Corrupted training data are ubiquitous. Corrective Machine Unlearning (CMU) seeks to remove the influence of such corruption post-training. Prior CMU typically assumes access to identified corrupted training samples (a "forget set"). However, in many real-world scenarios the training data are no longer accessible. We formalize source-free CMU, where the original training data are unavailable and, consequently, no forget set of identified corrupted training samples can be specified. Instead, we assume a small proxy (surrogate) set of corrupted samples that reflect the suspected corruption type without needing to be the original training samples. In this stricter setting, methods relying on forget set are ineffective or narrow in scope. We introduce Corrective Unlearning in Task Space (CUTS), a lightweight weight space correction method guided by the proxy set using task arithmetic principles. CUTS treats the clean and the corruption signal as distinct tasks. Specifically, we briefly fine-tune the corrupted model on the proxy to amplify the corruption mechanism in the weight space, compute the difference between the corrupted and fine-tuned weights as a proxy task vector, and subtract a calibrated multiple of this vector to cancel the corruption. Without access to clean data or a forget set, CUTS recovers a large fraction of the lost utility under label noise and, for backdoor triggers, nearly eliminates the attack with minimal damage to utility, outperforming state-of-the-art specialized CMU methods in source-free setting.

Despite impressive advances in large language models (LLMs), existing benchmarks often focus on single-turn or single-step tasks, failing to capture the kind of iterative reasoning required in real-world settings. To address this limitation, we introduce TurnBench, a novel benchmark that evaluates multi-turn, multi-step reasoning through an interactive code-breaking task inspired by the "Turing Machine Board Game." In each episode, a model must uncover hidden logical or arithmetic rules by making sequential guesses, receiving structured feedback, and integrating clues across multiple rounds. This dynamic setup requires models to reason over time, adapt based on past information, and maintain consistency across steps-capabilities underexplored in current benchmarks. TurnBench includes two modes: Classic, which tests standard reasoning, and Nightmare, which introduces increased complexity and requires robust inferential chains. To support fine-grained analysis, we provide ground-truth annotations for intermediate reasoning steps. Our evaluation of state-of-the-art LLMs reveals significant gaps: the best model achieves 84% accuracy in Classic mode, but performance drops to 18% in Nightmare mode. In contrast, human participants achieve 100% in both, underscoring the challenge TurnBench poses to current models. By incorporating feedback loops and hiding task rules, TurnBench reduces contamination risks and provides a rigorous testbed for diagnosing and advancing multi-step, multi-turn reasoning in LLMs.

Reinforcement learning (RL) has become a dominant paradigm for training large language models (LLMs), particularly for reasoning tasks. Effective RL for LLMs requires massive parallelization and poses an urgent need for efficient training systems. Most existing large-scale RL systems for LLMs are synchronous, alternating generation and training in a batch setting where rollouts in each training batch are generated by the same model. This approach stabilizes RL training but suffers from severe system-level inefficiency: generation must wait until the longest output in the batch is completed before model updates, resulting in GPU underutilization. We present AReaL, a fully asynchronous RL system that completely decouples generation from training. Rollout workers in AReaL continuously generate new outputs without waiting, while training workers update the model whenever a batch of data is collected. AReaL also incorporates a collection of system-level optimizations, leading to substantially higher GPU utilization. To stabilize RL training, AReaL balances the workload of rollout and training workers to control data staleness, and adopts a staleness-enhanced PPO variant to better handle outdated training samples. Extensive experiments on math and code reasoning benchmarks show that AReaL achieves up to 2.77$\times$ training speedup compared to synchronous systems with the same number of GPUs and matched or improved final performance. The code of AReaL is available at https://github.com/inclusionAI/AReaL/.

As AI models are deployed with increasing autonomy, it is important to ensure they do not take harmful actions unnoticed. As a potential mitigation, we investigate Chain-of-Thought (CoT) monitoring, wherein a weaker trusted monitor model continuously oversees the intermediate reasoning steps of a more powerful but untrusted model. We compare CoT monitoring to action-only monitoring, where only final outputs are reviewed, in a red-teaming setup where the untrusted model is instructed to pursue harmful side tasks while completing a coding problem. We find that while CoT monitoring is more effective than overseeing only model outputs in scenarios where action-only monitoring fails to reliably identify sabotage, reasoning traces can contain misleading rationalizations that deceive the CoT monitors, reducing performance in obvious sabotage cases. To address this, we introduce a hybrid protocol that independently scores model reasoning and actions, and combines them using a weighted average. Our hybrid monitor consistently outperforms both CoT and action-only monitors across all tested models and tasks, with detection rates twice higher than action-only monitoring for subtle deception scenarios.

The planning, management, and resource scheduling of cellular mobile networks require joint estimation of mobile traffic across different layers and nodes. Mobile traffic generation can proactively anticipate user demands and capture the dynamics of network load. However, existing methods mainly focus on generating traffic at a single spatiotemporal resolution, making it difficult to jointly model multi-scale traffic patterns. In this paper, we propose ZoomDiff, a diffusion-based model for multi-scale mobile traffic generation. ZoomDiff maps urban environmental context into mobile traffic with multiple spatial and temporal resolutions through a set of customized Denoising Refinement Diffusion Models (DRDM). DRDM employs a multi-stage noise-adding and denoising mechanism, enabling different stages to generate traffic at distinct spatiotemporal resolutions. This design aligns the progressive denoising process with hierarchical network layers, including base stations, cells, and grids of varying granularities. Experiments on real-world mobile traffic datasets show that ZoomDiff achieves at least an 18.4% improvement over state-of-the-art baselines in multi-scale traffic generation tasks. Moreover, ZoomDiff demonstrates strong efficiency and cross-city generalization, highlighting its potential as a powerful generative framework for modeling multi-scale mobile network dynamics.

Purpose: To investigate whether a vision-language foundation model can enhance undersampled MRI reconstruction by providing high-level contextual information beyond conventional priors. Methods: We proposed a semantic distribution-guided reconstruction framework that uses a pre-trained vision-language foundation model to encode both the reconstructed image and auxiliary information into high-level semantic features. A contrastive objective aligns the reconstructed representation with the target semantic distribution, ensuring consistency with high-level perceptual cues. The proposed objective works with various deep learning-based reconstruction methods and can flexibly incorporate semantic priors from multimodal sources. To test the effectiveness of these semantic priors, we evaluated reconstruction results guided by priors derived from either image-only or image-language auxiliary information. Results: Experiments on knee and brain datasets demonstrate that semantic priors from images preserve fine anatomical structures and achieve superior perceptual quality, as reflected in lower LPIPS values, higher Tenengrad scores, and improved scores in the reader study, compared with conventional regularization. The image-language information further expands the semantic distribution and enables high-level control over reconstruction attributes. Across all evaluations, the contrastive objective consistently guided the reconstructed features toward the desired semantic distributions while maintaining data fidelity, demonstrating the effectiveness of the proposed optimization framework. Conclusion: The study highlights that vision-language foundation models can improve undersampled MRI reconstruction through semantic-space optimization.

Multimodal Large Language Models (LLMs) hold promise for biomedical reasoning, but current benchmarks fail to capture the complexity of real-world clinical workflows. Existing evaluations primarily assess unimodal, decontextualized question-answering, overlooking multi-agent decision-making environments such as Molecular Tumor Boards (MTBs). MTBs bring together diverse experts in oncology, where diagnostic and prognostic tasks require integrating heterogeneous data and evolving insights over time. Current benchmarks lack this longitudinal and multimodal complexity. We introduce MTBBench, an agentic benchmark simulating MTB-style decision-making through clinically challenging, multimodal, and longitudinal oncology questions. Ground truth annotations are validated by clinicians via a co-developed app, ensuring clinical relevance. We benchmark multiple open and closed-source LLMs and show that, even at scale, they lack reliability -- frequently hallucinating, struggling with reasoning from time-resolved data, and failing to reconcile conflicting evidence or different modalities. To address these limitations, MTBBench goes beyond benchmarking by providing an agentic framework with foundation model-based tools that enhance multi-modal and longitudinal reasoning, leading to task-level performance gains of up to 9.0% and 11.2%, respectively. Overall, MTBBench offers a challenging and realistic testbed for advancing multimodal LLM reasoning, reliability, and tool-use with a focus on MTB environments in precision oncology.

Traditional object detection models in medical imaging operate within a closed-set paradigm, limiting their ability to detect objects of novel labels. Open-vocabulary object detection (OVOD) addresses this limitation but remains un-derexplored in medical imaging due to dataset scarcity and weak text-image alignment. T o bridge this gap, we introduce MedROV, the first Real-time Open V ocabulary detection model for medical imaging. T o enable open-vocabulary learning, we curate a large-scale dataset, Omnis, with 600K detection samples across nine imaging modalities and introduce a pseudo-labeling strategy to handle missing annotations from multi-source datasets. Additionally, we enhance generalization by incorporating knowledge from a large pre-trained foundation model. By leveraging contrastive learning and cross-modal representations, MedROV effectively detects both known and novel structures. Experimental results demonstrate that MedROV outperforms the previous state-of-the-art foundation model for medical image detection with an average absolute improvement of 40 mAP50, and surpasses closed-set detectors by more than 3 mAP50, while running at 70 FPS, setting a new benchmark in medical detection. Our source code, dataset, and trained model are available at MedROV.

What data should a vision-language model be trained on? To answer this question, many data curation efforts center on the quality of a dataset. However, most of these existing methods are (i) offline, i.e. they produce a static dataset from a set of predetermined filtering criteria, and (ii) concept-agnostic, i.e. they use model-based filters which induce additional data biases. In this work, we go beyond such offline, concept-agnostic methods and advocate for more flexible, task-adaptive online concept-based curation. Our first contribution is DataConcept, a collection of 128M web-crawled image-text pairs annotated with fine-grained details about their concept composition. Building on DataConcept, we introduce Concept-Aware Batch Sampling (CABS), a simple yet effective batch sampling framework that flexibly constructs batches on-the-fly based on specific target distributions. We propose two variants: (i) Diversity Maximization (CABS-DM) to curate batches with a broad coverage of available concepts, and (ii) Frequency Maximization (CABS-FM) to curate batches with high object multiplicity. Through extensive evaluations across 28 benchmarks, we demonstrate that our CABS method significantly benefits CLIP/SigLIP model classes and yields highly performant models. Overall, CABS represents a strong open-source alternative to proprietary online data curation algorithms, enabling practitioners to define custom concept distributions that optimize for specific downstream tasks.

Abstract--Long-tailed multi-label visual recognition poses a significant challenge, as images typically contain multiple labels with highly imbalanced class distributions, leading to biased models that favor head classes while underperforming on tail classes. Recent efforts have leveraged pre-trained vision-language models, such as CLIP, alongside long-tailed learning techniques to exploit rich visual-textual priors for improved performance. However, existing methods often derive semantic inter-class relationships directly from imbalanced datasets, resulting in unreliable correlations for tail classes due to data scarcity . Moreover, CLIP's zero-shot paradigm is optimized for single-label image-text matching, making it suboptimal for multi-label tasks. The framework incorporates a graph convolutional network for label-aware propagation and learnable soft prompts for refined embeddings. Moreover, it improves generalization through test-time ensembling and realigns visual-textual modalities using parameter-efficient fine-tuning to avert overfitting on tail classes without compromising head class performance. Recent progress in long-tailed visual recognition has mainly centered on the single-label multi-class setting. However, real-world images often encompass multiple objects and concepts, giving rise to long-tailed multi-label visual recognition (LTML), where each image is associated with multiple labels exhibiting long-tailed distributions [5], [6], [7], [8], [9], [10]. LTML introduces additional intricacies compared to its single-label multi-class counterpart, including label co-occurrences, which create interdependencies among classes, and intra-class imbalances between positive and negative instances. Existing LTML methods have primarily relied on strategies like class re-sampling, loss re-weighting, and specialized architectures, typically built upon ImageNet-pre-trained convolutional neural networks (CNNs) as backbones [5], [6], [7]. Wei Tang is with the School of Computer Science and Engineering, Southeast University, Nanjing 210096, China, the Key Laboratory of Computer Network and Information Integration (Southeast University), MoE, China, and with the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), Abu Dhabi, UAE (e-mail: tangw@seu.edu.cn).