system performance. You will operate at the intersection of robotics, machine learning, and human-in-the-loop systems, building the infrastructure and methodologies that connect teleoperation, evaluation, and learning. This includes developing evaluation policies, defining task structures, and contributing to operator-facing interfaces that enable scalable and reliable data collection. The ideal candidate is highly experimental, systems-oriented, and comfortable working across software, robotics, and data pipelines, with a strong focus on turning ambiguous capability goals into measurable and actionable evaluation systems. Key job responsibilities - Design and implement evaluation frameworks to measure robot capabilities across structured tasks, edge cases, and real-world scenarios - Develop task definitions, success criteria, and benchmarking methodologies that enable consistent and reproducible evaluation of policies - Create and refine data collection protocols that generate high-quality, task-relevant datasets aligned with model development needs - Build and iterate on teleoperation workflows and operator interfaces to support efficient, reliable, and scalable data collection - Analyze evaluation results and collected data to identify performance gaps, failure modes, and opportunities for targeted data collection - Collaborate with engineering teams to integrate evaluation tooling, logging systems, and data pipelines into the broader robotics stack - Stay current with advances in robotics, evaluation methodologies, and human-in-the-loop learning to continuously improve internal approaches - Lead technical projects from conception through production deployment - Mentor junior scientists and engineers
Senior Applied Scientist, Fauna
US, NY, New York
We are seeking an Applied Scientist to lead the development of evaluation frameworks and data collection protocols for robotic capabilities. In this role, you will focus on designing how we measure, stress-test, and improve robot behavior across a wide range of real-world tasks. Your work will play a critical role in shaping how policies are validated and how high-quality datasets are generated to accelerate system performance. You will operate at the intersection of robotics, machine learning, and human-in-the-loop systems, building the infrastructure and methodologies that connect teleoperation, evaluation, and learning. This includes developing evaluation policies, defining task structures, and contributing to operator-facing interfaces that enable scalable and reliable data collection. The ideal candidate is highly experimental, systems-oriented, and comfortable working across software, robotics, and data pipelines, with a strong focus on turning ambiguous capability goals into measurable and actionable evaluation systems. Key job responsibilities - Design and implement evaluation frameworks to measure robot capabilities across structured tasks, edge cases, and real-world scenarios - Develop task definitions, success criteria, and benchmarking methodologies that enable consistent and reproducible evaluation of policies - Create and refine data collection protocols that generate high-quality, task-relevant datasets aligned with model development needs - Build and iterate on teleoperation workflows and operator interfaces to support efficient, reliable, and scalable data collection - Analyze evaluation results and collected data to identify performance gaps, failure modes, and opportunities for targeted data collection - Collaborate with engineering teams to integrate evaluation tooling, logging systems, and data pipelines into the broader robotics stack - Stay current with advances in robotics, evaluation methodologies, and human-in-the-loop learning to continuously improve internal approaches - Lead technical projects from conception through production deployment - Mentor junior scientists and engineers
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