Scaling laws in language modeling traditionally quantify training loss as a function of dataset size and model parameters, providing compute-optimal estimates but often neglecting the impact of data quality on model generalization. In this paper, we extend the conventional understanding of scaling law by offering a microscopic view of data quality within the original formulation -- effective training tokens -- which we posit to be a critical determinant of performance for parameter-constrained language models. Specifically, we formulate the proposed term of effective training tokens to be a combination of two readily-computed indicators of text: (i) text diversity and (ii) syntheticity as measured by a teacher model. We pretrained over $200$ models of 25M to 1.5B parameters on a diverse set of sampled, synthetic data, and estimated the constants that relate text quality, model size, training tokens, and eight reasoning task accuracy scores. We demonstrated the estimated constants yield +0.83 Pearson correlation with true accuracies, and analyzed it in scenarios involving widely-used data techniques such as data sampling and synthesis which aim to improve data quality.

In the dynamic realm of social media, diverse topics are discussed daily, transcending linguistic boundaries. However, the complexities of understanding and categorising this content across various languages remain an important challenge with traditional techniques like topic modelling often struggling to accommodate this multilingual diversity. In this paper, we introduce X-Topic, a multilingual dataset featuring content in four distinct languages (English, Spanish, Japanese, and Greek), crafted for the purpose of tweet topic classification. Our dataset includes a wide range of topics, tailored for social media content, making it a valuable resource for scientists and professionals working on cross-linguistic analysis, the development of robust multilingual models, and computational scientists studying online dialogue. Finally, we leverage X-Topic to perform a comprehensive cross-linguistic and multilingual analysis, and compare the capabilities of current general- and domain-specific language models.

In order to train agents that can quickly adapt to new objectives or reward functions, efficient unsupervised representation learning in sequential decision-making environments can be important. Frameworks such as the Exogenous Block Markov Decision Process (Ex-BMDP) have been proposed to formalize this representation-learning problem (Efroni et al., 2022b). In the Ex-BMDP framework, the agent's high-dimensional observations of the environment have two latent factors: a controllable factor, which evolves deterministically within a small state space according to the agent's actions, and an exogenous factor, which represents time-correlated noise, and can be highly complex. The goal of the representation learning problem is to learn an encoder that maps from observations into the controllable latent space, as well as the dynamics of this space. Efroni et al. (2022b) has shown that this is possible with a sample complexity that depends only on the size of the controllable latent space, and not on the size of the noise factor. However, this prior work has focused on the episodic setting, where the controllable latent state resets to a specific start state after a finite horizon. By contrast, if the agent can only interact with the environment in a single continuous trajectory, prior works have not established sample-complexity bounds. We propose STEEL, the first provably sample-efficient algorithm for learning the controllable dynamics of an Ex-BMDP from a single trajectory, in the function approximation setting. STEEL has a sample complexity that depends only on the sizes of the controllable latent space and the encoder function class, and (at worst linearly) on the mixing time of the exogenous noise factor. We prove that STEEL is correct and sample-efficient, and demonstrate STEEL on two toy problems. Code is available at: https://github.com/midi-lab/steel.

While generative AI (GenAI) offers countless possibilities for creative and productive tasks, artificially generated media can be misused for fraud, manipulation, scams, misinformation campaigns, and more. To mitigate the risks associated with maliciously generated media, forensic classifiers are employed to identify AI-generated content. However, current forensic classifiers are often not evaluated in practically relevant scenarios, such as the presence of an attacker or when real-world artifacts like social media degradations affect images. In this paper, we evaluate state-of-the-art AI-generated image (AIGI) detectors under different attack scenarios. We demonstrate that forensic classifiers can be effectively attacked in realistic settings, even when the attacker does not have access to the target model and post-processing occurs after the adversarial examples are created, which is standard on social media platforms. These attacks can significantly reduce detection accuracy to the extent that the risks of relying on detectors outweigh their benefits. Finally, we propose a simple defense mechanism to make CLIP-based detectors, which are currently the best-performing detectors, robust against these attacks.

We investigate the internal representations of vision-language models (VLMs) to address hallucinations, a persistent challenge despite advances in model size and training. We project VLMs' internal image representations to their language vocabulary and observe more confident output probabilities on real objects than hallucinated objects. We additionally use these output probabilities to spatially localize real objects. Building on this approach, we introduce a knowledge erasure algorithm that removes hallucinations by linearly orthogonalizing image features with respect to hallucinated object features. We show that targeted edits to a model's latent representations can reduce hallucinations by up to 25.7% on the COCO2014 dataset while preserving performance. Our findings demonstrate how a deeper understanding of VLMs' latent representations can enhance reliability and enable novel capabilities, such as zero-shot segmentation.

LLM watermarks stand out as a promising way to attribute ownership of LLMgenerated text. One threat to watermark credibility comes from spoofing attacks, where an unauthorized third party forges the watermark, enabling it to falsely attribute arbitrary texts to a particular LLM. While recent works have demonstrated that state-of-the-art schemes are in fact vulnerable to spoofing, they lack deeper qualitative analysis of the texts produced by spoofing methods. In this work, we for the first time reveal that there are observable differences between genuine and spoofed watermark texts. Namely, we show that regardless of their underlying approach, all current spoofing methods consistently leave observable artifacts in spoofed texts, indicative of watermark forgery. We build upon these findings to propose rigorous statistical tests that reliably reveal the presence of such artifacts, effectively discovering that a watermark was spoofed. Our experimental evaluation shows high test power across all current spoofing methods, providing insights into their fundamental limitations, and suggesting a way to mitigate this threat. The improving abilities of large language models (LLMs) to generate human-like text at scale (Bubeck et al., 2023; Dubey et al., 2024) come with a growing risk of potential misuse. Hence, reliable detection of machine-generated text becomes increasingly important. Researchers have proposed the concept of watermarking: augmenting generated text with an imperceptible signal that can later be detected to attribute ownership of a text to a specific LLM (Kirchenbauer et al., 2023; Kuditipudi et al., 2023; Christ et al., 2024). Major LLM companies have pledged to watermark their models (Bartz & Hu, 2023), and regulators actively advocate for their use (Biden, 2023; CEU, 2024).

In strategic classification, agents manipulate their features, at a cost, to receive a positive classification outcome from the learner's classifier. The goal of the learner in such settings is to learn a classifier that is robust to strategic manipulations. While the majority of works in this domain consider accuracy as the primary objective of the learner, in this work, we consider learning objectives that have group fairness guarantees in addition to accuracy guarantees. We work with the minimax group fairness notion that asks for minimizing the maximal group error rate across population groups. We formalize a fairness-aware Stackelberg game between a population of agents consisting of several groups, with each group having its own cost function, and a learner in the agnostic PAC setting in which the learner is working with a hypothesis class H. When the cost functions of the agents are separable, we show the existence of an efficient algorithm that finds an approximately optimal deterministic classifier for the learner when the number of groups is small. This algorithm remains efficient, both statistically and computationally, even when H is the set of all classifiers. We then consider cost functions that are not necessarily separable and show the existence of oracle-efficient algorithms that find approximately optimal randomized classifiers for the learner when H has finite strategic VC dimension. These algorithms work under the assumption that the learner is fully transparent: the learner draws a classifier from its distribution (randomized classifier) before the agents respond by manipulating their feature vectors. We highlight the effectiveness of such transparency in developing oracle-efficient algorithms. We conclude with verifying the efficacy of our algorithms on real data by conducting an experimental analysis.

Capacitive touch sensors capture the two-dimensional spatial profile (referred to as a touch heatmap) of a finger's contact with a mobile touchscreen. However, the research and design of touchscreen mobile keyboards -- one of the most speed and accuracy demanding touch interfaces -- has focused on the location of the touch centroid derived from the touch image heatmap as the input, discarding the rest of the raw spatial signals. In this paper, we investigate whether touch heatmaps can be leveraged to further improve the tap decoding accuracy for mobile touchscreen keyboards. Specifically, we developed and evaluated machine-learning models that interpret user taps by using the centroids and/or the heatmaps as their input and studied the contribution of the heatmaps to model performance. The results show that adding the heatmap into the input feature set led to 21.4% relative reduction of character error rates on average, compared to using the centroid alone. Furthermore, we conducted a live user study with the centroid-based and heatmap-based decoders built into Pixel 6 Pro devices and observed lower error rate, faster typing speed, and higher self-reported satisfaction score based on the heatmap-based decoder than the centroid-based decoder. These findings underline the promise of utilizing touch heatmaps for improving typing experience in mobile keyboards.

Tree-based methods are powerful nonparametric techniques in statistics and machine learning. However, their effectiveness, particularly in finite-sample settings, is not fully understood. Recent applications have revealed their surprising ability to distinguish transformations (which we call symbolic feature selection) that remain obscure under current theoretical understanding. This work provides a finite-sample analysis of tree-based methods from a ranking perspective. We link oracle partitions in tree methods to response rankings at local splits, offering new insights into their finite-sample behavior in regression and feature selection tasks. Building on this local ranking perspective, we extend our analysis in two ways: (i) We examine the global ranking performance of individual trees and ensembles, including Classification and Regression Trees (CART) and Bayesian Additive Regression Trees (BART), providing finite-sample oracle bounds, ranking consistency, and posterior contraction results. (ii) Inspired by the ranking perspective, we propose concordant divergence statistics $\mathcal{T}_0$ to evaluate symbolic feature mappings and establish their properties. Numerical experiments demonstrate the competitive performance of these statistics in symbolic feature selection tasks compared to existing methods.

Feature selection can greatly improve performance and interpretability in machine learning problems. However, existing nonparametric feature selection methods either lack theoretical error control or fail to accurately control errors in practice. Many methods are also slow, especially in high dimensions. In this paper, we introduce a general feature selection method that applies integrated path stability selection to thresholding to control false positives and the false discovery rate. The method also estimates q-values, which are better suited to high-dimensional data than p-values. We focus on two special cases of the general method based on gradient boosting (IPSSGB) and random forests (IPSSRF). Extensive simulations with RNA sequencing data show that IPSSGB and IPSSRF have better error control, detect more true positives, and are faster than existing methods. We also use both methods to detect microRNAs and genes related to ovarian cancer, finding that they make better predictions with fewer features than other methods.