Both long-tailed and noisily labeled data frequently appear in real-world applications and impose significant challenges for learning. Most prior works treat either problem in an isolated way and do not explicitly consider the coupling effects of the two. Our empirical observation reveals that such solutions fail to consistently improve the learning when the dataset is long-tailed with label noise. Moreover, with the presence of label noise, existing methods do not observe universal improvements across different sub-populations; in other words, some sub-populations enjoyed the benefits of improved accuracy at the cost of hurting others. Based on these observations, we introduce the Fairness Regularizer (FR), inspired by regularizing the performance gap between any two sub-populations. We show that the introduced fairness regularizer improves the performances of sub-populations on the tail and the overall learning performance. Extensive experiments demonstrate the effectiveness of the proposed solution when complemented with certain existing popular robust or class-balanced methods.

Multimodal Large Language Models (MLLM) have enabled a wide range of advanced vision-language applications, including fine-grained object recognition and contextual understanding. When querying specific regions or objects in an image, human users naturally use "Visual Prompts" (VP) like bounding boxes to provide reference. However, no existing benchmark systematically evaluates the ability of MLLMs to interpret such VPs. This gap raises uncertainty about whether current MLLMs can effectively recognize VPs, an intuitive prompting method for humans, and utilize them to solve problems. To address this limitation, we introduce VP-Bench, aiming to assess MLLMs’ capability in VP perception and utilization. VP-Bench employs a two-stage evaluation framework: Stage 1 examines models’ ability to perceive VPs in natural scenes, utilizing 100K visualized prompts spanning 8 shapes and 355 attribute combinations. Stage 2 investigates the impact of VPs on downstream tasks, measuring their effectiveness in real-world problem-solving scenarios. Using VP-Bench, we evaluate 21 MLLMs, including proprietary systems (e.g., GPT-4o) and open-source models (e.g., InternVL-2.5 and Qwen2.5-VL). In addition, we conduct a comprehensive analysis of the factors influencing VP understanding, such as attribute variations and model scale. VP-Bench establishes a new reference framework for studying MLLMs’ ability to comprehend and resolve grounded referring questions.

Artificial intelligence (AI) models for computer vision trained with supervised machine learning are assumed to solve classification tasks by imitating human behavior learned from training labels. Most efforts in recent vision research focus on measuring the model task performance using standardized benchmarks such as accuracy. However limited work has sought to understand the perceptual difference between humans and machines. To fill this gap, this study first analyzes the statistical distributions of mistakes from the two sources, and then explores how task difficulty level affects these distributions. We find that even when AI learns an excellent model from the training data, one that outperforms humans in overall accuracy, these AI models have significant and consistent differences from human perception. We demonstrate the importance of studying these differences with a simple human-AI teaming algorithm that outperforms humans alone, AI alone, or AI-AI teaming.

Instruction tuning is critical for adapting large language models (LLMs) to downstream tasks, and recent studies have demonstrated that small amounts of human-curated data can outperform larger datasets, challenging traditional data scaling laws. While LLM-based data quality rating systems offer a cost-effective alternative to human annotation, they often suffer from inaccuracies and biases, even in powerful models like GPT-4. In this work, we introduce $DS^2$, a **D**iversity-aware **S**core curation method for **D**ata **S**election. By systematically modeling error patterns through a score transition matrix, $DS^2$ corrects LLM-based scores and promotes diversity in the selected data samples. Our approach shows that a curated subset (just 3.3\% of the original dataset) outperforms full-scale datasets (300k samples) across various machine-alignment benchmarks, and matches or surpasses human-aligned datasets such as LIMA with the same sample size (1k samples). These findings challenge conventional data scaling assumptions, highlighting that redundant, low-quality samples can degrade performance and reaffirming that ``more can be less''.

Unlearning in Large Language Models (LLMs) is essential for ensuring ethical and responsible AI use, especially in addressing privacy leak, bias, safety, and evolving regulations. Existing approaches to LLM unlearning often rely on retain data or a reference LLM, yet they struggle to adequately balance unlearning performance with overall model utility. This challenge arises because leveraging explicit retain data or implicit knowledge of retain data from a reference LLM to fine-tune the model tends to blur the boundaries between the forgotten and retain data, as different queries often elicit similar responses. In this work, we propose eliminating the need to retain data or the reference LLM for response calibration in LLM unlearning. Recognizing that directly applying gradient ascent on the forget data often leads to optimization instability and poor performance, our method guides the LLM on what not to respond to, and importantly, how to respond, based on the forget data. Hence, we introduce Forget data only Loss AjustmenT (FLAT), a "flat" loss adjustment approach which addresses these issues by maximizing $f$-divergence between the available template answer and the forget answer only w.r.t. the forget data. The variational form of the defined $f$-divergence theoretically provides a way of loss adjustment by assigning different importance weights for the learning w.r.t. template responses and the forgetting of responses subject to unlearning. Empirical results demonstrate that our approach not only achieves superior unlearning performance compared to existing methods but also minimizes the impact on the model’s retained capabilities, ensuring high utility across diverse tasks, including copyrighted content unlearning on Harry Potter dataset and MUSE Benchmark, and entity unlearning on the TOFU dataset.

Recent studies show that in supervised fine-tuning (SFT) of large language models (LLMs), data quality matters more than quantity. While most data cleaning methods concentrate on filtering entire samples, the quality of individual tokens within a sample can vary significantly. After pre-training, even in high-quality samples, patterns or phrases that are not task-related can be redundant, uninformative, or even harmful. Continuing to fine-tune on these patterns may offer limited benefit and even degrade downstream task performance. In this paper, we investigate token quality from a noisy-label perspective and propose a generic token cleaning pipeline for SFT tasks. Our method filters out uninformative tokens while preserving those carrying key task-specific information. Specifically, we first evaluate token quality by examining the influence of model updates on each token, then apply a threshold-based separation. The token influence can be measured in a single pass with a fixed reference model or iteratively with self-evolving reference models. The benefits and limitations of both methods are analyzed theoretically by error upper bounds. Extensive experiments show that our framework consistently improves downstream performance. Code is available at https://github.com/UCSC-REAL/TokenCleaning.

The rising popularity of large language models (LLMs) has raised concerns about potential abuse and harmful content. As a result, developing a highly generalizable and adaptable machine-generated text (MGT) detection system has become an urgent priority. Given that LLMs are most commonly misused in academic writing, this work investigates the generalization and adaptation capabilities of MGT detectors in three key aspects specific to academic writing: First, we construct MGT-Academic, a large-scale dataset comprising over 336M tokens and 749K samples. MGT-Academic focuses on academic writing, featuring human-written texts (HWTs) and MGTs across STEM, Humanities, and Social Sciences, paired with an extensible code framework for efficient benchmarking. Second, we benchmark the performance of various detectors for binary classification and text attribution tasks in both in-domain and cross-domain settings. This benchmark reveals the often-overlooked challenges of text attribution tasks. Third, we introduce a novel text attribution task in which models must adapt to new classes over time, with little or no access to prior training data, spanning both few-shot and many-shot scenarios. We implement a range of adaptation techniques to enhance performance across these settings. Our findings provide new insights into the generalization ability of MGT detectors and lay the foundation for building robust, adaptive detection systems. The code framework is available at https://github.com/Y-L-LIU/MGTBench-2.0.

Previous methods for image geo-localization have typically treated the task as either classification or retrieval, often relying on black-box decisions that lack interpretability. The rise of large vision-language models (LVLMs) has enabled a rethinking of geo-localization as a reasoning-driven task grounded in visual cues. However, two major challenges persist. On the data side, existing reasoning-focused datasets are primarily based on street-view imagery, offering limited scene diversity and constrained viewpoints. On the modeling side, current approaches predominantly rely on supervised fine-tuning, which yields only marginal improvements in reasoning capabilities. To address these challenges, we propose a novel pipeline that constructs a reasoning-oriented geo-localization dataset, $\textit{MP16-Reason}$, using diverse social media images. We introduce $\textit{GLOBE}$, $\textbf{G}$roup-relative policy optimization for $\textbf{L}$ocalizability assessment and $\textbf{O}$ptimized visual-cue reasoning, yielding $\textbf{B}$i-objective geo-$\textbf{E}$nhancement for the VLM in recognition and reasoning. $\textit{GLOBE}$ incorporates task-specific rewards that jointly enhance localizability assessment, visual-cue reasoning, and geolocation accuracy. Both qualitative and quantitative results demonstrate that $\textit{GLOBE}$ outperforms state-of-the-art open-source LVLMs on geo-localization tasks, particularly in diverse visual scenes, while also generating more insightful and interpretable reasoning trajectories. The data and code are available at https://github.com/lingli1996/GLOBE.

The widespread deployment of Large Language Models (LLMs) trained on massive, uncurated corpora has raised growing concerns about the inclusion of sensitive, copyrighted, or illegal content. This has led to increasing interest in LLM unlearning: the task of selectively removing specific information from a model without retraining from scratch or degrading overall utility. However, existing methods often rely on large-scale forget and retain datasets, and suffer from unnatural responses, poor generalization, or catastrophic utility loss. In this work, we propose $\textbf{R}$einforcement $\textbf{U}$n$\textbf{LE}$arning ($\textbf{RULE}$), an efficient framework that formulates unlearning as a refusal boundary optimization problem. RULE is trained with a small portion of forget set and synthesized boundary queries, using a verifiable reward function that encourages safe refusal on forget-related queries while preserving helpful responses on permissible inputs. We provide both theoretical and empirical evidence demonstrating the effectiveness of RULE in achieving targeted unlearning without compromising model utility. Experimental results show that, with only 12\% forget set and 8\% synthesized boundary data, RULE outperforms existing baselines by up to $17.4\%$ forget quality and $16.3\%$ naturalness response while maintaining general utility, achieving $\textit{forget-retain Pareto Optimality}$. Remarkably, we further observe that RULE improves the $\textit{naturalness}$ of model outputs, enhances training $\textit{efficiency}$, and exhibits strong $\textit{generalization ability}$, generalizing refusal behavior to semantically related but unseen queries.

The utilization of machine learning algorithms in distributed web applications is experiencing significant growth. One notable approach is Federated Learning (FL) Recent research has brought attention to the vulnerability of FL to gradient inversion attacks, which seek to reconstruct the original training samples, posing a substantial threat to client privacy. Most existing gradient inversion attacks, however, require control over the central server and rely on substantial prior knowledge, including information about batch normalization and data distribution. In this study, we introduce Poisoning Gradient Leakage from Client (PGLC), a novel attack method that operates from the clients’ side. For the first time, we demonstrate the feasibility of a client-side adversary with limited knowledge successfully recovering training samples from the aggregated global model. Our approach enables the adversary to employ a malicious model that increases the loss of a specific targeted class of interest. When honest clients employ the poisoned global model, the gradients of samples become distinct in the aggregated update. This allows the adversary to effectively reconstruct private inputs from other clients using the aggregated update. Furthermore, our PGLC attack exhibits stealthiness against Byzantine-robust aggregation rules (AGRs). Through the optimization of malicious updates and the blending of benign updates with a malicious replacement vector, our method remains undetected by these defense mechanisms. We conducted experiments across various benchmark datasets, considering representative Byzantine-robust AGRs and exploring different FL settings with varying levels of adversary knowledge about the data. Our results consistently demonstrate the ability of PGLC to extract training data in all tested scenarios.

Federated learning provides a promising paradigm for collecting machine learning models from distributed data sources without compromising users’ data privacy. The success of a credible federated learning system builds on the assumption that the decentralized and self-interested users will be willing to participate to share their local models in a trustworthy way. However, without proper incentives, users might simply opt out of the contribution cycle or will be mis-incentivized to contribute spam/false information. This paper introduces solutions to audit the truthful reporting of a local, user-side machine learning model for federated learning. Our results build on the literature of information elicitation but focus on the questions of verifying machine learning models (rather than human predictions). We provide a scoring rule-based framework that verifies and audits reporting of local models such that truthfully reporting is a Bayesian Nash Equilibrium. We study the market implementation, accuracy as well as robustness properties of our proposed solution. We verify the effectiveness of our methods using MNIST and CIFAR-10 datasets. Particularly we show that by reporting low-quality hypotheses, users will receive decreasing scores (rewards, payment, or higher punishment).

The rawly collected training data often comes with separate noisy labels collected from multiple imperfect annotators (e.g., via crowdsourcing). A typical way of using these separate labels is to first aggregate them into one and apply standard training methods. The literature has also studied extensively on effective aggregation approaches. This paper revisits this choice and aims to provide an answer to the question of whether one should aggregate separate noisy labels into single ones or use them separately as given. We theoretically analyze the performance of both approaches under the empirical risk minimization framework for a number of popular loss functions, including the ones designed specifically for the problem of learning with noisy labels. Our theorems conclude that label separation is preferred over label aggregation when the noise rates are high, or the number of labelers/annotations is insufficient. Extensive empirical results validate our conclusions.

Label smoothing (LS) is an arising learning paradigm that uses the positively weighted average of both the hard training labels and uniformly distributed soft labels. It was shown that LS serves as a regularizer for training data with hard labels and therefore improves the generalization of the model. Later it was reported LS even helps with improving robustness when learning with noisy labels. However, we observed that the advantage of LS vanishes when we operate in a high label noise regime. Intuitively speaking, this is due to the increased entropy of P(noisy label|X) when the noise rate is high, in which case, further applying LS tends to “over-smooth” the estimated posterior. We proceeded to discover that several learning-with-noisy-labels solutions in the literature instead relate more closely to negative/not label smoothing (NLS), which acts counter to LS and defines as using a negative weight to combine the hard and soft labels! We provide understandings for the properties of LS and NLS when learning with noisy labels. Among other established properties, we theoretically show NLS is considered more beneficial when the label noise rates are high. We provide extensive experimental results on multiple benchmarks to support our findings too. Code is publicly available at https://github.com/UCSC-REAL/negative-label-smoothing.

It is important to collect credible training samples $(x,y)$ for building data-intensive learning systems (e.g., a deep learning system). Asking people to report complex distribution $p(x)$, though theoretically viable, is challenging in practice. This is primarily due to the cognitive loads required for human agents to form the report of this highly complicated information. While classical elicitation mechanisms apply to eliciting a complex and generative (and continuous) distribution $p(x)$, we are interested in eliciting samples $x_i \sim p(x)$ from agents directly. We coin the above problem sample elicitation. This paper introduces a deep learning aided method to incentivize credible sample contributions from self-interested and rational agents. We show that with an accurate estimation of a certain $f$-divergence function we can achieve approximate incentive compatibility in eliciting truthful samples. We then present an efficient estimator with theoretical guarantees via studying the variational forms of the $f$-divergence function. We also show a connection between this sample elicitation problem and $f$-GAN, and how this connection can help reconstruct an estimator of the distribution based on collected samples. Experiments on synthetic data, MNIST, and CIFAR-10 datasets demonstrate that our mechanism elicits truthful samples. Our implementation is available at https://github.com/weijiaheng/Credible-sample-elicitation.git.