Generative AI systems like foundation models (FMs) must align well with human values to ensure their behavior is helpful and trustworthy. While Reinforcement Learning from Human Feedback (RLHF) has shown promise for optimizing model performance using human judgments, existing RLHF pipelines predominantly rely on immediate feedback, which can fail to accurately reflect the downstream impact of an interaction on users' utility. We demonstrate that feedback based on evaluators' foresight estimates of downstream consequences systematically induces Goodhart's Law dynamics, incentivizing misaligned behaviors like sycophancy and deception and ultimately degrading user outcomes. To alleviate this, we propose decoupling evaluation from prediction by refocusing RLHF on hindsight feedback. Our theoretical analysis reveals that conditioning evaluator feedback on downstream observations mitigates misalignment and improves expected human utility, even when these observations are simulated by the AI system itself. To leverage this insight in a practical alignment algorithm, we introduce Reinforcement Learning from Hindsight Simulation (RLHS), which first simulates plausible consequences and then elicits feedback to assess what behaviors were genuinely beneficial in hindsight. We apply RLHS to two widely-employed online and offline preference optimization methods -- Proximal Policy Optimization (PPO) and Direct Preference Optimization (DPO) -- and show empirically that misalignment is significantly reduced with both methods. Through an online human user study, we show that RLHS consistently outperforms RLHF in helping users achieve their goals and earns higher satisfaction ratings, despite being trained solely with simulated hindsight feedback. These results underscore the importance of focusing on long-term consequences, even simulated ones, to mitigate misalignment in RLHF.

An email written by our team members has a 65% satisfaction rating from customers,

Since its inception in 2016, the Alexa Prize program has enabled hundreds of university students to explore and compete to develop conversational agents through the SocialBot Grand Challenge. The goal of the challenge is to build agents capable of conversing coherently and engagingly with humans on popular topics for 20 minutes, while achieving an average rating of at least 4.0/5.0. However, as conversational agents attempt to assist users with increasingly complex tasks, new conversational AI techniques and evaluation platforms are needed. The Alexa Prize TaskBot challenge, established in 2021, builds on the success of the SocialBot challenge by introducing the requirements of interactively assisting humans with real-world Cooking and Do-It-Yourself tasks, while making use of both voice and visual modalities. This challenge requires the TaskBots to identify and understand the user's need, identify and integrate task and domain knowledge into the interaction, and develop new ways of engaging the user without distracting them from the task at hand, among other challenges. This paper provides an overview of the TaskBot challenge, describes the infrastructure support provided to the teams with the CoBot Toolkit, and summarizes the approaches the participating teams took to overcome the research challenges.

Dialogue level quality estimation is vital for optimizing data driven dialogue management. Current automated methods to estimate turn and dialogue level user satisfaction employ hand-crafted features and rely on complex annotation schemes, which reduce the generalizability of the trained models. We propose a novel user satisfaction estimation approach which minimizes an adaptive multi-task loss function in order to jointly predict turn-level Response Quality labels provided by experts and explicit dialogue-level ratings provided by end users. The proposed BiLSTM based deep neural net model automatically weighs each turn's contribution towards the estimated dialogue-level rating, implicitly encodes temporal dependencies, and removes the need to hand-craft features. On dialogues sampled from 28 Alexa domains, two dialogue systems and three user groups, the joint dialogue-level satisfaction estimation model achieved up to an absolute 27% (0.43->0.70) and 7% (0.63->0.70) improvement in linear correlation performance over baseline deep neural net and benchmark Gradient boosting regression models, respectively.

Computing with words (CWW) has emerged as a powerful tool for processing the linguistic information, especially the one generated by human beings. Various CWW approaches have emerged since the inception of CWW, such as perceptual computing, extension principle based CWW approach, symbolic method based CWW approach, and 2-tuple based CWW approach. Furthermore, perceptual computing can use interval approach (IA), enhanced interval approach (EIA), or Hao-Mendel approach (HMA), for data processing. There have been numerous works in which HMA was shown to be better at word modelling than EIA, and EIA better than IA. But, a deeper study of these works reveals that HMA captures lesser fuzziness than the EIA or IA. Thus, we feel that EIA is more suited for word modelling in multi-person systems and HMA for single-person systems (as EIA is an improvement over IA). Furthermore, another set of works, compared the performances perceptual computing to the other above said CWW approaches. In all these works, perceptual computing was shown to be better than other CWW approaches. However, none of the works tried to investigate the reason behind this observed better performance of perceptual computing. Also, no comparison has been performed for scenarios where the inputs are differentially weighted. Thus, the aim of this work is to empirically establish that EIA is suitable for multi-person systems and HMA for single-person systems. Another dimension of this work is also to empirically prove that perceptual computing gives better performance than other CWW approaches based on extension principle, symbolic method and 2-tuple especially in scenarios where inputs are differentially weighted.

An automated metric to evaluate dialogue quality is vital for optimizing data driven dialogue management. The common approach of relying on explicit user feedback during a conversation is intrusive and sparse. Current models to estimate user satisfaction use limited feature sets and employ annotation schemes with limited generalizability to conversations spanning multiple domains. To address these gaps, we created a new Response Quality annotation scheme, introduced five new domain-independent feature sets and experimented with six machine learning models to estimate User Satisfaction at both turn and dialogue level. Response Quality ratings achieved significantly high correlation (0.76) with explicit turn-level user ratings. Using the new feature sets we introduced, Gradient Boosting Regression model achieved best (rating [1-5]) prediction performance on 26 seen (linear correlation ~0.79) and one new multi-turn domain (linear correlation 0.67). We observed a 16% relative improvement (68% -> 79%) in binary ("satisfactory/dissatisfactory") class prediction accuracy of a domain-independent dialogue-level satisfaction estimation model after including predicted turn-level satisfaction ratings as features.

An automated metric to evaluate dialogue quality is vital for optimizing data driven dialogue management. The common approach of relying on explicit user feedback during a conversation is intrusive and sparse. Current models to estimate user satisfaction use limited feature sets and rely on annotation schemes with low inter-rater reliability, limiting generalizability to conversations spanning multiple domains. To address these gaps, we created a new Response Quality annotation scheme, based on which we developed turn-level User Satisfaction metric. We introduced five new domain-independent feature sets and experimented with six machine learning models to estimate the new satisfaction metric. Using Response Quality annotation scheme, across randomly sampled single and multi-turn conversations from 26 domains, we achieved high inter-annotator agreement (Spearman's rho 0.94). The Response Quality labels were highly correlated (0.76) with explicit turn-level user ratings. Gradient boosting regression achieved best correlation of ~0.79 between predicted and annotated user satisfaction labels. Multi Layer Perceptron and Gradient Boosting regression models generalized to an unseen domain better (linear correlation 0.67) than other models. Finally, our ablation study verified that our novel features significantly improved model performance.

The supply of able ML designers has yet to catch up to this demand. A major reason for this is that ML is just plain tricky. This tutorial introduces the basics of Machine Learning theory, laying down the common themes and concepts, making it easy to follow the logic and get comfortable with the topic. So what exactly is "machine learning" anyway? ML is actually a lot of things. The field is quite vast and is expanding rapidly, being continually partitioned and sub-partitioned ad nauseam into different sub-specialties and types of machine learning.

Machine Learning (ML) is coming into its own, with a growing recognition that ML can play a key role in a wide range of critical applications, such as data mining, natural language processing, image recognition, and expert systems. ML provides potential solutions in all these domains and more, and is set to be a pillar of our future civilization. The supply of able ML designers has yet to catch up to this demand. A major reason for this is that ML is just plain tricky. This tutorial introduces the basics of Machine Learning theory, laying down the common themes and concepts, making it easy to follow the logic and get comfortable with the topic. So what exactly is "machine learning" anyway?

No discussion of ML would be complete without at least mentioning neural networks. Not only do neural nets offer an extremely powerful tool to solve very tough problems, but they also offer fascinating hints at the workings of our own brains, and intriguing possibilities for one day creating truly intelligent machines. Neural networks are well suited to machine learning problems where the number of inputs is gigantic. The computational cost of handling such a problem is just too overwhelming for the types of systems we've discussed above. As it turns out, however, neural networks can be effectively tuned using techniques that are strikingly similar to gradient descent in principle. A thorough discussion of neural networks is beyond the scope of this tutorial, but I recommend checking out our previous post on the subject.