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

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.

Related content

US, WA, Seattle
The Amazon Economics Team is hiring Economist Interns. We are looking for detail-oriented, organized, and responsible individuals who are eager to learn how to work with large and complicated data sets to solve real-world business problems. Some knowledge of econometrics, as well as basic familiarity with Stata, R, or Python is necessary. Experience with SQL, UNIX, Sawtooth, and Spark would be a plus. These are full-time positions at 40 hours per week, with compensation being awarded on an hourly basis. You will learn how to build data sets and perform applied econometric analysis at Internet speed collaborating with economists, data scientists and MBAʼs. These skills will translate well into writing applied chapters in your dissertation and provide you with work experience that may help you with future job market placement. Roughly 85% of interns from previous cohorts have converted to full-time economics employment at Amazon. If you are interested, please send your CV to our mailing list at econ-internship@amazon.com. We are open to hiring candidates to work out of one of the following locations: Seattle, WA, USA
GB, Cambridge
Our team undertakes research together with multiple organizations to advance the state-of-the-art in speech technologies. We not only work on giving Alexa, the ground-breaking service that powers Echo, her voice, but we also develop cutting-edge technologies with Amazon Studios, the provider of original content for Prime Video. Do you want to be part of the team developing the latest technology that impacts the customer experience of ground-breaking products? Then come join us and make history. We are looking for a passionate, talented, and inventive Senior Applied Scientist with a background in Machine Learning to help build industry-leading Speech, Language and Video technology. As a Senior Applied Scientist at Amazon you will work with talented peers to develop novel algorithms and modelling techniques to drive the state of the art in speech and vocal arts synthesis. Position Responsibilities: - Participate in the design, development, evaluation, deployment and updating of data-driven models for digital vocal arts applications. - Participate in research activities including the application and evaluation and digital vocal and video arts techniques for novel applications. - Research and implement novel ML and statistical approaches to add value to the business. - Mentor junior engineers and scientists. We are open to hiring candidates to work out of one of the following locations: Cambridge, GBR
US, VA, Arlington
The People eXperience and Technology Central Science Team (PXTCS) uses economics, behavioral science, statistics, and machine learning to proactively identify mechanisms and process improvements which simultaneously improve Amazon and the lives, wellbeing, and the value of work to Amazonians. We are an interdisciplinary team that combines the talents of science and engineering to develop and deliver solutions that measurably achieve this goal. We are looking for economists who are able to apply economic methods to address business problems. The ideal candidate will work with engineers and computer scientists to estimate models and algorithms on large scale data, design pilots and measure their impact, and transform successful prototypes into improved policies and programs at scale. We are looking for creative thinkers who can combine a strong technical economic toolbox with a desire to learn from other disciplines, and who know how to execute and deliver on big ideas as part of an interdisciplinary technical team. Ideal candidates will work in a team setting with individuals from diverse disciplines and backgrounds. They will work with teammates to develop scientific models and conduct the data analysis, modeling, and experimentation that is necessary for estimating and validating models. They will work closely with engineering teams to develop scalable data resources to support rapid insights, and take successful models and findings into production as new products and services. They will be customer-centric and will communicate scientific approaches and findings to business leaders, listening to and incorporate their feedback, and delivering successful scientific solutions. Key job responsibilities Use reduced-form causal analysis and/or structural economic modeling methods to evaluate the impact of policies on employee outcomes, and examine how external labor market and economic conditions impact Amazon's ability to hire and retain talent. A day in the life Work with teammates to apply economic methods to business problems. This might include identifying the appropriate research questions, writing code to implement a DID analysis or estimate a structural model, or writing and presenting a document with findings to business leaders. Our economists also collaborate with partner teams throughout the process, from understanding their challenges, to developing a research agenda that will address those challenges, to help them implement solutions. About the team We are a multidisciplinary team that combines the talents of science and engineering to develop innovative solutions to make Amazon Earth's Best Employer. We are open to hiring candidates to work out of one of the following locations: Arlington, VA, USA
US, WA, Seattle
We are expanding our Global Risk Management & Claims team and insurance program support for Amazon’s growing risk portfolio. This role will partner with our risk managers to develop pricing models, determine rate adequacy, build underwriting and claims dashboards, estimate reserves, and provide other analytical support for financially prudent decision making. As a member of the Global Risk Management team, this role will provide actuarial support for Amazon’s worldwide operation. Key job responsibilities ● Collaborate with risk management and claims team to identify insurance gaps, propose solutions, and measure impacts insurance brings to the business ● Develop pricing mechanisms for new and existing insurance programs utilizing actuarial skills and training in innovative ways ● Build actuarial forecasts and analyses for businesses under rapid growth, including trend studies, loss distribution analysis, ILF development, and industry benchmarks ● Design actual vs expected and other metrics dashboards to assist decision makings in pricing analysis ● Create processes to monitor loss cost and trends ● Propose and implement loss prevention initiatives with impact on insurance pricing in mind ● Advise underwriting decisions with analysis on driver risk profile ● Support insurance cost budgeting activities ● Collaborate with external vendors and other internal analytics teams to extract insurance insight ● Conduct other ad hoc pricing analyses and risk modeling as needed We are open to hiring candidates to work out of one of the following locations: Arlington, VA, USA | New York, NY, USA | Seattle, WA, USA
US, NY, New York
The Amazon SCOT Forecasting team seeks a Senior Applied Scientist to join our team. Our research team conducts research into the theory and application of reinforcement learning. This research is shared in top journals and conferences and has a significant impact on the field. Through our launch of several Deep RL models into production, our work also affects decision making in the real world. Members of our group have varied interests—from the mathematical foundations of reinforcement learning, to language modeling, to maintaining the performance of generative models in the face of copyrights, and more. Recent work has focused on sample efficiency of RL algorithms, treatment effect estimation, and RL agents integrating real-world constraints, as applied in supply chains. Previous publications include: - Linear Reinforcement Learning with Ball Structure Action Space - Meta-Analysis of Randomized Experiments with Applications to Heavy-Tailed Response Data - A Few Expert Queries Suffices for Sample-Efficient RL with Resets and Linear Value Approximation - Deep Inventory Management - What are the Statistical Limits of Offline RL with Linear Function Approximation? Working collaboratively with a group of fellow scientists and engineers, you will identify complex problems and develop solutions in the RL space. We encourage collaboration across teammates and their areas of specialty, leading to creative and ambitious projects with the goal of publication and production. Key job responsibilities - Drive collaborative research and creative problem solving - Constructively critique peer research; mentor junior scientists - Create experiments and prototype implementations of new algorithms and techniques - Collaborate with engineering teams to design and implement software built on these new algorithms - Contribute to progress of the Amazon and broader research communities by producing publications We are open to hiring candidates to work out of one of the following locations: New York, NY, USA
US, CA, Virtual Location - California
If you are interested in this position, please apply on Twitch's Career site https://www.twitch.tv/jobs/en/ About Us: Launched in 2011, Twitch is a global community that comes together each day to create multiplayer entertainment: unique, live, unpredictable experiences created by the interactions of millions. We bring the joy of co-op to everything, from casual gaming to world-class esports to anime marathons, music, and art streams. Twitch also hosts TwitchCon, where we bring everyone together to celebrate and grow their personal interests and passions. We're always live at Twitch. About the Role: As a Data Scientist, Analytics member of the Data Platform - Insights team, you'll provide data analysis and support for platform, service, and operational engineering teams at Twitch, shaping the way success is measured. Defining what questions should be asked and scaling analytics methods and tools to support our growing business. Additionally, you will help support the vision for business analytics, solutions architecture for data related business constructs, as well as tactical execution such as experiment analysis and campaign performance reporting. You are paving the way for high-quality, high-velocity decisions and will report to the Manager, Data Science. For this role, we're looking for an experienced data staff who will oversee data instrumentation, dashboard/report building, metrics reviews, inform team investments, guidance on success/failure metrics and ad-hoc analysis. You will also work with technical and non-technical staff members throughout the company, and your effort will have an impact on hundreds of partners at Twitch You Will: - Work with members of Platforms & Services to guide them towards better decision making from the available data. - Promote data knowledge and insights through managing communications with partners and other teams, collaborate with colleagues to complete data projects and ensure all parties can use the insights to further improve. - Maintain a customer-centric focus while being a domain and product expert through data, develop trust amongst peers, and ensure that the teams and programs have access to data to make decisions - Manage ambiguous problems and adapt tools to answer complicated questions. - Identify the trade-offs between speed and quality of different approaches. - Create analytical frameworks to measure team success by partnering with teams to establish success metrics, create approaches to track the data and troubleshoot errors, measure and evaluate the data to develop a common language for all colleagues to understand these metrics. - Operationalize data processes to provide partners with ad-hoc analysis, automated dashboards, and self-service reporting tools so that everyone gets a good sense of the state of the business Perks: - Medical, Dental, Vision & Disability Insurance - 401(k), Maternity & Parental Leave - Flexible PTO - Commuter Benefits - Amazon Employee Discount - Monthly Contribution & Discounts for Wellness Related Activities & Programs (e.g., gym memberships, off-site massages), -Breakfast, Lunch & Dinner Served Daily - Free Snacks & Beverages We are open to hiring candidates to work out of one of the following locations: Irvine, CA, USA | Seattle, WA, USA | Virtual Location - CA
US, WA, Bellevue
Have you ever ordered a product on Amazon and when that box with the smile arrived you wondered how it got to you so fast? Have you wondered where it came from and how much it cost Amazon to deliver it to you? Have you also wondered what are different ways that the transportation assets can be used to delight the customer even more. If so, the Amazon transportation Services, Product and Science is for you . We manage the delivery of tens of millions of products every week to Amazon’s customers, achieving on-time delivery in a cost-effective manner. We are looking for an enthusiastic, customer obsessed Applied Scientist with strong scientific thinking, good software and statistics experience, skills to help manage projects and operations, improve metrics, and develop scalable processes and tools. The primary role of an Applied Scientist within Amazon is to address business challenges through building a compelling case, and using data to influence change across the organization. This individual will be given responsibility on their first day to own those business challenges and the autonomy to think strategically and make data driven decisions. Decisions and tools made in this role will have significant impact to the customer experience, as it will have a major impact on how we operate the middle mile network. Ideal candidates will be a high potential, strategic and analytic graduate with a PhD in (Operations Research, Statistics, Engineering, and Supply Chain) ready for challenging opportunities in the core of our world class operations space. Great candidates have a history of operations research, machine learning , and the ability to use data and research to make changes. This role requires robust skills in research and implementation of scalable products and models . This individual will need to be able to work with a team, but also be comfortable making decisions independently, in what is often times an ambiguous environment. Responsibilities may include: - Develop input and assumptions based preexisting models to estimate the costs and savings opportunities associated with varying levels of network growth and operations - Creating metrics to measure business performance, identify root causes and trends, and prescribe action plans - Managing multiple projects simultaneously - Working with technology teams and product managers to develop new tools and systems to support the growth of the business - Communicating with and supporting various internal stakeholders and external audiences We are open to hiring candidates to work out of one of the following locations: Bellevue, WA, USA
US, CA, Los Angeles
The Alexa team is looking for a passionate, talented, and inventive Applied Scientist with a strong machine learning background, to help build industry-leading Speech and Language technology. Key job responsibilities As an Applied Scientist with the Alexa team, you will work with talented peers to develop novel algorithms and modeling techniques to advance the state of the art in spoken language understanding. Your work will directly impact our customers in the form of products and services that make use of speech and language technology. You will leverage Amazon’s heterogeneous data sources and large-scale computing resources to accelerate advances in spoken language understanding. About the team The Alexa team has a mission to push the envelope in Automatic Speech Recognition (ASR), Natural Language Understanding (NLU), and Audio Signal Processing, in order to provide the best-possible experience for our customers. We are open to hiring candidates to work out of one of the following locations: Los Angeles, CA, USA
US, WA, Seattle
Are you fascinated by the power of Natural Language Processing (NLP) and Large Language Models (LLM) to transform the way we interact with technology? Are you passionate about applying advanced machine learning techniques to solve complex challenges in the e-commerce space? If so, Amazon's International Seller Services team has an exciting opportunity for you as an Applied Scientist. At Amazon, we strive to be Earth's most customer-centric company, where customers can find and discover anything they want to buy online. Our International Seller Services team plays a pivotal role in expanding the reach of our marketplace to sellers worldwide, ensuring customers have access to a vast selection of products. As an Applied Scientist, you will join a talented and collaborative team that is dedicated to driving innovation and delivering exceptional experiences for our customers and sellers. You will be part of a global team that is focused on acquiring new merchants from around the world to sell on Amazon’s global marketplaces around the world. The position is based in Seattle but will interact with global leaders and teams in Europe, Japan, China, Australia, and other regions. Join us at the Central Science Team of Amazon's International Seller Services and become part of a global team that is redefining the future of e-commerce. With access to vast amounts of data, cutting-edge technology, and a diverse community of talented individuals, you will have the opportunity to make a meaningful impact on the way sellers engage with our platform and customers worldwide. Together, we will drive innovation, solve complex problems, and shape the future of e-commerce. Please visit https://www.amazon.science for more information Key job responsibilities - Apply your expertise in LLM models to design, develop, and implement scalable machine learning solutions that address complex language-related challenges in the international seller services domain. - Collaborate with cross-functional teams, including software engineers, data scientists, and product managers, to define project requirements, establish success metrics, and deliver high-quality solutions. - Conduct thorough data analysis to gain insights, identify patterns, and drive actionable recommendations that enhance seller performance and customer experiences across various international marketplaces. - Continuously explore and evaluate state-of-the-art NLP techniques and methodologies to improve the accuracy and efficiency of language-related systems. - Communicate complex technical concepts effectively to both technical and non-technical stakeholders, providing clear explanations and guidance on proposed solutions and their potential impact. We are open to hiring candidates to work out of one of the following locations: Seattle, WA, USA
US, CA, Palo Alto
We’re working to improve shopping on Amazon using the conversational capabilities of large language models. We are open to hiring candidates to work out of one of the following locations: Palo Alto, CA, USA