Beyond Surface Structure: A Causal Assessment of LLMs' Comprehension Ability

Large language models (LLMs) have shown remarkable capability in natural language tasks, yet debate persists on whether they truly comprehend deep structure (i.e., core semantics) or merely rely on surface structure (e.g., presentation format). Prior studies observe that LLMs' performance declines when intervening on surface structure, arguing their success relies on surface structure recognition. However, surface structure sensitivity does not prevent deep structure comprehension. Rigorously evaluating LLMs' capability requires analyzing both, yet deep structure is often overlooked. To this end, we assess LLMs' comprehension ability using causal mediation analysis, aiming to fully discover the capability of using both deep and surface structures. Specifically, we formulate the comprehension of deep structure as direct causal effect (DCE) and that of surface structure as indirect causal effect (ICE), respectively. To address the non-estimability of original DCE and ICE -- stemming from the infeasibility of isolating mutual influences of deep and surface structures, we develop the corresponding quantifiable surrogates, including approximated DCE (ADCE) and approximated ICE (AICE). We further apply the ADCE to evaluate a series of mainstream LLMs, showing that most of them exhibit deep structure comprehension ability, which grows along with the prediction accuracy. Comparing ADCE and AICE demonstrates closed-source LLMs rely more on deep structure, while open-source LLMs are more surface-sensitive, which decreases with model scale. Theoretically, ADCE is a bidirectional evaluation, which measures both the sufficiency and necessity of deep structure changes in causing output variations, thus offering a more comprehensive assessment than accuracy, a common evaluation in LLMs. Our work provides new insights into LLMs' deep structure comprehension and offers novel methods for LLMs evaluation.

Slight Corruption in Pre-training Data Makes Better Diffusion Models

Diffusion models (DMs) have shown remarkable capabilities in generating realistic high-quality images, audios, and videos. They benefit significantly from extensive pre-training on large-scale datasets, including web-crawled data with paired data and conditions, such as image-text and image-class pairs. Despite rigorous filtering, these pre-training datasets often inevitably contain corrupted pairs where conditions do not accurately describe the data. This paper presents the first comprehensive study on the impact of such corruption in pre-training data of DMs. We synthetically corrupt ImageNet-1K and CC3M to pre-train and evaluate over 50 conditional DMs. Our empirical findings reveal that various types of slight corruption in pre-training can significantly enhance the quality, diversity, and fidelity of the generated images across different DMs, both during pre-training and downstream adaptation stages. Theoretically, we consider a Gaussian mixture model and prove that slight corruption in the condition leads to higher entropy and a reduced 2-Wasserstein distance to the ground truth of the data distribution generated by the corruptly trained DMs. Inspired by our analysis, we propose a simple method to improve the training of DMs on practical datasets by adding condition embedding perturbations (CEP). CEP significantly improves the performance of various DMs in both pre-training and downstream tasks. We hope that our study provides new insights into understanding the data and pre-training processes of DMs.

Improving Group Robustness on Spurious Correlation Requires Preciser Group Inference

Standard empirical risk minimization (ERM) models may prioritize learning spurious correlations between spurious features and true labels, leading to poor accuracy on groups where these correlations do not hold. Mitigating this issue often requires expensive spurious attribute (group) labels or relies on trained ERM models to infer group labels when group information is unavailable. However, the significant performance gap in worst-group accuracy between using pseudo group labels and using oracle group labels inspires us to consider further improving group robustness through preciser group inference. Therefore, we propose GIC, a novel method that accurately infers group labels, resulting in improved worst-group performance. GIC trains a spurious attribute classifier based on two key properties of spurious correlations: (1) high correlation between spurious attributes and true labels, and (2) variability in this correlation between datasets with different group distributions. Empirical studies on multiple datasets demonstrate the effectiveness of GIC in inferring group labels, and combining GIC with various downstream invariant learning methods improves worst-group accuracy, showcasing its powerful flexibility. Additionally, through analyzing the misclassifications in GIC, we identify an interesting phenomenon called semantic consistency, which may contribute to better decoupling the association between spurious attributes and labels, thereby mitigating spurious correlation.

Conformalized semi-supervised random forest for classification and abnormality detection

Traditional classifiers infer labels under the premise that the training and test samples are generated from the same distribution. This assumption can be problematic for safety-critical applications such as medical diagnosis and network attack detection. In this paper, we consider the multi-class classification problem when the training data and the test data may have different distributions. We propose conformalized semi-supervised random forest (CSForest), which constructs set-valued predictions $C(x)$ to include the correct class label with desired probability while detecting outliers efficiently. We compare the proposed method to other state-of-art methods in both a synthetic example and a real data application to demonstrate the strength of our proposal.

GAN-based federated learning for label protection in binary classification

As an emerging technique, vertical federated learning collaborates with different data sources to jointly train a machine learning model without data exchange. However, federated learning is computationally expensive and inefficient in modeling due to complex encryption algorithms and secure computation protocols. Split learning offers an alternative solution to circumvent these challenges. Despite this, vanilla split learning still suffers privacy leakage. Here, we propose the Generative Adversarial Federated Model (GAFM), which integrates the vanilla split learning framework with the Generative Adversarial Network (GAN) for protection against label leakage from gradients in binary classification tasks. We compare our proposal to existing models, including Marvell, Max Norm, and SplitNN, on three publicly available datasets, where GAFM shows significant improvement regarding the trade-off between classification accuracy and label privacy protection. We also provide heuristic justification for why GAFM can improve over baselines and demonstrate that GAFM offers label protection through gradient perturbation compared to SplitNN.

FedGBF: An efficient vertical federated learning framework via gradient boosting and bagging

Federated learning, conducive to solving data privacy and security problems, has attracted increasing attention recently. However, the existing federated boosting model sequentially builds a decision tree model with the weak base learner, resulting in redundant boosting steps and high interactive communication costs. In contrast, the federated bagging model saves time by building multi-decision trees in parallel, but it suffers from performance loss. With the aim of obtaining an outstanding performance with less time cost, we propose a novel model in a vertically federated setting termed as Federated Gradient Boosting Forest (FedGBF). FedGBF simultaneously integrates the boosting and bagging's preponderance by building the decision trees in parallel as a base learner for boosting. Subsequent to FedGBF, the problem of hyperparameters tuning is rising. Then we propose the Dynamic FedGBF, which dynamically changes each forest's parameters and thus reduces the complexity. Finally, the experiments based on the benchmark datasets demonstrate the superiority of our method.