Felipe Maia Polo

Hongyi Wang, Felipe Maia Polo, Yuekai Sun et al.

Training AI models that generalize across tasks and domains has long been among the open problems driving AI research. The emergence of Foundation Models made it easier to obtain expert models for a given task, but the heterogeneity of data that may be encountered at test time often means that any single expert is insufficient. We consider the Fusion of Experts (FoE) problem of fusing outputs of expert models with complementary knowledge of the data distribution and formulate it as an instance of supervised learning. Our method is applicable to both discriminative and generative tasks and leads to significant performance improvements in image and text classification, text summarization, multiple-choice QA, and automatic evaluation of generated text. We also extend our method to the "frugal" setting where it is desired to reduce the number of expert model evaluations at test time. Our implementation is publicly available at https://github.com/hwang595/FoE-ICLR2024.

Peiyao Cai, Chengyu Cui, Felipe Maia Polo et al.

We propose a statistical framework built on latent variable modeling for scaling laws of large language models (LLMs). Our work is motivated by the rapid emergence of numerous new LLM families with distinct architectures and training strategies, evaluated on an increasing number of benchmarks. This heterogeneity makes a single global scaling curve inadequate for capturing how performance varies across families and benchmarks. To address this, we propose a latent variable modeling framework in which each LLM family is associated with a latent variable that captures the common underlying features in that family. An LLM's performance on different benchmarks is then driven by its latent skills, which are jointly determined by the latent variable and the model's own observable features. We develop an estimation procedure for this latent variable model and establish its statistical properties. We also design efficient numerical algorithms that support estimation and various downstream tasks. Empirically, we evaluate the approach on 12 widely used benchmarks from the Open LLM Leaderboard (v1/v2).

Felipe Maia Polo, Rafael Izbicki, Evanildo Gomes Lacerda et al.

Supervised learning techniques typically assume training data originates from the target population. Yet, in reality, dataset shift frequently arises, which, if not adequately taken into account, may decrease the performance of their predictors. In this work, we propose a novel and flexible framework called DetectShift that quantifies and tests for multiple dataset shifts, encompassing shifts in the distributions of $(X, Y)$, $X$, $Y$, $X|Y$, and $Y|X$. DetectShift equips practitioners with insights into data shifts, facilitating the adaptation or retraining of predictors using both source and target data. This proves extremely valuable when labeled samples in the target domain are limited. The framework utilizes test statistics with the same nature to quantify the magnitude of the various shifts, making results more interpretable. It is versatile, suitable for regression and classification tasks, and accommodates diverse data forms - tabular, text, or image. Experimental results demonstrate the effectiveness of DetectShift in detecting dataset shifts even in higher dimensions.

Felipe Maia Polo, Yuekai Sun, Moulinath Banerjee

Conditional independence (CI) testing is a fundamental and challenging task in modern statistics and machine learning. Many modern methods for CI testing rely on powerful supervised learning methods to learn regression functions or Bayes predictors as an intermediate step; we refer to this class of tests as regression-based tests. Although these methods are guaranteed to control Type-I error when the supervised learning methods accurately estimate the regression functions or Bayes predictors of interest, their behavior is less understood when they fail due to misspecified inductive biases; in other words, when the employed models are not flexible enough or when the training algorithm does not induce the desired predictors. Then, we study the performance of regression-based CI tests under misspecified inductive biases. Namely, we propose new approximations or upper bounds for the testing errors of three regression-based tests that depend on misspecification errors. Moreover, we introduce the Rao-Blackwellized Predictor Test (RBPT), a regression-based CI test robust against misspecified inductive biases. Finally, we conduct experiments with artificial and real data, showcasing the usefulness of our theory and methods.

Nimrod Shabtay, Felipe Maia Polo, Sivan Doveh et al.

The large-scale training of multi-modal models on data scraped from the web has shown outstanding utility in infusing these models with the required world knowledge to perform effectively on multiple downstream tasks. However, one downside of scraping data from the web can be the potential sacrifice of the benchmarks on which the abilities of these models are often evaluated. To safeguard against test data contamination and to truly test the abilities of these foundation models we propose LiveXiv: A scalable evolving live benchmark based on scientific ArXiv papers. LiveXiv accesses domain-specific manuscripts at any given timestamp and proposes to automatically generate visual question-answer pairs (VQA). This is done without any human-in-the-loop, using the multi-modal content in the manuscripts, like graphs, charts, and tables. Moreover, we introduce an efficient evaluation approach that estimates the performance of all models on the evolving benchmark using evaluations of only a subset of models. This significantly reduces the overall evaluation cost. We benchmark multiple open and proprietary Large Multi-modal Models (LMMs) on the first version of our benchmark, showing its challenging nature and exposing the models true abilities, avoiding contamination. Lastly, in our commitment to high quality, we have collected and evaluated a manually verified subset. By comparing its overall results to our automatic annotations, we have found that the performance variance is indeed minimal (<2.5%). Our dataset is available online on HuggingFace, and our code will be available here.

Felipe Maia Polo, Aida Nematzadeh, Virginia Aglietti et al.

Moving beyond evaluations that collapse performance across heterogeneous prompts toward fine-grained evaluation at the prompt level, or within relatively homogeneous subsets, is necessary to diagnose generative models' strengths and weaknesses. Such fine-grained evaluations, however, suffer from a data bottleneck: human gold-standard labels are too costly at this scale, while automated ratings are often misaligned with human judgment. To resolve this challenge, we propose a novel statistical model based on tensor factorization that merges cheap autorater data with a limited set of human gold-standard labels. Specifically, our approach uses autorater scores to pretrain latent representations of prompts and generative models, and then aligns those pretrained representations to human preferences using a small calibration set. This sample-efficient methodology is robust to autorater quality, more accurately predicts human preferences on a per-prompt basis than standard baselines, and provides tight confidence intervals for key statistical parameters of interest. We also showcase the practical utility of our method by constructing granular leaderboards based on prompt qualities and by estimating model performance solely from autorater scores, eliminating the need for additional human annotations.

Felipe Maia Polo, Lucas Weber, Leshem Choshen et al.

The versatility of large language models (LLMs) led to the creation of diverse benchmarks that thoroughly test a variety of language models' abilities. These benchmarks consist of tens of thousands of examples making evaluation of LLMs very expensive. In this paper, we investigate strategies to reduce the number of evaluations needed to assess the performance of an LLM on several key benchmarks. For example, we show that to accurately estimate the performance of an LLM on MMLU, a popular multiple-choice QA benchmark consisting of 14K examples, it is sufficient to evaluate this LLM on 100 curated examples. We release evaluation tools and tiny versions of popular benchmarks: Open LLM Leaderboard, MMLU, HELM, and AlpacaEval 2.0. Our empirical analysis demonstrates that these tools and tiny benchmarks are sufficient to reliably and efficiently reproduce the original evaluation results.

Felipe Maia Polo, Seamus Somerstep, Leshem Choshen et al.

Scaling laws for large language models (LLMs) predict model performance based on parameters like size and training data. However, differences in training configurations and data processing across model families lead to significant variations in benchmark performance, making it difficult for a single scaling law to generalize across all LLMs. On the other hand, training family-specific scaling laws requires training models of varying sizes for every family. In this work, we propose Skills Scaling Laws (SSLaws, pronounced as Sloth), a novel scaling law that leverages publicly available benchmark data and assumes LLM performance is driven by low-dimensional latent skills, such as reasoning and instruction following. These latent skills are influenced by computational resources like model size and training tokens but with varying efficiencies across model families. Sloth exploits correlations across benchmarks to provide more accurate and interpretable predictions while alleviating the need to train multiple LLMs per family. We present both theoretical results on parameter identification and empirical evaluations on 12 prominent benchmarks, from Open LLM Leaderboard v1/v2, demonstrating that Sloth predicts LLM performance efficiently and offers insights into scaling behaviors for complex downstream tasks and increased test-time compute.

Seamus Somerstep, Felipe Maia Polo, Allysson Flavio Melo de Oliveira et al.

With the rapid growth in the number of Large Language Models (LLMs), there has been a recent interest in LLM routing, or directing queries to the cheapest LLM that can deliver a suitable response. We conduct a minimax analysis of the routing problem, providing a lower bound and finding that a simple router that predicts both cost and accuracy for each question can be minimax optimal. Inspired by this, we introduce CARROT, a Cost AwaRe Rate Optimal rouTer that selects a model based on estimates of the models' cost and performance. Alongside CARROT, we also introduce the Smart Price-aware ROUTing (SPROUT) dataset to facilitate routing on a wide spectrum of queries with the latest state-of-the-art LLMs. Using SPROUT and prior benchmarks such as Routerbench and open-LLM-leaderboard-v2 we empirically validate CARROT's performance against several alternative routers.

Gabriel Patron, Zhiwei Xu, Ishan Kapnadak et al.

Personalization is central to human-AI interaction, yet current diffusion-based image generation systems remain largely insensitive to user diversity. Existing attempts to address this often rely on costly paired preference data or introduce latency through Large Language Models. In this work, we introduce REBECA (REcommendations BEyond CAtalogs), a lightweight and scalable framework for personalized image generation that learns directly from implicit feedback signals such as likes, ratings, and clicks. Instead of fine-tuning the underlying diffusion model, REBECA employs a two-stage process: training a conditional diffusion model to sample user- and rating-specific image embeddings, which are subsequently decoded into images using a pretrained diffusion backbone. This approach enables efficient, fine-tuning-free personalization across large user bases. We rigorously evaluate REBECA on real-world datasets, proposing a novel statistical personalization verifier and a permutation-based hypothesis test to assess preference alignment. Our results demonstrate that REBECA consistently produces high-fidelity images tailored to individual tastes, outperforming baselines while maintaining computational efficiency.

Daniele Bracale, Subha Maity, Felipe Maia Polo et al.

Often in prediction tasks, the predictive model itself can influence the distribution of the target variable, a phenomenon termed performative prediction. Generally, this influence stems from strategic actions taken by stakeholders with a vested interest in predictive models. A key challenge that hinders the widespread adaptation of performative prediction in machine learning is that practitioners are generally unaware of the social impacts of their predictions. To address this gap, we propose a methodology for learning the distribution map that encapsulates the long-term impacts of predictive models on the population. Specifically, we model agents' responses as a cost-adjusted utility maximization problem and propose estimates for said cost. Our approach leverages optimal transport to align pre-model exposure (ex ante) and post-model exposure (ex post) distributions. We provide a rate of convergence for this proposed estimate and assess its quality through empirical demonstrations on a credit-scoring dataset.

Maike Züfle, Vilém Zouhar, Tu Anh Dinh et al.

Automated metrics for machine translation attempt to replicate human judgment. Unlike humans, who often assess a translation in the context of multiple alternatives, these metrics typically consider only the source sentence and a single translation. This discrepancy in the evaluation setup may negatively impact the performance of automated metrics. We propose two automated metrics that incorporate additional information beyond the single translation. COMET-polycand uses alternative translations of the same source sentence to compare and contrast with the translation at hand, thereby providing a more informed assessment of its quality. COMET-polyic, inspired by retrieval-based in-context learning, takes in translations of similar source texts along with their human-labeled quality scores to guide the evaluation. We find that including a single additional translation in COMET-polycand improves the segment-level metric performance (0.079 to 0.118 Kendall's tau-b correlation), with further gains when more translations are added. Incorporating retrieved examples in COMET-polyic yields similar improvements (0.079 to 0.116 Kendall's tau-b correlation). We release our models publicly.

Felipe Maia Polo, Xinhe Wang, Mikhail Yurochkin et al.

Large language models are increasingly used as judges (LLM-as-a-judge) to evaluate model outputs at scale, but their assessments often diverge systematically from human judgments. We present Bridge, a unified statistical framework that explicitly bridges human and LLM evaluations under both absolute scoring and pairwise comparison paradigms. Bridge posits a latent human preference score for each prompt-response pair and models LLM deviations as linear transformations of covariates that capture sources of discrepancies. This offers a simple and principled framework for refining LLM ratings and characterizing systematic discrepancies between humans and LLMs. We provide an efficient fitting algorithm with asymptotic guarantees for statistical inference. Using six LLM judges and two benchmarks (BigGen Bench and Chatbot Arena), Bridge achieves higher agreement with human ratings (accuracy, calibration, and KL divergence) and exposes systematic human-LLM gaps.

Felipe Maia Polo, Subha Maity, Mikhail Yurochkin et al.

Programmatic Weak Supervision (PWS) enables supervised model training without direct access to ground truth labels, utilizing weak labels from heuristics, crowdsourcing, or pre-trained models. However, the absence of ground truth complicates model evaluation, as traditional metrics such as accuracy, precision, and recall cannot be directly calculated. In this work, we present a novel method to address this challenge by framing model evaluation as a partial identification problem and estimating performance bounds using Fréchet bounds. Our approach derives reliable bounds on key metrics without requiring labeled data, overcoming core limitations in current weak supervision evaluation techniques. Through scalable convex optimization, we obtain accurate and computationally efficient bounds for metrics including accuracy, precision, recall, and F1-score, even in high-dimensional settings. This framework offers a robust approach to assessing model quality without ground truth labels, enhancing the practicality of weakly supervised learning for real-world applications.

Felipe Maia Polo, Gabriel Caiaffa Floriano Mendonça, Kauê Capellato J. Parreira et al.

We present and make available pre-trained language models (Phraser, Word2Vec, Doc2Vec, FastText, and BERT) for the Brazilian legal language, a Python package with functions to facilitate their use, and a set of demonstrations/tutorials containing some applications involving them. Given that our material is built upon legal texts coming from several Brazilian courts, this initiative is extremely helpful for the Brazilian legal field, which lacks other open and specific tools and language models. Our main objective is to catalyze the use of natural language processing tools for legal texts analysis by the Brazilian industry, government, and academia, providing the necessary tools and accessible material.

Carmen Melo Toledo, Guilherme Mendes Bassedon, Jonathan Batista Ferreira et al.

Student's grade retention is a key issue faced by many education systems, especially those in developing countries. In this paper, we seek to gauge the relevance of students' personality traits in predicting grade retention in Brazil. For that, we used data collected in 2012 and 2017, in the city of Sertaozinho, countryside of the state of Sao Paulo, Brazil. The surveys taken in Sertaozinho included several socioeconomic questions, standardized tests, and a personality test. Moreover, students were in grades 4, 5, and 6 in 2012. Our approach was based on training machine learning models on the surveys' data to predict grade retention between 2012 and 2017 using information from 2012 or before, and then using some strategies to quantify personality traits' predictive power. We concluded that, besides proving to be fairly better than a random classifier when isolated, personality traits contribute to prediction even when using socioeconomic variables and standardized tests results.

Felipe Maia Polo, Renato Vicente

In supervised learning, training and test datasets are often sampled from distinct distributions. Domain adaptation techniques are thus required. Covariate shift adaptation yields good generalization performance when domains differ only by the marginal distribution of features. Covariate shift adaptation is usually implemented using importance weighting, which may fail, according to common wisdom, due to small effective sample sizes (ESS). Previous research argues this scenario is more common in high-dimensional settings. However, how effective sample size, dimensionality, and model performance/generalization are formally related in supervised learning, considering the context of covariate shift adaptation, is still somewhat obscure in the literature. Thus, a main challenge is presenting a unified theory connecting those points. Hence, in this paper, we focus on building a unified view connecting the ESS, data dimensionality, and generalization in the context of covariate shift adaptation. Moreover, we also demonstrate how dimensionality reduction or feature selection can increase the ESS, and argue that our results support dimensionality reduction before covariate shift adaptation as a good practice.

Felipe Maia Polo, Itamar Ciochetti, Emerson Bertolo

The objective of this paper is to develop predictive models to classify Brazilian legal proceedings in three possible classes of status: (i) archived proceedings, (ii) active proceedings, and (iii) suspended proceedings. This problem's resolution is intended to assist public and private institutions in managing large portfolios of legal proceedings, providing gains in scale and efficiency. In this paper, legal proceedings are made up of sequences of short texts called "motions." We combined several natural language processing (NLP) and machine learning techniques to solve the problem. Although working with Portuguese NLP, which can be challenging due to lack of resources, our approaches performed remarkably well in the classification task, achieving maximum accuracy of .93 and top average F1 Scores of .89 (macro) and .93 (weighted). Furthermore, we could extract and interpret the patterns learned by one of our models besides quantifying how those patterns relate to the classification task. The interpretability step is important among machine learning legal applications and gives us an exciting insight into how black-box models make decisions.