André F. Cruz

1. Preprint 2026 ICLR 2026 Workshop

   A Description is Worth a Million Clicks

   André F Cruz, Andreas Bjerre-Nielsen, Jon Kleinberg, and Rediet Abebe PDF

   Details: A Description is Worth a Million Clicks

   Preprint, 2026

   Earlier version: Text as the Richest Preference Signal, ICLR 2026 Workshop on AI for Mechanism Design and Strategic Decision Making.

   Representations of human preferences guide search, content feeds, advertising, and large language model (LLM) alignment. For decades, these representations have been learned from fixed-form signals (ratings, likes, clicks), and mature, scalable algorithms such as collaborative filtering were built around them. Yet people express their preferences far more naturally in language. LLMs now make it possible to work with such text directly. On six large-scale datasets from Amazon and Goodreads, we compare collaborative filtering, trained and heavily tuned on abundant interaction data, against a zero-shot LLM prompted with item descriptions and users’ past reviews. This simple text baseline is competitive with collaborative filtering trained on around one million interactions. To probe why, we use LLMs to distill text into thousands of interpretable item characteristics and each user’s stated preferences over them. These features predict choices just as well as the text they come from. This is consistent with Lancaster’s view of goods as bundles of characteristics: a review speaks to characteristics shared by many items, whereas an interaction records a single choice. Fixed-form and free-form representations are also complementary, jointly outperforming any single model. Together, these findings invite revisiting how preferences are elicited and encoded.

   BibTeX

   @article{cruz2026description,
     title = {A Description is Worth a Million Clicks},
     author = {Cruz, Andr{\'e} F and Bjerre-Nielsen, Andreas and Kleinberg, Jon and Abebe, Rediet},
     journal = {Preprint},
     year = {2026},
   }

2. ICLR 2026

   ROC-n-reroll: How verifier imperfection affects test-time scaling

   Florian E Dorner, Yatong Chen*, André F Cruz*, and Fanny Yang arXivCode

   Details: ROC-n-reroll: How verifier imperfection affects test-time scaling

   In The Fourteenth International Conference on Learning Representations (ICLR), 2026

   Test-time scaling aims to improve language model performance by leveraging additional compute during inference. While many works have empirically studied techniques like Best-of-N (BoN) and rejection sampling that make use of a verifier to enable test-time scaling, there is little theoretical understanding of how verifier imperfection affects performance. In this work, we address this gap. Specifically, we prove how instance-level accuracy of these methods is precisely characterized by the geometry of the verifier’s ROC curve. Interestingly, while scaling is determined by the local geometry of the ROC curve for rejection sampling, it depends on global properties of the ROC curve for BoN. As a consequence when the ROC curve is unknown, it is impossible to extrapolate the performance of rejection sampling based on the low-compute regime. Furthermore, while rejection sampling outperforms BoN for fixed compute, in the infinite-compute limit both methods converge to the same level of accuracy, determined by the slope of the ROC curve near the origin. Our theoretical results are confirmed by experiments on GSM8K using different versions of Llama and Qwen to generate and verify solutions.

   OpenReview

   BibTeX

   @inproceedings{dorner2026rocnreroll,
     title = {{ROC}-n-reroll: How verifier imperfection affects test-time scaling},
     author = {Dorner, Florian E and Chen{*}, Yatong and Cruz{*}, Andr{\'e} F and Yang, Fanny},
     booktitle = {The Fourteenth International Conference on Learning Representations (ICLR)},
     year = {2026},
     url = {https://openreview.net/forum?id=3Gy5mmyuxn},
   }

3. NeurIPS 2024

   Evaluating language models as risk scores

   André F. Cruz, Moritz Hardt, and Celestine Mendler-Dünner arXivCode

   Details: Evaluating language models as risk scores

   In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track (NeurIPS), 2024

   Current question-answering benchmarks predominantly focus on accuracy in realizable prediction tasks. Conditioned on a question and answer-key, does the most likely token match the ground truth? Such benchmarks necessarily fail to evaluate language models’ ability to quantify outcome uncertainty. In this work, we focus on the use of language models as risk scores for unrealizable prediction tasks. We introduce folktexts, a software package to systematically generate risk scores using large language models, and evaluate them against benchmark prediction tasks. Specifically, the package derives natural language tasks from US Census data products, inspired by popular tabular data benchmarks. A flexible API allows for any task to be constructed out of 28 census features whose values are mapped to prompt-completion pairs. We demonstrate the utility of folktexts through a sweep of empirical insights on 16 recent large language models, inspecting risk scores, calibration curves, and diverse evaluation metrics. We find that zero-shot risk sores have high predictive signal while being widely miscalibrated: base models overestimate outcome uncertainty, while instruction-tuned models underestimate uncertainty and generate over-confident risk scores.

   OpenReview

   BibTeX

   @inproceedings{cruz2024evaluating,
     title = {Evaluating language models as risk scores},
     author = {Cruz, Andr\'{e} F. and Hardt, Moritz and Mendler-Dünner, Celestine},
     booktitle = {The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track (NeurIPS)},
     year = {2024},
     url = {https://openreview.net/forum?id=qrZxL3Bto9},
   }

4. ICLR 2024 Oral

   Unprocessing Seven Years of Algorithmic Fairness

   André F Cruz and Moritz Hardt arXivCode

   Details: Unprocessing Seven Years of Algorithmic Fairness

   In The Twelfth International Conference on Learning Representations (ICLR), 2024 · Oral presentation, top 1%

   Seven years ago, researchers proposed a postprocessing method to equalize the error rates of a model across different demographic groups. The work launched hundreds of papers purporting to improve over the postprocessing baseline. We empirically evaluate these claims through thousands of model evaluations on several tabular datasets. We find that the fairness-accuracy Pareto frontier achieved by postprocessing contains all other methods we were feasibly able to evaluate. In doing so, we address two common methodological errors that have confounded previous observations. One relates to the comparison of methods with different unconstrained base models. The other concerns methods achieving different levels of constraint relaxation. At the heart of our study is a simple idea we call unprocessing that roughly corresponds to the inverse of postprocessing. Unprocessing allows for a direct comparison of methods using different underlying models and levels of relaxation. Interpreting our findings, we recall a widely overlooked theoretical argument, present seven years ago, that accurately predicted what we observe.

   OpenReview

   BibTeX

   @inproceedings{cruz2024unprocessing,
     title = {Unprocessing Seven Years of Algorithmic Fairness},
     author = {Cruz, Andr\'{e} F and Hardt, Moritz},
     booktitle = {The Twelfth International Conference on Learning Representations (ICLR)},
     year = {2024},
     url = {https://openreview.net/forum?id=jr03SfWsBS},
   }

5. ICLR 2023

   FairGBM: Gradient Boosting with Fairness Constraints

   André F Cruz, Catarina Belém, Sérgio Jesus, +3 more arXivCode

   Details: FairGBM: Gradient Boosting with Fairness Constraints

   In The Eleventh International Conference on Learning Representations (ICLR), 2023

   Machine Learning (ML) algorithms based on gradient boosted decision trees (GBDT) are still favored on many tabular data tasks across various mission critical applications, from healthcare to finance. However, GBDT algorithms are not free of the risk of bias and discriminatory decision-making. Despite GBDT’s popularity and the rapid pace of research in fair ML, existing in-processing fair ML methods are either inapplicable to GBDT, incur in significant train time overhead, or are inadequate for problems with high class imbalance. We present FairGBM, a learning framework for training GBDT under fairness constraints with little to no impact on predictive performance when compared to unconstrained LightGBM. Since common fairness metrics are non-differentiable, we employ a "proxy-Lagrangian" formulation using smooth convex error rate proxies to enable gradient-based optimization. Additionally, our open-source implementation shows an order of magnitude speedup in training time when compared with related work, a pivotal aspect to foster the widespread adoption of FairGBM by real-world practitioners.

   OpenReview

   BibTeX

   @inproceedings{cruz2023fairgbm,
     title = {Fair{GBM}: Gradient Boosting with Fairness Constraints},
     author = {Cruz, Andr\'{e} F and Bel{\'{e}}m, Catarina and Jesus, S{\'{e}}rgio and Bravo, Jo{\~{a}}o and Saleiro, Pedro and Bizarro, Pedro},
     booktitle = {The Eleventh International Conference on Learning Representations (ICLR)},
     year = {2023},
     url = {https://openreview.net/forum?id=x-mXzBgCX3a},
   }

6. ICDM 2021

   Promoting Fairness through Hyperparameter Optimization

   André F Cruz, Pedro Saleiro, Catarina Belém, +2 more arXivCode

   Details: Promoting Fairness through Hyperparameter Optimization

   In 2021 IEEE International Conference on Data Mining (ICDM), 2021

   Considerable research effort has been guided towards algorithmic fairness but real-world adoption of bias reduction techniques is still scarce. Existing methods are either metric- or model-specific, require access to sensitive attributes at inference time, or carry high development or deployment costs. This work explores the unfairness that emerges when optimizing ML models solely for predictive performance, and how to mitigate it with a simple and easily deployed intervention: fairness-aware hyperparameter optimization (HO). We propose and evaluate fairness-aware variants of three popular HO algorithms: Fair Random Search, Fair TPE, and Fairband. We validate our approach on a real-world bank account opening fraud case-study, as well as on three datasets from the fairness literature. Results show that, without extra training cost, it is feasible to find models with 111% mean fairness increase and just 6% decrease in performance when compared with fairness-blind HO.

   DOI

   BibTeX

   @inproceedings{cruz2021promoting,
     author = {Cruz, Andr\'{e} F and Saleiro, Pedro and Bel\'{e}m, Catarina and Soares, Carlos and Bizarro, Pedro},
     booktitle = {2021 IEEE International Conference on Data Mining (ICDM)},
     title = {Promoting Fairness through Hyperparameter Optimization},
     year = {2021},
     url = {https://doi.org/10.1109/ICDM51629.2021.00119},
     volume = {},
     number = {},
     pages = {1036-1041},
     doi = {10.1109/ICDM51629.2021.00119},
   }

7. KDD 2021

   TimeSHAP: Explaining Recurrent Models through Sequence Perturbations

   João Bento, Pedro Saleiro, André F Cruz, +2 more arXivCode

   Details: TimeSHAP: Explaining Recurrent Models through Sequence Perturbations

   In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, 2021

   Although recurrent neural networks (RNNs) are state-of-the-art in numerous sequential decision-making tasks, there has been little research on explaining their predictions. In this work, we present TimeSHAP, a model-agnostic recurrent explainer that builds upon KernelSHAP and extends it to the sequential domain. TimeSHAP computes feature-, timestep-, and cell-level attributions. As sequences may be arbitrarily long, we further propose a pruning method that is shown to dramatically decrease both its computational cost and the variance of its attributions. We use TimeSHAP to explain the predictions of a real-world bank account takeover fraud detection RNN model, and draw key insights from its explanations: i) the model identifies important features and events aligned with what fraud analysts consider cues for account takeover; ii) positive predicted sequences can be pruned to only 10% of the original length, as older events have residual attribution values; iii) the most recent input event of positive predictions only contributes on average to 41% of the model’s score; iv) notably high attribution to client’s age, upheld on higher false positive rates for older clients.

   DOIWebsite

   BibTeX

   @inproceedings{bento2021timeshap,
     author = {Bento, Jo\~{a}o and Saleiro, Pedro and Cruz, Andr\'{e} F and Figueiredo, M\'{a}rio A.T. and Bizarro, Pedro},
     title = {TimeSHAP: Explaining Recurrent Models through Sequence Perturbations},
     year = {2021},
     isbn = {9781450383325},
     publisher = {Association for Computing Machinery},
     address = {New York, NY, USA},
     url = {https://doi.org/10.1145/3447548.3467166},
     doi = {10.1145/3447548.3467166},
     booktitle = {Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery \& Data Mining},
     pages = {2565–2573},
     numpages = {9},
     keywords = {Shapley values, TimeSHAP, RNN, XAI, SHAP, explainability},
     location = {Virtual Event, Singapore},
     series = {KDD '21},
   }

All publications→