Search

As large language models (LLMs) expand into performing as agents for real-world applications beyond traditional NLP tasks, evaluating their robustness becomes increasingly important. However, existing benchmarks often overlook critical dimensions like cultural and social awareness. To address these, we introduce CASA, a benchmark designed to assess LLM agents' sensitivity to cultural and social norms across two web-based tasks: online shopping and social discussion forums. Our approach evaluates LLM agents' ability to detect and appropriately respond to norm-violating user queries and observations. Furthermore, we propose a comprehensive evaluation framework that measures awareness coverage, helpfulness in managing user queries, and the violation rate when facing misleading web content. Experiments show that current LLMs perform significantly better in non-agent than in web-based agent environments, with agents achieving less than 10% awareness coverage and over 40% violation rates. To improve performance, we explore two methods: prompting and fine-tuning, and find that combining both methods can offer complementary advantages -- fine-tuning on culture-specific datasets significantly enhances the agents' ability to generalize across different regions, while prompting boosts the agents' ability to navigate complex tasks. These findings highlight the importance of constantly benchmarking LLM agents' cultural and social awareness during the development cycle., Comment: NAACL 2025 Findings

Despite strong advisory against it, large generative models (LMs) are already being used for decision making tasks that were previously done by predictive models or humans. We put popular LMs to the test in a high-stakes decision making task: recidivism prediction. Studying three closed-access and open-source LMs, we analyze the LMs not exclusively in terms of accuracy, but also in terms of agreement with (imperfect, noisy, and sometimes biased) human predictions or existing predictive models. We conduct experiments that assess how providing different types of information, including distractor information such as photos, can influence LM decisions. We also stress test techniques designed to either increase accuracy or mitigate bias in LMs, and find that some to have unintended consequences on LM decisions. Our results provide additional quantitative evidence to the wisdom that current LMs are not the right tools for these types of tasks.

Time series forecasting aids decision-making, especially for stakeholders who rely on accurate predictions, making it very important to understand and explain these models to ensure informed decisions. Traditional explainable AI (XAI) methods, which underline feature or temporal importance, often require expert knowledge. In contrast, natural language explanations (NLEs) are more accessible to laypeople. However, evaluating forecast NLEs is difficult due to the complex causal relationships in time series data. To address this, we introduce two new performance metrics based on simulatability, assessing how well a human surrogate can predict model forecasts using the explanations. Experiments show these metrics differentiate good from poor explanations and align with human judgments. Utilizing these metrics, we further evaluate the ability of state-of-the-art large language models (LLMs) to generate explanations for time series data, finding that numerical reasoning, rather than model size, is the main factor influencing explanation quality.

The deployment of Large Language Models (LLMs) and Large Multimodal Models (LMMs) on mobile devices has gained significant attention due to the benefits of enhanced privacy, stability, and personalization. However, the hardware constraints of mobile devices necessitate the use of models with fewer parameters and model compression techniques like quantization. Currently, there is limited understanding of quantization's impact on various task performances, including LLM tasks, LMM tasks, and, critically, trust and safety. There is a lack of adequate tools for systematically testing these models on mobile devices. To address these gaps, we introduce MobileAIBench, a comprehensive benchmarking framework for evaluating mobile-optimized LLMs and LMMs. MobileAIBench assesses models across different sizes, quantization levels, and tasks, measuring latency and resource consumption on real devices. Our two-part open-source framework includes a library for running evaluations on desktops and an iOS app for on-device latency and hardware utilization measurements. Our thorough analysis aims to accelerate mobile AI research and deployment by providing insights into the performance and feasibility of deploying LLMs and LMMs on mobile platforms.

We present a method for identifying groups of test examples -- slices -- on which a model under-performs, a task now known as slice discovery. We formalize coherence -- a requirement that erroneous predictions, within a slice, should be wrong for the same reason -- as a key property that any slice discovery method should satisfy. We then use influence functions to derive a new slice discovery method, InfEmbed, which satisfies coherence by returning slices whose examples are influenced similarly by the training data. InfEmbed is simple, and consists of applying K-Means clustering to a novel representation we deem influence embeddings. We show InfEmbed outperforms current state-of-the-art methods on 2 benchmarks, and is effective for model debugging across several case studies., Comment: NeuRIPs 2023 conference paper

Missing values are a fundamental problem in data science. Many datasets have missing values that must be properly handled because the way missing values are treated can have large impact on the resulting machine learning model. In medical applications, the consequences may affect healthcare decisions. There are many methods in the literature for dealing with missing values, including state-of-the-art methods which often depend on black-box models for imputation. In this work, we show how recent advances in interpretable machine learning provide a new perspective for understanding and tackling the missing value problem. We propose methods based on high-accuracy glass-box Explainable Boosting Machines (EBMs) that can help users (1) gain new insights on missingness mechanisms and better understand the causes of missingness, and (2) detect -- or even alleviate -- potential risks introduced by imputation algorithms. Experiments on real-world medical datasets illustrate the effectiveness of the proposed methods., Comment: Preprint of a paper accepted by CHIL 2023

Many organizations measure treatment effects via an experimentation platform to evaluate the casual effect of product variations prior to full-scale deployment. However, standard experimentation platforms do not perform optimally for end user populations that exhibit heterogeneous treatment effects (HTEs). Here we present a personalized experimentation framework, Personalized Experiments (PEX), which optimizes treatment group assignment at the user level via HTE modeling and sequential decision policy optimization to optimize multiple short-term and long-term outcomes simultaneously. We describe an end-to-end workflow that has proven to be successful in practice and can be readily implemented using open-source software., Comment: 5 pages, 2 figures

Learning heterogeneous treatment effects (HTEs) is an important problem across many fields. Most existing methods consider the setting with a single treatment arm and a single outcome metric. However, in many real world domains, experiments are run consistently - for example, in internet companies, A/B tests are run every day to measure the impacts of potential changes across many different metrics of interest. We show that even if an analyst cares only about the HTEs in one experiment for one metric, precision can be improved greatly by analyzing all of the data together to take advantage of cross-experiment and cross-outcome metric correlations. We formalize this idea in a tensor factorization framework and propose a simple and scalable model which we refer to as the low rank or LR-learner. Experiments in both synthetic and real data suggest that the LR-learner can be much more precise than independent HTE estimation.

Black-box heterogeneous treatment effect (HTE) models are increasingly being used to create personalized policies that assign individuals to their optimal treatments. However, they are difficult to understand, and can be burdensome to maintain in a production environment. In this paper, we present a scalable, interpretable personalized experimentation system, implemented and deployed in production at Meta. The system works in a multiple treatment, multiple outcome setting typical at Meta to: (1) learn explanations for black-box HTE models; (2) generate interpretable personalized policies. We evaluate the methods used in the system on publicly available data and Meta use cases, and discuss lessons learnt during the development of the system., Comment: Camera-ready version for KDD 2022. Previously titled "Distilling Heterogeneity: From Explanations of Heterogeneous Treatment Effect Models to Interpretable Policies". A short version was presented at MIT CODE 2021

Generalized additive models (GAMs) have become a leading modelclass for interpretable machine learning. However, there are many algorithms for training GAMs, and these can learn different or even contradictory models, while being equally accurate. Which GAM should we trust? In this paper, we quantitatively and qualitatively investigate a variety of GAM algorithms on real and simulated datasets. We find that GAMs with high feature sparsity (only using afew variables to make predictions) can miss patterns in the data and be unfair to rare subpopulations. Our results suggest that inductive bias plays a crucial role in what interpretable models learn and that tree-based GAMs represent the best balance of sparsity, fidelity and accuracy and thus appear to be the most trustworthy GAM., Comment: Accepted in 2021 KDD

Understanding how racial information impacts human decision making in online systems is critical in today's world. Prior work revealed that race information of criminal defendants, when presented as a text field, had no significant impact on users' judgements of recidivism. We replicated and extended this work to explore how and when race information influences users' judgements, with respect to the saliency of presentation. Our results showed that adding photos to the race labels had a significant impact on recidivism predictions for users who identified as female, but not for those who identified as male. The race of the defendant also impacted these results, with black defendants being less likely to be predicted to recidivate compared to white defendants. These results have strong implications for how system-designers choose to display race information, and cautions researchers to be aware of gender and race effects when using Amazon Mechanical Turk workers., Comment: This paper has been accepted for publication at CHI2020

Models which estimate main effects of individual variables alongside interaction effects have an identifiability challenge: effects can be freely moved between main effects and interaction effects without changing the model prediction. This is a critical problem for interpretability because it permits "contradictory" models to represent the same function. To solve this problem, we propose pure interaction effects: variance in the outcome which cannot be represented by any smaller subset of features. This definition has an equivalence with the Functional ANOVA decomposition. To compute this decomposition, we present a fast, exact algorithm that transforms any piecewise-constant function (such as a tree-based model) into a purified, canonical representation. We apply this algorithm to Generalized Additive Models with interactions trained on several datasets and show large disparity, including contradictions, between the effects before and after purification. These results underscore the need to specify data distributions and ensure identifiability before interpreting model parameters., Comment: AISTATS 2020

Methods for interpreting machine learning black-box models increase the outcomes' transparency and in turn generates insight into the reliability and fairness of the algorithms. However, the interpretations themselves could contain significant uncertainty that undermines the trust in the outcomes and raises concern about the model's reliability. Focusing on the method "Local Interpretable Model-agnostic Explanations" (LIME), we demonstrate the presence of two sources of uncertainty, namely the randomness in its sampling procedure and the variation of interpretation quality across different input data points. Such uncertainty is present even in models with high training and test accuracy. We apply LIME to synthetic data and two public data sets, text classification in 20 Newsgroup and recidivism risk-scoring in COMPAS, to support our argument.

Generalized additive models (GAMs) are favored in many regression and binary classification problems because they are able to fit complex, nonlinear functions while still remaining interpretable. In the first part of this paper, we generalize a state-of-the-art GAM learning algorithm based on boosted trees to the multiclass setting, and show that this multiclass algorithm outperforms existing GAM learning algorithms and sometimes matches the performance of full complexity models such as gradient boosted trees. In the second part, we turn our attention to the interpretability of GAMs in the multiclass setting. Surprisingly, the natural interpretability of GAMs breaks down when there are more than two classes. Naive interpretation of multiclass GAMs can lead to false conclusions. Inspired by binary GAMs, we identify two axioms that any additive model must satisfy in order to not be visually misleading. We then develop a technique called Additive Post-Processing for Interpretability (API), that provably transforms a pre-trained additive model to satisfy the interpretability axioms without sacrificing accuracy. The technique works not just on models trained with our learning algorithm, but on any multiclass additive model, including multiclass linear and logistic regression. We demonstrate the effectiveness of API on a 12-class infant mortality dataset., Comment: KDD 2019

When might human input help (or not) when assessing risk in fairness domains? Dressel and Farid (2018) asked Mechanical Turk workers to evaluate a subset of defendants in the ProPublica COMPAS data for risk of recidivism, and concluded that COMPAS predictions were no more accurate or fair than predictions made by humans. We delve deeper into this claim to explore differences in human and algorithmic decision making. We construct a Human Risk Score based on the predictions made by multiple Turk workers, characterize the features that determine agreement and disagreement between COMPAS and Human Scores, and construct hybrid Human+Machine models to predict recidivism. Our key finding is that on this data set, Human and COMPAS decision making differed, but not in ways that could be leveraged to significantly improve ground-truth prediction. We present the results of our analyses and suggestions for data collection best practices to leverage complementary strengths of human and machines in the fairness domain.

Existing methods to estimate the prevalence of chronic hepatitis C (HCV) in New York City (NYC) are limited in scope and fail to assess hard-to-reach subpopulations with highest risk such as injecting drug users (IDUs). To address these limitations, we employ a Bayesian multi-parameter evidence synthesis model to systematically combine multiple sources of data, account for bias in certain data sources, and provide unbiased HCV prevalence estimates with associated uncertainty. Our approach improves on previous estimates by explicitly accounting for injecting drug use and including data from high-risk subpopulations such as the incarcerated, and is more inclusive, utilizing ten NYC data sources. In addition, we derive two new equations to allow age at first injecting drug use data for former and current IDUs to be incorporated into the Bayesian evidence synthesis, a first for this type of model. Our estimated overall HCV prevalence as of 2012 among NYC adults aged 20-59 years is 2.78% (95% CI 2.61-2.94%), which represents between 124,900 and 140,000 chronic HCV cases. These estimates suggest that HCV prevalence in NYC is higher than previously indicated from household surveys (2.2%) and the surveillance system (2.37%), and that HCV transmission is increasing among young injecting adults in NYC. An ancillary benefit from our results is an estimate of current IDUs aged 20-59 in NYC: 0.58% or 27,600 individuals.

Non-negative matrix factorization (NMF) is a technique for finding latent representations of data. The method has been applied to corpora to construct topic models. However, NMF has likelihood assumptions which are often violated by real document corpora. We present a double parametric bootstrap test for evaluating the fit of an NMF-based topic model based on the duality of the KL divergence and Poisson maximum likelihood estimation. The test correctly identifies whether a topic model based on an NMF approach yields reliable results in simulated and real data., Comment: Presented at NIPS 2017 Symposium on Interpretable Machine Learning

Black-box risk scoring models permeate our lives, yet are typically proprietary or opaque. We propose Distill-and-Compare, a model distillation and comparison approach to audit such models. To gain insight into black-box models, we treat them as teachers, training transparent student models to mimic the risk scores assigned by black-box models. We compare the student model trained with distillation to a second un-distilled transparent model trained on ground-truth outcomes, and use differences between the two models to gain insight into the black-box model. Our approach can be applied in a realistic setting, without probing the black-box model API. We demonstrate the approach on four public data sets: COMPAS, Stop-and-Frisk, Chicago Police, and Lending Club. We also propose a statistical test to determine if a data set is missing key features used to train the black-box model. Our test finds that the ProPublica data is likely missing key feature(s) used in COMPAS., Comment: Camera-ready version for AAAI/ACM AIES 2018. Data and pseudocode at https://github.com/shftan/auditblackbox. Previously titled "Detecting Bias in Black-Box Models Using Transparent Model Distillation". A short version was presented at NIPS 2017 Symposium on Interpretable Machine Learning

Ensembles of decision trees perform well on many problems, but are not interpretable. In contrast to existing approaches in interpretability that focus on explaining relationships between features and predictions, we propose an alternative approach to interpret tree ensemble classifiers by surfacing representative points for each class -- prototypes. We introduce a new distance for Gradient Boosted Tree models, and propose new, adaptive prototype selection methods with theoretical guarantees, with the flexibility to choose a different number of prototypes in each class. We demonstrate our methods on random forests and gradient boosted trees, showing that the prototypes can perform as well as or even better than the original tree ensemble when used as a nearest-prototype classifier. In a user study, humans were better at predicting the output of a tree ensemble classifier when using prototypes than when using Shapley values, a popular feature attribution method. Hence, prototypes present a viable alternative to feature-based explanations for tree ensembles., Comment: Camera-ready version for ACM-IMS FODS 2020. A short version was presented at NIPS 2016 Workshop on Interpretable Machine Learning for Complex Systems