Despite advances in large language models (LLMs) on reasoning and instruction-following benchmarks, it is unclear whether they can reliably produce outputs aligned with a variety of user goals, a concept called steerability. We highlight two gaps in current LLM evaluations for assessing steerability. First, many benchmarks are built with past LLM chats and text scraped from the Internet, which may skew towards common requests, underrepresenting less-common requests by potential users. Second, prior work measures performance as a scalar, which could conceal behavioral shifts in LLM outputs in open-ended generation. To mitigate these gaps, we introduce a framework based on a multi-dimensional goal space that models user goals and LLM outputs as vectors with dimensions corresponding to text attributes (e.g., reading difficulty). Applied to a text-rewriting task, we find that current LLMs induce intended changes or "side-effects" to text attributes, impeding steerability. Interventions to improve steerability, such as prompt engineering, best-of-N sampling, and reinforcement learning fine-tuning, have varying effectiveness, yet side effects remain problematic. Our findings suggest that even strong LLMs struggle with steerability, and existing alignment strategies may be insufficient.

Despite their many successes, transformer-based large language models (LLMs) continue to struggle with tasks that require complex reasoning over large parts of their input. We argue that these failures arise due to capacity limits on the accurate flow of information within LLMs. To formalize this issue, we introduce the bounded attention prefix oracle (BAPO) model, a new computational framework that models bandwidth constraints on attention heads, the mechanism for internal communication in LLMs. We show that several important reasoning problems like graph reachability require high communication bandwidth for BAPOs to solve; we call these problems BAPO-hard. Our experiments corroborate our theoretical predictions: GPT-4o, Claude, and Gemini succeed on BAPO-easy tasks and fail even on relatively small BAPO-hard tasks. BAPOs also reveal another benefit of chain of thought (CoT): we prove that breaking down a task using CoT can turn any BAPO-hard problem into a BAPO-easy one. Our results offer principled explanations for key LLM failures and suggest directions for architectures and inference methods that mitigate bandwidth limits.

We explore the use of Large Language Model (LLM-based) chatbots to power recommender systems. We observe that the chatbots respond poorly when they encounter under-specified requests (e.g., they make incorrect assumptions, hedge with a long response, or refuse to answer). We conjecture that such miscalibrated response tendencies (i.e., conversational priors) can be attributed to LLM fine-tuning by annotators — single-turn annotations may not capture multi-turn conversation utility, and the annotators’ preferences may not even be representative of users interacting with a recommender system. We first analyze public LLM chat logs to conclude that query under-specification is common. Next, we study synthetic recommendation problems with known but latent item utilities, and frame them as Partially Observed Decision Processes (PODP). We find that pre-trained LLMs can be sub-optimal for PODPs and derive better policies that clarify under-specified queries when appropriate. Then, we re-calibrate LLMs by prompting them with learned control messages to approximate the improved policy. Finally, we show empirically that our lightweight learning approach effectively uses logged conversation data to re-calibrate the response strategies of LLM-based chatbots for recommendation tasks.

A significant proportion of machine learning methodologies for recommendation systems are grounded in the fundamental principle of matching, utilizing perceptual and similarity-based learning approaches. These methods include both the extraction of features from data through representation learning and the derivation of similarity matching functions via neural function learning. While these models are important for recommendation systems, their foundational design philosophy primarily captures correlational signals within the data. Transitioning from correlation-based learning to causal learning in recommendation systems represents a critical area to explore, as causal models enable extrapolation beyond observational data in both representation learning and ranking tasks. Specifically, causal learning offers potential enhancements to the recommender system community across multiple dimensions, including, but not limited to, explainable, unbiased, fairness-aware, robust, and cognitive reasoning models for recommendation. This special issue is dedicated to exploring the research and practical applications of causal inference within the realms of recommendation and broader ranking scenarios. It has attracted interest from an array of researchers and practitioners on disseminating the latest developments in causal modeling for recommender systems. Moreover, it has attracted the interest of professionals from various fields such as Information Retrieval, Machine Learning, Artificial Intelligence, Natural Language Processing, Data Science, and others.

In recent years, neural models have been repeatedly touted to exhibit state-of-the-art performance in recommendation. Nevertheless, multiple recent studies have revealed that the reported state-of-the-art results of many neural recommendation models cannot be reliably replicated. A primary reason is that existing evaluations are performed under various inconsistent protocols. Correspondingly, these replicability issues make it difficult to understand how much benefit we can actually gain from these neural models. It then becomes clear that a fair and comprehensive performance comparison between traditional and neural models is needed.

The dynamic nature of AI technologies makes testing human-AI interaction and collaboration challenging – especially before such features are deployed in the wild. This presents a challenge for designers and AI practitioners as early feedback for iteration is often unavailable in the development phase. In this paper, we take inspiration from integration testing concepts in software development and present HINT (Human-AI INtegration Testing), a crowd-based framework for testing AI-based experiences integrated with a humans-in-the-loop workflow. HINT supports early testing of AI-based features within the context of realistic user tasks and makes use of successive sessions to simulate AI experiences that evolve over-time. Finally, it provides practitioners with reports to evaluate and compare aspects of these experiences.

Abstract In the summarization domain, a key requirement for summaries is to be factually consistent with the input document. Previous work has found that natural language inference (NLI) models do not perform competitively when applied to inconsistency detection. In this work, we revisit the use of NLI for inconsistency detection, finding that past work suffered from a mismatch in input granularity between NLI datasets (sentence-level), and inconsistency detection (document level). We provide a highly effective and light-weight method called SummaCConv that enables NLI models to be successfully used for this task by segmenting documents into sentence units and aggregating scores between pairs of sentences. We furthermore introduce a new benchmark called SummaC (Summary Consistency) which consists of six large inconsistency detection datasets. On this dataset, SummaCConv obtains state-of-the-art results with a balanced accuracy of 74.4%, a 5% improvement compared with prior work.

Philippe Laban, Tobias Schnabel, Paul Bennett, Marti A. Hearst. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

In many domains, user preferences are similar locally within like-minded subgroups of users, but typically differ globally between those subgroups. Local recommendation models were shown to substantially improve top-K recommendation performance in such settings. However, existing local models do not scale to large-scale datasets with an increasing number of subgroups and do not support inductive recommendations for users not appearing in the training set. Key reasons for this are that subgroup detection and recommendation get implemented as separate steps in the model or that local models are explicitly instantiated for each subgroup. In this paper, we propose an End-to-end Local Factor Model (Elfm) which overcomes these limitations by combining both steps and incorporating local structures through an inductive bias. Our model can be optimized end-to-end and supports incremental inference, does not require a full separate model for each subgroup, and has overall small memory and computational costs for incorporating local structures. Empirical results show that our method substantially improves recommendation performance on large-scale datasets with millions of users and items with considerably smaller model size. Our user study also shows that our approach produces coherent item subgroups which could aid in the generation of explainable recommendations.

Item-to-item recommendation (e.g., “People who like this also like...”) is a ubiquitous and important type of recommendation in real-world systems. Observational data from historical interaction logs abound in these settings. However, since virtually all observational data exhibit biases, such as time-in-inventory or interface biases, it is crucial that recommender algorithms account for these biases. In this paper, we develop a principled approach for item-to-item recommendation based on causal inference and present a practical and highly effective method for estimating the causal parameters from a small annotated dataset. Empirically, we find that our approach substantially improves upon existing methods while requiring only small amounts of annotated data.

Real-world recommender systems often allow users to adjust the presented content through a variety of preference elicitation techniques such as “liking” or interest profiles. These elicitation techniques trade-off time and effort to users with the richness of the signal they provide to learning component driving the recommendations. In this paper, we explore this trade-off, seeking new ways for people to express their preferences with the goal of improving communication channels between users and the recommender system. Through a need-finding study, we observe the patterns in how people express their preferences during curation task, propose a taxonomy for organizing them, and point out research opportunities. We present a case study that illustrates how using this taxonomy to design an onboarding experience can lead to more accurate machine-learned recommendations while maintaining user satisfaction under low effort.

Many interactive online systems, such as social media platforms or news sites, provide personalized experiences through recommendations or news feed customization based on people's feedback and engagement on individual items (e.g., liking items). In this paper, we investigate how we can support a greater degree of user control in such systems by changing the way the system allows people to gauge the consequences of their feedback actions. To this end, we consider two important aspects of how the system responds to feedback actions: (i) immediacy, i.e., how quickly the system responds with an update, and (ii) visibility, i.e., whether or not changes will get highlighted. We used both an in-lab qualitative study and a large-scale crowd-sourced study to examine the impact of these factors on people's reported preferences and observed behavioral metrics. We demonstrate that UX design which enables people to preview the impact of their actions and highlights changes results in a higher reported transparency, an overall preference for this design, and a greater selectivity in which items are liked.

Instrumental variable analysis is a powerful tool for estimating causal effects when randomization or full control of confounders is not possible. The application of standard methods such as 2SLS, GMM, and more recent variants are significantly impeded when the causal effects are complex, the instruments are high-dimensional, and/or the treatment is high-dimensional. In this paper, we propose the DeepGMM algorithm to overcome this. Our algorithm is based on a new variational reformulation of GMM with optimal inverse-covariance weighting that allows us to efficiently control very many moment conditions. We further develop practical techniques for optimization and model selection that make it particularly successful in practice. Our algorithm is also computationally tractable and can handle large-scale datasets. Numerical results show our algorithm matches the performance of the best tuned methods in standard settings and continues to work in high-dimensional settings where even recent methods break.

Recommender systems rely heavily on the predictive accuracy of the learning algorithm. Most work on improving accuracy has focused on the learning algorithm itself. We argue that this algorithmic focus is myopic. In particular, since learning algorithms generally improve with more and better data, we propose shaping the feedback generation process as an alternate and complementary route to improving accuracy. To this effect, we explore how changes to the user interface can impact the quality and quantity of feedback data -- and therefore the learning accuracy. Motivated by information foraging theory, we study how feedback quality and quantity are influenced by interface design choices along two axes: information scent and information access cost. We present a user study of these interface factors for the common task of picking a movie to watch, showing that these factors can effectively shape and improve the implicit feedback data that is generated while maintaining the user experience.

Implicit feedback (e.g., clicks, dwell times, etc.) is an abundant source of data in human-interactive systems. While implicit feedback has many advantages (e.g., it is inexpensive to collect, user-centric, and timely), its inherent biases are a key obstacle to its effective use. For example, position bias in search rankings strongly influences how many clicks a result receives, so that directly using click data as a training signal in Learning-to-Rank (LTR) methods yields sub-optimal results. To overcome this bias problem, we present a counterfactual inference framework that provides the theoretical basis for unbiased LTR via Empirical Risk Minimization despite biased data. Using this framework, we derive a propensity-weighted ranking SVM for discriminative learning from implicit feedback, where click models take the role of the propensity estimator. Beyond the theoretical support, we show empirically that the proposed learning method is highly effective in dealing with biases, that it is robust to noise and propensity model mis-specification, and that it scales efficiently. We also demonstrate the real-world applicability of our approach on an operational search engine, where it substantially improves retrieval performance.

Any learning algorithm for recommendation faces a fundamental trade-off between exploiting partial knowledge of a user»s interests to maximize satisfaction in the short term and discovering additional user interests to maximize satisfaction in the long term. To enable discovery, a machine learning algorithm typically elicits feedback on items it is uncertain about, which is termed algorithmic exploration in machine learning. This exploration comes with a cost to the user, since the items an algorithm chooses for exploration frequently turn out to not match the user»s interests. In this paper, we study how users tolerate such exploration and how presentation strategies can mitigate the exploration cost. To this end, we conduct a behavioral study with over 600 people, where we vary how algorithmic exploration is mixed into the set of recommendations. We find that users respond non-linearly to the amount of exploration, where some exploration mixed into the set of recommendations has little effect on short-term satisfaction and behavior. For long-term satisfaction, the overall goal is to learn via exploration about the items presented. We therefore also analyze the quantity and quality of implicit feedback signals such as clicks and hovers, and how they vary with different amounts of mix-in exploration. Our findings provide insights into how to design presentation strategies for algorithmic exploration in interactive recommender systems, mitigating the short-term costs of algorithmic exploration while aiming to elicit informative feedback data for learning.

Accurately evaluating new policies (e.g. ad-placement models, ranking functions, recommendation functions) is one of the key prerequisites for improving interactive systems. While the conventional approach to evaluation relies on online A/B tests, recent work has shown that counterfactual estimators can provide an inexpensive and fast alternative, since they can be applied offline using log data that was collected from a different policy fielded in the past. In this paper, we address the question of how to estimate the performance of a new target policy when we have log data from multiple historic policies. This question is of great relevance in practice, since policies get updated frequently in most online systems. We show that naively combining data from multiple logging policies can be highly suboptimal. In particular, we find that the standard Inverse Propensity Score (IPS) estimator suffers especially when logging and target policies diverge -- to a point where throwing away data improves the variance of the estimator. We therefore propose two alternative estimators which we characterize theoretically and compare experimentally. We find that the new estimators can provide substantially improved estimation accuracy.

Implicit feedback (e.g., clicks, dwell times, etc.) is an abundant source of data in human-interactive systems. While implicit feedback has many advantages (e.g., it is inexpensive to collect, user centric, and timely), its inherent biases are a key obstacle to its effective use. For example, position bias in search rankings strongly influences how many clicks a result receives, so that directly using click data as a training signal in Learning-to-Rank (LTR) methods yields sub-optimal results. To overcome this bias problem, we present a counterfactual inference framework that provides the theoretical basis for unbiased LTR via Empirical Risk Minimization despite biased data. Using this framework, we derive a Propensity-Weighted Ranking SVM for discriminative learning from implicit feedback, where click models take the role of the propensity estimator. In contrast to most conventional approaches to de-biasing the data using click models, this allows training of ranking functions even in settings where queries do not repeat. Beyond the theoretical support, we show empirically that the proposed learning method is highly effective in dealing with biases, that it is robust to noise and propensity model misspecification, and that it scales efficiently. We also demonstrate the real-world applicability of our approach on an operational search engine, where it substantially improves retrieval performance.

Most data for evaluating and training recommender systems is subject to selection biases, either through self-selection by the users or through the actions of the recommendation system itself. In this paper, we provide a principled approach to handle selection biases by adapting models and estimation techniques from causal inference. The approach leads to unbiased performance estimators despite biased data, and to a matrix factorization method that provides substantially improved prediction performance on real-world data. We theoretically and empirically characterize the robustness of the approach, and find that it is highly practical and scalable.

In this paper, we study shortlists as an interface component for recommender systems with the dual goal of supporting the user's decision process, as well as improving implicit feedback elicitation for increased recommendation quality. A shortlist is a temporary list of candidates that the user is currently considering, e.g., a list of a few movies the user is currently considering for viewing. From a cognitive perspective, shortlists serve as digital short-term memory where users can off-load the items under consideration -- thereby decreasing their cognitive load. From a machine learning perspective, adding items to the shortlist generates a new implicit feedback signal as a by-product of exploration and decision making which can improve recommendation quality. Shortlisting therefore provides additional data for training recommendation systems without the increases in cognitive load that requesting explicit feedback would incur.

We present a comprehensive study of evaluation methods for unsupervised embedding techniques that obtain meaningful representations of words from text.Different evaluations result in different orderings of embedding methods, calling into question the common assumption that there is one single optimal vector representation.We present new evaluation techniques that directly compare embeddings with respect to specific queries.These methods reduce bias, provide greater insight, and allow us to solicit data-driven relevance judgments rapidly and accurately through crowdsourcing.

We propose online unsupervised domain adaptation (DA), which is performed incrementally as data comes in and is applicable when batch DA is not possible.In a part-of-speech (POS) tagging evaluation, we find that online unsupervised DA performs as well as batch DA.

We present FLORS, a new part-of-speech tagger for domain adaptation. FLORS uses robust representations that work especially well for unknown words and for known words with unseen tags. FLORS is simpler and faster than previous domain adaptation methods, yet it has significantly better accuracy than several baselines.

Coactive Learning is a model of interaction between a learning system (e.g. search engine) and its human users, wherein the system learns from (typically implicit) user feedback during operational use. User feedback takes the form of preferences, and recent work has introduced online algorithms that learn from this weak feedback. However, we show that these algorithms can be unstable and ineffective in real-world settings where biases and noise in the feedback are significant. In this paper, we propose the first coactive learning algorithm that can learn robustly despite bias and noise. In particular, we explore how presenting users with slightly perturbed objects (e.g., rankings) can stabilize the learning process. We theoretically validate the algorithm by proving bounds on the average regret. We also provide extensive empirical evidence on benchmarks and from a live search engine user study, showing that the new algorithm substantially outperforms existing methods.