Ruhi Sarikaya

Shubhashis Roy Dipta, Daniel Bis, Kun Zhou et al.

Conversational assistants powered by large language models (LLMs) excel at tool-use tasks but struggle with adhering to complex, business-specific rules. While models can reason over business rules provided in context, including all policies for every query introduces high latency and wastes compute. Furthermore, these lengthy prompts lead to long contexts, harming overall performance due to the "needle-in-the-haystack" problem. To address these challenges, we propose a multi-stage alignment method that teaches models to recall and apply relevant business policies during chain-of-thought reasoning at inference time, without including the full business policy in-context. Furthermore, we introduce a novel PolicyRecall reward based on the Jaccard score and a Hallucination Penalty for GRPO training. Altogether, our best model outperforms the baseline by 16 points and surpasses comparable in-context baselines of similar model size by 3 points, while using 40% fewer words.

Revanth Gangi Reddy, Tanay Dixit, Jiaxin Qin et al.

Wikipedia, a vast and continuously consulted knowledge base, faces significant challenges in maintaining up-to-date content due to its reliance on manual human editors. Inspired by the vision of continuous knowledge acquisition in NELL and fueled by advances in LLM-based agents, this paper introduces WiNELL, an agentic framework for continuously updating Wikipedia articles. Our approach employs a multi-agent framework to aggregate online information, select new and important knowledge for a target entity in Wikipedia, and then generate precise edit suggestions for human review. Our fine-grained editing models, trained on Wikipedia's extensive history of human edits, enable incorporating updates in a manner consistent with human editing behavior. Our editor models outperform both open-source instruction-following baselines and closed-source LLMs (e.g., GPT-4o) in key information coverage and editing efficiency. End-to-end evaluation on high-activity Wikipedia pages demonstrates WiNELL's ability to identify and suggest timely factual updates. This opens up a promising research direction in LLM agents for automatically updating knowledge bases in a never-ending fashion.

Zhenhailong Wang, Jiateng Liu, Amin Fazel et al.

Modern conversational agents like ChatGPT and Alexa+ rely on predefined policies specifying metadata, response styles, and tool-usage rules. As these LLM-based systems expand to support diverse business and user queries, such policies, often implemented as in-context prompts, are becoming increasingly complex and lengthy, making faithful adherence difficult and imposing large fixed computational costs. With the rise of multimodal agents, policies that govern visual and multimodal behaviors are critical but remain understudied. Prior prompt-compression work mainly shortens task templates and demonstrations, while existing policy-alignment studies focus only on text-based safety rules. We introduce Multimodal Policy Internalization (MPI), a new task that internalizes reasoning-intensive multimodal policies into model parameters, enabling stronger policy-following without including the policy during inference. MPI poses unique data and algorithmic challenges. We build two datasets spanning synthetic and real-world decision-making and tool-using tasks and propose TriMPI, a three-stage training framework. TriMPI first injects policy knowledge via continual pretraining, then performs supervised finetuning, and finally applies PolicyRollout, a GRPO-style reinforcement learning extension that augments rollouts with policy-aware responses for grounded exploration. TriMPI achieves notable gains in end-to-end accuracy, generalization, and robustness to forgetting. As the first work on multimodal policy internalization, we provide datasets, training recipes, and comprehensive evaluations to foster future research. Project page: https://mikewangwzhl.github.io/TriMPI.

Jiateng Liu, Zhenhailong Wang, Xiaojiang Huang et al.

Large Language Model (LLM)-based agentic systems rely on in-context policy documents encoding diverse business rules. As requirements grow, these documents expand rapidly, causing high computational overhead. This motivates developing internalization methods that embed policy documents into model priors while preserving performance. Prior prompt compression work targets generic prompts, but agentic policy documents span multiple complexity levels and require deeper reasoning, making internalization harder. We introduce CC-Gen, an agentic benchmark generator with Controllable Complexity across four levels, enabling systematic evaluation of agents' ability to handle complexity and offering a unified framework for assessing policy internalization. Our analysis shows that complex policy specifications governing workflows pose major reasoning challenges. Supporting internalization with gold user agent interaction trajectories containing chain-of-thought (CoT) annotations via supervised fine-tuning (SFT) is data-intensive and degrades sharply as policy complexity increases. To mitigate data and reasoning burdens, we propose Category-Aware Policy Continued Pretraining (CAP-CPT). Our automated pipeline parses policy documents to extract key specifications, grouping them into factual, behavioral, and conditional categories, and isolating complex conditions that drive workflow complexity. This guides targeted data synthesis and enables agents to internalize policy information through an autoregressive pretraining loss. Experiments show CAP-CPT improves SFT baselines in all settings, with up to 41% and 22% gains on Qwen-3-32B, achieving 97.3% prompt length reduction on CC-Gen and further enhancing tau-Bench with minimal SFT data.

Laura Panfili, Steve Duman, Andrew Nave et al.

Grice's Cooperative Principle (1975) describes the implicit maxims that guide conversation between humans. As humans begin to interact with non-human dialogue systems more frequently and in a broader scope, an important question emerges: what principles govern those interactions? The present study addresses this question by evaluating human-AI interactions using Grice's four maxims; we demonstrate that humans do, indeed, apply these maxims to interactions with AI, even making explicit references to the AI's performance through a Gricean lens. Twenty-three participants interacted with an American English-speaking Alexa and rated and discussed their experience with an in-lab researcher. Researchers then reviewed each exchange, identifying those that might relate to Grice's maxims: Quantity, Quality, Manner, and Relevance. Many instances of explicit user frustration stemmed from violations of Grice's maxims. Quantity violations were noted for too little but not too much information, while Quality violations were rare, indicating trust in Alexa's responses. Manner violations focused on speed and humanness. Relevance violations were the most frequent, and they appear to be the most frustrating. While the maxims help describe many of the issues participants encountered, other issues do not fit neatly into Grice's framework. Participants were particularly averse to Alexa initiating exchanges or making unsolicited suggestions. To address this gap, we propose the addition of human Priority to describe human-AI interaction. Humans and AIs are not conversational equals, and human initiative takes priority. We suggest that the application of Grice's Cooperative Principles to human-AI interactions is beneficial both from an AI development perspective and as a tool for describing an emerging form of interaction.

Cheng Wang, Sungjin Lee, Sunghyun Park et al.

Real-world machine learning systems are achieving remarkable performance in terms of coarse-grained metrics like overall accuracy and F-1 score. However, model improvement and development often require fine-grained modeling on individual data subsets or slices, for instance, the data slices where the models have unsatisfactory results. In practice, it gives tangible values for developing such models that can pay extra attention to critical or interested slices while retaining the original overall performance. This work extends the recent slice-based learning (SBL)~\cite{chen2019slice} with a mixture of attentions (MoA) to learn slice-aware dual attentive representations. We empirically show that the MoA approach outperforms the baseline method as well as the original SBL approach on monitored slices with two natural language understanding (NLU) tasks.

Cheng Wang, Sun Kim, Taiwoo Park et al.

We have been witnessing the usefulness of conversational AI systems such as Siri and Alexa, directly impacting our daily lives. These systems normally rely on machine learning models evolving over time to provide quality user experience. However, the development and improvement of the models are challenging because they need to support both high (head) and low (tail) usage scenarios, requiring fine-grained modeling strategies for specific data subsets or slices. In this paper, we explore the recent concept of slice-based learning (SBL) (Chen et al., 2019) to improve our baseline conversational skill routing system on the tail yet critical query traffic. We first define a set of labeling functions to generate weak supervision data for the tail intents. We then extend the baseline model towards a slice-aware architecture, which monitors and improves the model performance on the selected tail intents. Applied to de-identified live traffic from a commercial conversational AI system, our experiments show that the slice-aware model is beneficial in improving model performance for the tail intents while maintaining the overall performance.

Han Li, Sunghyun Park, Aswarth Dara et al.

Current state-of-the-art large-scale conversational AI or intelligent digital assistant systems in industry comprises a set of components such as Automatic Speech Recognition (ASR) and Natural Language Understanding (NLU). For some of these systems that leverage a shared NLU ontology (e.g., a centralized intent/slot schema), there exists a separate skill routing component to correctly route a request to an appropriate skill, which is either a first-party or third-party application that actually executes on a user request. The skill routing component is needed as there are thousands of skills that can either subscribe to the same intent and/or subscribe to an intent under specific contextual conditions (e.g., device has a screen). Ensuring model robustness or resilience in the skill routing component is an important problem since skills may dynamically change their subscription in the ontology after the skill routing model has been deployed to production. We show how different modeling design choices impact the model robustness in the context of skill routing on a state-of-the-art commercial conversational AI system, specifically on the choices around data augmentation, model architecture, and optimization method. We show that applying data augmentation can be a very effective and practical way to drastically improve model robustness.

Sunghyun Park, Han Li, Ameen Patel et al.

Natural Language Understanding (NLU) is an established component within a conversational AI or digital assistant system, and it is responsible for producing semantic understanding of a user request. We propose a scalable and automatic approach for improving NLU in a large-scale conversational AI system by leveraging implicit user feedback, with an insight that user interaction data and dialog context have rich information embedded from which user satisfaction and intention can be inferred. In particular, we propose a general domain-agnostic framework for curating new supervision data for improving NLU from live production traffic. With an extensive set of experiments, we show the results of applying the framework and improving NLU for a large-scale production system and show its impact across 10 domains.

Dookun Park, Hao Yuan, Dongmin Kim et al.

Measuring user satisfaction level is a challenging task, and a critical component in developing large-scale conversational agent systems serving the needs of real users. An widely used approach to tackle this is to collect human annotation data and use them for evaluation or modeling. Human annotation based approaches are easier to control, but hard to scale. A novel alternative approach is to collect user's direct feedback via a feedback elicitation system embedded to the conversational agent system, and use the collected user feedback to train a machine-learned model for generalization. User feedback is the best proxy for user satisfaction, but is not available for some ineligible intents and certain situations. Thus, these two types of approaches are complementary to each other. In this work, we tackle the user satisfaction assessment problem with a hybrid approach that fuses explicit user feedback, user satisfaction predictions inferred by two machine-learned models, one trained on user feedback data and the other human annotation data. The hybrid approach is based on a waterfall policy, and the experimental results with Amazon Alexa's large-scale datasets show significant improvements in inferring user satisfaction. A detailed hybrid architecture, an in-depth analysis on user feedback data, and an algorithm that generates data sets to properly simulate the live traffic are presented in this paper.

Pragaash Ponnusamy, Alireza Roshan Ghias, Chenlei Guo et al.

Today, most large-scale conversational AI agents (e.g. Alexa, Siri, or Google Assistant) are built using manually annotated data to train the different components of the system. Typically, the accuracy of the ML models in these components are improved by manually transcribing and annotating data. As the scope of these systems increase to cover more scenarios and domains, manual annotation to improve the accuracy of these components becomes prohibitively costly and time consuming. In this paper, we propose a system that leverages user-system interaction feedback signals to automate learning without any manual annotation. Users here tend to modify a previous query in hopes of fixing an error in the previous turn to get the right results. These reformulations, which are often preceded by defective experiences caused by errors in ASR, NLU, ER or the application. In some cases, users may not properly formulate their requests (e.g. providing partial title of a song), but gleaning across a wider pool of users and sessions reveals the underlying recurrent patterns. Our proposed self-learning system automatically detects the errors, generate reformulations and deploys fixes to the runtime system to correct different types of errors occurring in different components of the system. In particular, we propose leveraging an absorbing Markov Chain model as a collaborative filtering mechanism in a novel attempt to mine these patterns. We show that our approach is highly scalable, and able to learn reformulations that reduce Alexa-user errors by pooling anonymized data across millions of customers. The proposed self-learning system achieves a win/loss ratio of 11.8 and effectively reduces the defect rate by more than 30% on utterance level reformulations in our production A/B tests. To the best of our knowledge, this is the first self-learning large-scale conversational AI system in production.

Jihwan Lee, Ruhi Sarikaya, Young-Bum Kim

In this paper, we introduce an approach for leveraging available data across multiple locales sharing the same language to 1) improve domain classification model accuracy in Spoken Language Understanding and user experience even if new locales do not have sufficient data and 2) reduce the cost of scaling the domain classifier to a large number of locales. We propose a locale-agnostic universal domain classification model based on selective multi-task learning that learns a joint representation of an utterance over locales with different sets of domains and allows locales to share knowledge selectively depending on the domains. The experimental results demonstrate the effectiveness of our approach on domain classification task in the scenario of multiple locales with imbalanced data and disparate domain sets. The proposed approach outperforms other baselines models especially when classifying locale-specific domains and also low-resourced domains.

Han Li, Jihwan Lee, Sidharth Mudgal et al.

Domain classification is the task of mapping spoken language utterances to one of the natural language understanding domains in intelligent personal digital assistants (IPDAs). This is a major component in mainstream IPDAs in industry. Apart from official domains, thousands of third-party domains are also created by external developers to enhance the capability of IPDAs. As more domains are developed rapidly, the question of how to continuously accommodate the new domains still remains challenging. Moreover, existing continual learning approaches do not address the problem of incorporating personalized information dynamically for better domain classification. In this paper, we propose CoNDA, a neural network based approach for domain classification that supports incremental learning of new classes. Empirical evaluation shows that CoNDA achieves high accuracy and outperforms baselines by a large margin on both incrementally added new domains and existing domains.

JIhwan Lee, Dongchan Kim, Ruhi Sarikaya et al.

Representation learning is an essential problem in a wide range of applications and it is important for performing downstream tasks successfully. In this paper, we propose a new model that learns coupled representations of domains, intents, and slots by taking advantage of their hierarchical dependency in a Spoken Language Understanding system. Our proposed model learns the vector representation of intents based on the slots tied to these intents by aggregating the representations of the slots. Similarly, the vector representation of a domain is learned by aggregating the representations of the intents tied to a specific domain. To the best of our knowledge, it is the first approach to jointly learning the representations of domains, intents, and slots using their hierarchical relationships. The experimental results demonstrate the effectiveness of the representations learned by our model, as evidenced by improved performance on the contextual cross-domain reranking task.

Thomas Powers, Rasool Fakoor, Siamak Shakeri et al.

Optimal selection of a subset of items from a given set is a hard problem that requires combinatorial optimization. In this paper, we propose a subset selection algorithm that is trainable with gradient-based methods yet achieves near-optimal performance via submodular optimization. We focus on the task of identifying a relevant set of sentences for claim verification in the context of the FEVER task. Conventional methods for this task look at sentences on their individual merit and thus do not optimize the informativeness of sentences as a set. We show that our proposed method which builds on the idea of unfolding a greedy algorithm into a computational graph allows both interpretability and gradient-based training. The proposed differentiable greedy network (DGN) outperforms discrete optimization algorithms as well as other baseline methods in terms of precision and recall.

Rasool Fakoor, Amanjit Kainth, Siamak Shakeri et al.

Recent advances in spoken language technologies and the introduction of many customer facing products, have given rise to a wide customer reliance on smart personal assistants for many of their daily tasks. In this paper, we present a system to reduce users' cognitive load by extending personal assistants with long-term personal memory where users can store and retrieve by voice, arbitrary pieces of information. The problem is framed as a neural retrieval based question answering system where answers are selected from previously stored user memories. We propose to directly optimize the end-to-end retrieval performance, measured by the F1-score, using reinforcement learning, leading to better performance on our experimental test set(s).

Chetan Naik, Arpit Gupta, Hancheng Ge et al.

In the slot-filling paradigm, where a user can refer back to slots in the context during a conversation, the goal of the contextual understanding system is to resolve the referring expressions to the appropriate slots in the context. In large-scale multi-domain systems, this presents two challenges - scaling to a very large and potentially unbounded set of slot values, and dealing with diverse schemas. We present a neural network architecture that addresses the slot value scalability challenge by reformulating the contextual interpretation as a decision to carryover a slot from a set of possible candidates. To deal with heterogenous schemas, we introduce a simple data-driven method for trans- forming the candidate slots. Our experiments show that our approach can scale to multiple domains and provides competitive results over a strong baseline.

Young-Bum Kim, Dongchan Kim, Anjishnu Kumar et al.

In this paper, we explore the task of mapping spoken language utterances to one of thousands of natural language understanding domains in intelligent personal digital assistants (IPDAs). This scenario is observed for many mainstream IPDAs in industry that allow third parties to develop thousands of new domains to augment built-in ones to rapidly increase domain coverage and overall IPDA capabilities. We propose a scalable neural model architecture with a shared encoder, a novel attention mechanism that incorporates personalization information and domain-specific classifiers that solves the problem efficiently. Our architecture is designed to efficiently accommodate new domains that appear in-between full model retraining cycles with a rapid bootstrapping mechanism two orders of magnitude faster than retraining. We account for practical constraints in real-time production systems, and design to minimize memory footprint and runtime latency. We demonstrate that incorporating personalization results in significantly more accurate domain classification in the setting with thousands of overlapping domains.

Young-Bum Kim, Dongchan Kim, Joo-Kyung Kim et al.

Intelligent personal digital assistants (IPDAs), a popular real-life application with spoken language understanding capabilities, can cover potentially thousands of overlapping domains for natural language understanding, and the task of finding the best domain to handle an utterance becomes a challenging problem on a large scale. In this paper, we propose a set of efficient and scalable neural shortlisting-reranking models for large-scale domain classification in IPDAs. The shortlisting stage focuses on efficiently trimming all domains down to a list of k-best candidate domains, and the reranking stage performs a list-wise reranking of the initial k-best domains with additional contextual information. We show the effectiveness of our approach with extensive experiments on 1,500 IPDA domains.

Young-Bum Kim, Sungjin Lee, Ruhi Sarikaya

In multi-turn dialogs, natural language understanding models can introduce obvious errors by being blind to contextual information. To incorporate dialog history, we present a neural architecture with Speaker-Sensitive Dual Memory Networks which encode utterances differently depending on the speaker. This addresses the different extents of information available to the system - the system knows only the surface form of user utterances while it has the exact semantics of system output. We performed experiments on real user data from Microsoft Cortana, a commercial personal assistant. The result showed a significant performance improvement over the state-of-the-art slot tagging models using contextual information.