Teaching language models to reason consistently

At this year’s ACL, Amazon researchers won an outstanding-paper award for showing that knowledge distillation using contrastive decoding in the teacher model and counterfactual reasoning in the student model improves the consistency of “chain of thought” reasoning.

Teaching large language models (LLMs) to reason is an active topic of research in natural-language processing, and a popular approach to that problem is the so-called chain-of-thought paradigm, in which a model is prompted not just to answer questions but to provide rationales for its answers.

In-context learning.png
The structure of the type of prompt used to induce chain-of-thought reasoning in a large language model.

However, given LLMs’ tendency to hallucinate (that is, make spurious factual assertions), the generated rationales may be inconsistent with the predicted answers, making them untrustworthy.

In a paper we presented at this year’s meeting of the Association for Computational Linguistics (ACL), we show how to improve the consistency of chain-of-thought reasoning through knowledge distillation: given pairs of questions and answers from a training set, an LLM — the “teacher” — generates rationales for a smaller “student” model, which learns to both answer questions and provide rationales for its answers. Our paper received one of the conference’s outstanding-paper awards, reserved for 39 of the 1,074 papers accepted to the main conference.

Sample rationale.png
Example of a student model outputting a rationale together with the answer to a question.

With knowledge distillation (KD), we still have to contend with the possibility that the rationales generated by the teacher are spurious or vacuous. On the student side, the risk is that while the model may learn to produce rationales, and it may learn to deliver answers, it won’t learn the crucial logical relationships between the two; it might, for instance, learn inferential short cuts between questions and answers that bypass the whole reasoning process.

False rationales.png
In a study involving a leading LLM, we found that 42% of generated rationales were vacuous (top), and 37% were irrelevant (bottom).

To curb hallucination, on the teacher side, we use contrastive decoding, which ensures that the rationales generated for true assertions differ as much as possible from the rationales generated for false assertions.

To train the student model, we use counterfactual reasoning, in which the model is trained on both true and false rationales and must learn to provide the answer that corresponds to the rationale, even if it’s wrong. To ensure that this doesn’t compromise model performance, during training, we label true rationales “factual” and false rationales “counterfactual”.

Counterfactual training.png
Counterfactual training eliminates reasoning short cuts, in which the student model uses incidental features of the input question to leap to an answer, without performing the intervening inferential steps.

To evaluate our model, we compared it to a chain-of-thought model built using ordinary knowledge distillation, on datasets for four different reasoning tasks. We asked human reviewers to evaluate the rationales generated by the teacher models. To evaluate the student models, we used the leakage-adjusted simulatability (LAS) metric, which measures the ability of a simulator (an external model) to predict the student’s outputs from the generated rationales. Across the board, our models outperformed the baselines, while preserving accuracy on the reasoning tasks.

Contrastive decoding

As our teacher model, we use a trained LLM whose parameters are frozen. To generate training examples for the student model, we use in-context learning, in which we provide the teacher with a handful of examples of questions, answers, and human-annotated rationales, then supply a final question-answer pair. The model generates the rationale for the final pair.

Related content
Methods for controlling the outputs of large generative models and integrating symbolic reasoning with machine learning are among the conference’s hot topics.

During training, LLMs learn the probabilities of sequences of words. At generation time, they either select the single most probable word to continue a sequence or sample from the top-ranked words. This is the standard decoding step, which doesn’t guarantee that the generated rationales justify the model’s answers.

We can control the decoding process without making any adjustments to the LLM parameters. With contrastive decoding, we perform the same in-context rationale generation twice, once with the true answer in the final question-answer pair and once with a perturbed answer.

Then, when we’re decoding the true question-answer pair, we select words that are not only probable given the true pair but relatively improbable given the false pair. In other words, we force the rationale for the true pair to diverge from the rationale for the false pair. In this way, we ensure that the output skews toward rationales particularized to the answers in the question-answer pairs.

In our experiments, we considered two types of perturbation to the true answers: null answers, where no answer at all was supplied, and false answers. We found that contrastive decoding using false answers consistently yielded better rationales than contrastive decoding using null answers.

Counterfactual reasoning

Past research has shown that question-answering models will often exploit short cuts in their training data to improve performance. For instance, answering “who?” questions with the first proper name encountered in a source document will yield the right answer with surprising frequency.

Similarly, a chain-of-thought model might learn to use shortcuts in answering questions and generate rationales as a parallel task, without learning the crucial connection between the two. The goal of training our model on a counterfactual-reasoning objective is to break that short cut.

Related content
Amazon’s Dan Roth on a hot new research topic — that he’s been studying for more than 25 years.

To generate counterfactual training data, we randomly vary the answers in question-answer pairs and generate the corresponding rationales, just as we did for contrastive decoding. Then we train the student model using the questions and rationales as input, and it must generate the corresponding answers.

This means that the student model may very well see the same question multiple times during training, but with different answers (and rationales). The “factual” and “counterfactual” tags prevent it from getting confused about its task.

In our experiments, we compared our approach to one that also uses in-context learning but uses greedy decoding to produce rationales — that is, a decoding method that always selects the highest-probability word. We also used two other baselines: an LLM that directly generates rationales from in-context learning and a model trained on human-annotated rationales.

Our study with human evaluators showed that in-context learning with contrastive decoding generated more persuasive rationales than in-context learning with greedy decoding:

Teacher ModelGrammaticalityNew InfoSupports Answer
Greedy0.990.650.48
Contrast.-Empty0.970.770.58
Contrast.-Wrong0.970.820.63

Table: Human evaluation of data generated with greedy decoding, contrastive decoding using empty answers, and contrastive decoding using incorrect answers.

In the experiments using the LAS metric, knowledge distillation using contrastive decoding alone consistently outperformed all three baselines, and knowledge distillation with counterfactual reasoning and contrastive decoding consistently outperformed knowledge distillation with contrastive decoding alone. The model trained on the human-annotated dataset yielded the most-accurate results on downstream tasks, but its rationales fared badly. On average, our model was slightly more accurate than the one trained using greedy decoding.

LAS results.png
Experimental results, measured according to leakage-adjusted simulatability (LAS) and question-answering accuracy.

Research areas

Related content

GB, MLN, Edinburgh
We’re looking for a Machine Learning Scientist in the Personalization team for our Edinburgh office experienced in generative AI and large models. You will be responsible for developing and disseminating customer-facing personalized recommendation models. This is a hands-on role with global impact working with a team of world-class engineers and scientists across the Edinburgh offices and wider organization. You will lead the design of machine learning models that scale to very large quantities of data, and serve high-scale low-latency recommendations to all customers worldwide. You will embody scientific rigor, designing and executing experiments to demonstrate the technical efficacy and business value of your methods. You will work alongside a science team to delight customers by aiding in recommendations relevancy, and raise the profile of Amazon as a global leader in machine learning and personalization. Successful candidates will have strong technical ability, focus on customers by applying a customer-first approach, excellent teamwork and communication skills, and a motivation to achieve results in a fast-paced environment. Our position offers exceptional opportunities for every candidate to grow their technical and non-technical skills. If you are selected, you have the opportunity to make a difference to our business by designing and building state of the art machine learning systems on big data, leveraging Amazon’s vast computing resources (AWS), working on exciting and challenging projects, and delivering meaningful results to customers world-wide. Key job responsibilities Develop machine learning algorithms for high-scale recommendations problems. Rapidly design, prototype and test many possible hypotheses in a high-ambiguity environment, making use of both quantitative analysis and business judgement. Collaborate with software engineers to integrate successful experimental results into large-scale, highly complex Amazon production systems capable of handling 100,000s of transactions per second at low latency. Report results in a manner which is both statistically rigorous and compellingly relevant, exemplifying good scientific practice in a business environment.
IN, TS, Hyderabad
Welcome to the Worldwide Returns & ReCommerce team (WWR&R) at Amazon.com. WWR&R is an agile, innovative organization dedicated to ‘making zero happen’ to benefit our customers, our company, and the environment. Our goal is to achieve the three zeroes: zero cost of returns, zero waste, and zero defects. We do this by developing products and driving truly innovative operational excellence to help customers keep what they buy, recover returned and damaged product value, keep thousands of tons of waste from landfills, and create the best customer returns experience in the world. We have an eye to the future – we create long-term value at Amazon by focusing not just on the bottom line, but on the planet. We are building the most sustainable re-use channel we can by driving multiple aspects of the Circular Economy for Amazon – Returns & ReCommerce. Amazon WWR&R is comprised of business, product, operational, program, software engineering and data teams that manage the life of a returned or damaged product from a customer to the warehouse and on to its next best use. Our work is broad and deep: we train machine learning models to automate routing and find signals to optimize re-use; we invent new channels to give products a second life; we develop highly respected product support to help customers love what they buy; we pilot smarter product evaluations; we work from the customer backward to find ways to make the return experience remarkably delightful and easy; and we do it all while scrutinizing our business with laser focus. You will help create everything from customer-facing and vendor-facing websites to the internal software and tools behind the reverse-logistics process. You can develop scalable, high-availability solutions to solve complex and broad business problems. We are a group that has fun at work while driving incredible customer, business, and environmental impact. We are backed by a strong leadership group dedicated to operational excellence that empowers a reasonable work-life balance. As an established, experienced team, we offer the scope and support needed for substantial career growth. Amazon is earth’s most customer-centric company and through WWR&R, the earth is our customer too. Come join us and innovate with the Amazon Worldwide Returns & ReCommerce team!
US, WA, Seattle
Prime Video is a first-stop entertainment destination offering customers a vast collection of premium programming in one app available across thousands of devices. Prime members can customize their viewing experience and find their favorite movies, series, documentaries, and live sports – including Amazon MGM Studios-produced series and movies; licensed fan favorites; and programming from Prime Video add-on subscriptions such as Apple TV+, Max, Crunchyroll and MGM+. All customers, regardless of whether they have a Prime membership or not, can rent or buy titles via the Prime Video Store, and can enjoy even more content for free with ads. Are you interested in shaping the future of entertainment? Prime Video's technology teams are creating best-in-class digital video experience. As a Prime Video technologist, you’ll have end-to-end ownership of the product, user experience, design, and technology required to deliver state-of-the-art experiences for our customers. You’ll get to work on projects that are fast-paced, challenging, and varied. You’ll also be able to experiment with new possibilities, take risks, and collaborate with remarkable people. We’ll look for you to bring your diverse perspectives, ideas, and skill-sets to make Prime Video even better for our customers. With global opportunities for talented technologists, you can decide where a career Prime Video Tech takes you! In Prime Video READI, our mission is to automate infrastructure scaling and operational readiness. We are growing a team specialized in time series modeling, forecasting, and release safety. This team will invent and develop algorithms for forecasting multi-dimensional related time series. The team will develop forecasts on key business dimensions with optimization recommendations related to performance and efficiency opportunities across our global software environment. As a founding member of the core team, you will apply your deep coding, modeling and statistical knowledge to concrete problems that have broad cross-organizational, global, and technology impact. Your work will focus on retrieving, cleansing and preparing large scale datasets, training and evaluating models and deploying them to production where we continuously monitor and evaluate. You will work on large engineering efforts that solve significantly complex problems facing global customers. You will be trusted to operate with complete independence and are often assigned to focus on areas where the business and/or architectural strategy has not yet been defined. You must be equally comfortable digging in to business requirements as you are drilling into design with development teams and developing production ready learning models. You consistently bring strong, data-driven business and technical judgment to decisions. You will work with internal and external stakeholders, cross-functional partners, and end-users around the world at all levels. Our team makes a big impact because nothing is more important to us than delivering for our customers, continually earning their trust, and thinking long term. You are empowered to bring new technologies to your solutions. If you crave a sense of ownership, this is the place to be.
US, WA, Seattle
Amazon Advertising operates at the intersection of eCommerce and advertising, and is investing heavily in building a world-class advertising business. We are defining and delivering a collection of self-service performance advertising products that drive discovery and sales. Our products are strategically important to our Retail and Marketplace businesses driving long-term growth. We deliver billions of ad impressions and millions of clicks daily and are breaking fresh ground to create world-class products to improve both shopper and advertiser experience. With a broad mandate to experiment and innovate, we grow at an unprecedented rate with a seemingly endless range of new opportunities. The Ad Response Prediction team in Sponsored Products organization build advanced deep-learning models, large-scale machine-learning pipelines, and real-time serving infra to match shoppers’ intent to relevant ads on all devices, for all contexts and in all marketplaces. Through precise estimation of shoppers’ interaction with ads and their long-term value, we aim to drive optimal ads allocation and pricing, and help to deliver a relevant, engaging and delightful ads experience to Amazon shoppers. As the business and the complexity of various new initiatives we take continues to grow, we are looking for talented Applied Scientists to join the team. Key job responsibilities As a Applied Scientist II, you will: * Conduct hands-on data analysis, build large-scale machine-learning models and pipelines * Work closely with software engineers on detailed requirements, technical designs and implementation of end-to-end solutions in production * Run regular A/B experiments, gather data, perform statistical analysis, and communicate the impact to senior management * Establish scalable, efficient, automated processes for large-scale data analysis, machine-learning model development, model validation and serving * Provide technical leadership, research new machine learning approaches to drive continued scientific innovation * Be a member of the Amazon-wide Machine Learning Community, participating in internal and external MeetUps, Hackathons and Conferences
US, CA, Palo Alto
Amazon’s Advertising Technology team builds the technology infrastructure and ad serving systems to manage billions of advertising queries every day. The result is better quality advertising for publishers and more relevant ads for customers. In this organization you’ll experience the benefits of working in a dynamic, entrepreneurial environment, while leveraging the resources of Amazon.com (AMZN), one of the world's leading companies. Amazon Publisher Services (APS) helps publishers of all sizes and on all channels better monetize their content through effective advertising. APS unites publishers with advertisers across devices and media channels. We work with Amazon teams across the globe to solve complex problems for our customers. The end results are Amazon products that let publishers focus on what they do best - publishing. The APS Publisher Products Engineering team is responsible for building cloud-based advertising technology services that help Web, Mobile, Streaming TV broadcasters and Audio publishers grow their business. The engineering team focuses on unlocking our ad tech on the most impactful Desktop, mobile and Connected TV devices in the home, bringing real-time capabilities to this medium for the first time. As a successful Data Scientist in our team, · You are an analytical problem solver who enjoys diving into data, is excited about investigations and algorithms, and can credibly interface between technical teams and business stakeholders. You will collaborate directly with product managers, BIEs and our data infra team. · You will analyze large amounts of business data, automate and scale the analysis, and develop metrics (e.g., user recognition, ROAS, Share of Wallet) that will enable us to continually measure the impact of our initiatives and refine the product strategy. · Your analytical abilities, business understanding, and technical aptitude will be used to identify specific and actionable opportunities to solve existing business problems and look around corners for future opportunities. Your expertise in synthesizing and communicating insights and recommendations to audiences of varying levels of technical sophistication will enable you to answer specific business questions and innovate for the future. · You will have direct exposure to senior leadership as we communicate results and provide scientific guidance to the business. Major responsibilities include: · Utilizing code (Apache, Spark, Python, R, Scala, etc.) for analyzing data and building statistical models to solve specific business problems. · Collaborate with product, BIEs, software developers, and business leaders to define product requirements and provide analytical support · Build customer-facing reporting to provide insights and metrics which track system performance · Influence the product strategy directly through your analytical insights · Communicating verbally and in writing to business customers and leadership team with various levels of technical knowledge, educating them about our systems, as well as sharing insights and recommendations
US, WA, Bellevue
mmPROS Surface Research Science seeks an exceptional Applied Scientist with expertise in optimization and machine learning to optimize Amazon's middle mile transportation network, the backbone of its logistics operations. Amazon's middle mile transportation network utilizes a fleet of semi-trucks, trains, and airplanes to transport millions of packages and other freight between warehouses, vendor facilities, and customers, on time and at low cost. The Surface Research Science team delivers innovation, models, algorithms, and other scientific solutions to efficiently plan and operate the middle mile surface (truck and rail) transportation network. The team focuses on large-scale problems in vehicle route planning, capacity procurement, network design, forecasting, and equipment re-balancing. Your role will be to build innovative optimization and machine learning models to improve driver routing and procurement efficiency. Your models will impact business decisions worth billions of dollars and improve the delivery experience for millions of customers. You will operate as part of a team of innovative, experienced scientists working on optimization and machine learning. You will work in close collaboration with partners across product, engineering, business intelligence, and operations. Key job responsibilities - Design and develop optimization and machine learning models to inform our hardest planning decisions. - Implement models and algorithms in Amazon's production software. - Lead and partner with product, engineering, and operations teams to drive modeling and technical design for complex business problems. - Lead complex modeling and data analyses to aid management in making key business decisions and set new policies. - Write documentation for scientific and business audiences. About the team This role is part of mmPROS Surface Research Science. Our mission is to build the most efficient and optimal transportation network on the planet, using our science and technology as our biggest advantage. We leverage technologies in optimization, operations research, and machine learning to grow our businesses and solve Amazon's unique logistical challenges. Scientists in the team work in close collaboration with each other and with partners across product, software engineering, business intelligence, and operations. They regularly interact with software engineering teams and business leadership.
IL, Tel Aviv
Come join the AWS Agentic AI science team in building the next generation models for intelligent automation. AWS, the world-leading provider of cloud services, has fostered the creation and growth of countless new businesses, and is a positive force for good. Our customers bring problems that will give Applied Scientists like you endless opportunities to see your research have a positive and immediate impact in the world. You will have the opportunity to partner with technology and business teams to solve real-world problems, have access to virtually endless data and computational resources, and to world-class engineers and developers that can help bring your ideas into the world. As part of the team, we expect that you will develop innovative solutions to hard problems, and publish your findings at peer reviewed conferences and workshops. We are looking for world class researchers with experience in one or more of the following areas - autonomous agents, API orchestration, Planning, large multimodal models (especially vision-language models), reinforcement learning (RL) and sequential decision making. Key job responsibilities PhD, or Master's degree and 4+ years of CS, CE, ML or related field experience 3+ years of building models for business application experience Experience in patents or publications at top-tier peer-reviewed conferences or journals Experience programming in Java, C++, Python or related language Experience in any of the following areas: algorithms and data structures, parsing, numerical optimization, data mining, parallel and distributed computing, high-performance computing
IL, Haifa
Come join the AWS Agentic AI science team in building the next generation models for intelligent automation. AWS, the world-leading provider of cloud services, has fostered the creation and growth of countless new businesses, and is a positive force for good. Our customers bring problems that will give Applied Scientists like you endless opportunities to see your research have a positive and immediate impact in the world. You will have the opportunity to partner with technology and business teams to solve real-world problems, have access to virtually endless data and computational resources, and to world-class engineers and developers that can help bring your ideas into the world. As part of the team, we expect that you will develop innovative solutions to hard problems, and publish your findings at peer reviewed conferences and workshops. We are looking for world class researchers with experience in one or more of the following areas - autonomous agents, API orchestration, Planning, large multimodal models (especially vision-language models), reinforcement learning (RL) and sequential decision making.
US, MA, Westborough
We are seeking a Principal Applied Scientist to lead the development of our autonomous driving stack for last-mile delivery vehicles. In this role, you will drive technical innovation, architect advanced autonomous systems, and lead a team of researchers and engineers in pushing the boundaries of what's possible in autonomous delivery. Key job responsibilities As the Principal Applied Scientist, you will architect and evolve LMDA's autonomous driving stack for last-mile delivery vehicles. Your role involves driving research and development in key areas such as perception, prediction, planning, and control. You will develop novel algorithms and approaches to solve complex challenges in urban autonomous navigation. A critical aspect of your role will be leading system-level architecture decisions and setting technical direction for the autonomous systems team. You will mentor and develop a team of scientists and engineers, fostering a culture of innovation and excellence. This involves close collaboration with cross-functional teams including hardware, safety, and operations to ensure seamless integration of autonomous systems. As a senior technical leader, you will represent LMDA's technical capabilities to partners, customers, and at industry conferences. In this role, you will define and execute the technical roadmap for LMDA's autonomous systems. This includes identifying key research areas and technological advancements that will drive LMDA's competitive advantage. A crucial aspect of your role will be balancing long-term research goals with near-term product delivery needs. You will lead the integration of various autonomous subsystems into a cohesive, performant stack. This includes developing and implementing strategies for optimizing system performance across hardware and software. You will also design and oversee testing and validation frameworks for autonomous systems. About the team Last Mile Delivery Automation (LMDA) is at the forefront of revolutionizing the logistics industry through advanced autonomous vehicle technology. Our mission is to create safe, efficient, and scalable autonomous solutions for last-mile delivery, reducing costs and environmental impact while improving delivery speed and reliability.
US, NY, New York
Join us in a historic endeavor to make Generative AI accessible to the world with breakthrough research! The AWS AI team has a world-leading team of researchers and academics, and we are looking for world-class colleagues to join us and make the AI revolution happen. Our team of scientists drives the innovation that enables external and internal SageMaker customers to train their next generation models on both GPU and Trainium instances. As part of the team, we expect that you will develop innovative solutions to hard problems, and publish your findings at peer reviewed conferences and workshops. AWS is the world-leading provider of cloud services, has fostered the creation and growth of countless new businesses, and is a positive force for good. Our customers bring problems which will give Applied Scientists like you endless opportunities to see your research have a positive and immediate impact in the world. You will have the opportunity to partner with technology and business teams to solve real-world problems, have access to virtually endless data and computational resources, and to world-class engineers and developers that can help bring your ideas into the world. About the team Why AWS Amazon Web Services (AWS) is the world’s most comprehensive and broadly adopted cloud platform. We pioneered cloud computing and never stopped innovating — that’s why customers from the most successful startups to Global 500 companies trust our robust suite of products and services to power their businesses. Utility Computing (UC) AWS Utility Computing (UC) provides product innovations — from foundational services such as Amazon’s Simple Storage Service (S3) and Amazon Elastic Compute Cloud (EC2), to consistently released new product innovations that continue to set AWS’s services and features apart in the industry. As a member of the UC organization, you’ll support the development and management of Compute, Database, Storage, Internet of Things (IoT), Platform, and Productivity Apps services in AWS, including support for customers who require specialized security solutions for their cloud services. Inclusive Team Culture Here at AWS, it’s in our nature to learn and be curious. Our employee-led affinity groups foster a culture of inclusion that empower us to be proud of our differences. Ongoing events and learning experiences, including our Conversations on Race and Ethnicity (CORE) and AmazeCon (gender diversity) conferences, inspire us to never stop embracing our uniqueness. 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 we strive for flexibility as part of our working culture. When we feel supported in the workplace and at home, there’s nothing we can’t achieve in the cloud. Mentorship and 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, mentorship and other career-advancing resources here to help you develop into a better-rounded professional. Diverse Experiences Amazon values diverse experiences. Even if you do not meet all of the preferred 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.