Recent Multimodal Large Language Models (MLLMs) excel in vision-language understanding but face challenges in adapting to dynamic real-world scenarios that require continuous integration of new knowledge and skills. While continual learning (CL) offers a potential solution, existing benchmarks and methods suffer from critical limitations. In this paper, we introduce MLLM-CL, a novel benchmark encompassing domain and ability continual learning, where the former focuses on independently and identically distributed (IID) evaluation across evolving mainstream domains, whereas the latter evaluates on non-IID scenarios with new model abilities. Methodologically, we propose preventing catastrophic interference through parameter isolation and an MLLM-based routing mechanism. Extensive experiments demonstrate that our approach can integrate domain-specific knowledge and functional abilities with minimal forgetting, significantly outperforming existing methods. Our benchmark and code are available at https://github.com/bjzhb666/MLLM-CL.

Knowledge distillation (KD) is a widely used framework for training compact, task-specific models by transferring the knowledge from teacher models. However, its application to active learning (AL), which aims to minimize annotation costs through iterative sample selection, remains underexplored. This gap stems from the fact that KD typically assumes access to sufficient labeled data, whereas AL operates in data-scarce scenarios where task-specific teacher models are often unavailable. In this paper, we first introduce ActiveKD, a framework that integrates AL with KD by leveraging the zero- and few-shot capabilities of large vision-language models (VLMs). A key aspect of ActiveKD is the structured prediction bias of VLMs-i.e., their predictions form clusters in the probability space. We regard this structure as an inductive bias of the teacher model, capturing generalizable output patterns beneficial to student learning. To exploit this bias, we propose Probabilistic CoreSet (PCoreSet), a selection strategy that maximizes coverage in the probability space rather than the feature space. PCoreSet strategically selects probabilistically diverse unlabeled samples, facilitating more efficient transfer of teacher knowledge under limited annotation budgets. Extensive evaluations on 11 datasets show that ActiveKD consistently improves performance across selection methods (e.g., +29.07% on ImageNet, averaged over methods). Under ActiveKD, PCoreSet ranks first in 64/73 settings (approximately 87.7%) across 5 student and 3 teacher networks, always achieving the best performance except for first 2 AL rounds. Our code is available at https://github.com/erjui/PCoreSet.

Large Multimodal Models (LMMs) typically build on ViTs (e.g., CLIP), yet their training with simple random in-batch negatives limits the ability to capture fine-grained visual differences, particularly in geometric scenarios. To address this challenge, we propose a novel hard negative contrastive learning framework for the vision encoder, which combines image-based contrastive learning using generation-based hard negatives created by perturbing diagram generation code, and text-based contrastive learning using rule-based negatives derived from modified geometric descriptions and retrieval-based negatives selected based on caption similarity. We train a vision encoder (CLIP) using our hard negative training method, namely MMCLIP (Multimodal Math CLIP), and subsequently train an LMM for geometric problem-solving. Experiments show that our trained model, MMGeoLM, significantly outperforms other open-source models on three geometric reasoning benchmarks. Even with a size of 7B, it can rival powerful closed-source models like GPT -4o. We further conduct ablation studies to analyze three key factors: hard negative types, the efficiency of image-based negatives, and training configurations. These analyses yield important insights into optimizing the training pipeline of vision encoder for fine-grained geometric reasoning tasks. Geometric mathematical reasoning has garnered significant attention as an essential capability for large multimodal models (Anthropic, 2024; OpenAI, 2023; Bai et al., 2023). It requires fine-grained identification of visual elements (Lu et al., 2023) within the given images, such as geometric shapes, spatial configurations, and the relationships between them (He et al., 2024). However, the "eyes" of most existing LMMs, i.e., their pretrained vision encoders such as CLIP (Patel et al., 2024; Y ang et al., 2023; Goel et al., 2022), are primarily trained on general visual datasets that do not emphasize the intricate features necessary for specialized mathematical reasoning. Therefore, these models often fail to understand the nuanced geometric information accurately and produce incorrect reasoning and answers. As shown in Figure 1, facing a simple parallel line problem, the leading LMMs such as GPT -4o (OpenAI, 2024a), Claude-3 (Anthropic, 2024), and Qwen2.5-VL

Ensuring faithful interpretability in large language models is imperative for trustworthy and reliable AI. A key obstacle is self-repair, a phenomenon where networks compensate for reduced signal in one component by amplifying others, masking the true importance of the ablated component. While prior work attributes self-repair to layer normalization and back-up components that compensate for ablated components, we identify a novel form occurring within the attention mechanism, where softmax redistribution conceals the influence of important attention scores. This leads traditional ablation and gradient-based methods to underestimate the significance of all components contributing to these attention scores. We introduce Gradient Interaction Modifications (GIM), a technique that accounts for self-repair during backpropagation. Extensive experiments across multiple large language models (Gemma 2B/9B, LLAMA 1B/3B/8B, Qwen 1.5B/3B) and diverse tasks demonstrate that GIM significantly improves faithfulness over existing circuit identification and feature attribution methods. Our work is a significant step toward better understanding the inner mechanisms of LLMs, which is crucial for improving them and ensuring their safety. Our code is available at https://github.com/JoakimEdin/gim.

Recent advances in AI and ML have transformed data science, yet increasing complexity and expertise requirements continue to hinder progress. Although crowd-sourcing platforms alleviate some challenges, high-level machine learning engineering (MLE) tasks remain labor-intensive and iterative. We introduce R&D-Agent, a comprehensive, decoupled, and extensible framework that formalizes the MLE process. R&D-Agent defines the MLE workflow into two phases and six components, turning agent design for MLE from ad-hoc craftsmanship into a principled, testable process. Although several existing agents report promising gains on their chosen components, they can mostly be summarized as a partial optimization from our framework's simple baseline. Inspired by human experts, we designed efficient and effective agents within this framework that achieve state-of-the-art performance. Evaluated on MLE-Bench, the agent built on R&D-Agent ranks as the top-performing machine learning engineering agent, achieving 35.1% any medal rate, demonstrating the ability of the framework to speed up innovation and improve accuracy across a wide range of data science applications. We have open-sourced R&D-Agent on GitHub: https://github.com/microsoft/RD-Agent.

Evaluating the programming robustness of large language models (LLMs) is paramount for ensuring their reliability in AI-based software development. However, adversarial attacks exhibit fundamental limitations that compromise fair robustness assessment: they demonstrate contradictory evaluation outcomes where different attack strategies tend to favor different models, and more critically, they operate solely through external perturbations, failing to capture the intrinsic stability essential for autonomous coding agents where subsequent inputs are endogenously generated by the model itself. We introduce EVALOOOP, a novel assessment framework that evaluates robustness from a self-consistency perspective, leveraging the natural duality inherent in software engineering tasks (e.g., code generation and code summarization). EVALOOOP establishes a self-contained feedback loop where an LLM iteratively transforms between code and natural language until functional failure occurs, with robustness quantified by a novel Average Sustainable Loops (ASL) metric-the mean number of iterations maintaining functional correctness across benchmark tasks. This cyclical strategy intrinsically evaluates robustness without relying on external attack configurations, providing a unified metric that reveals how effectively LLMs preserve semantic integrity through sustained self-referential transformations. We evaluate 96 popular LLMs, ranging from 0.5B to 685B parameters, on EVALOOOP equipped with the MBPP Plus benchmark, and found that EVALOOOP typically induces a 2.65%-47.62% absolute drop in pass@1 accuracy within ten loops. Intriguingly, robustness does not always align with initial performance (i.e., one-time query); for instance, Qwen3-235B-A22B-Instruct-2507, despite inferior initial code generation compared to OpenAI's o-series models and DeepSeek-V3, demonstrated the superior robustness (ASL score).

Reinforcement learning (RL) has proven effective in strengthening the reasoning capabilities of large language models (LLMs). A widely adopted method, Group Relative Policy Optimization (GRPO), has shown strong empirical results in training DeepSeek-R1. However, GRPO fails to update the policy when all responses within a group are incorrect (i.e., \emph{all-negative-sample} groups). This limitation underscores a key gap between artificial and human intelligence: unlike humans, who can learn from mistakes, GRPO discards these signals. Our first contribution is to introduce a simple framework that mitigates the all-negative-sample issue by incorporating response diversity within groups using a \textit{step-wise} judge model, which can be either directly trained or adapted from existing LLMs. We prove that this diversification can accelerate GRPO's learning dynamics in a simplified setting. We also empirically validate the proposed stepwise guided policy optimization (SGPO) method, demonstrating consistent gains across model sizes (7B, 14B, 32B) in offline and online training on 9 benchmarks, including base and distilled variants. Our results highlight two advantages: (i) SGPO surpasses GRPO, especially in the early and mid-training stages where all-negative-sample groups are prevalent; and (ii) SGPO does not require judge models to generate correct answers, differentiating it from knowledge distillation methods.

Existing efforts in building graphical user interface (GUI) agents largely rely on the training paradigm of supervised fine-tuning (SFT) on large vision-language models (L VLMs). However, this approach not only demands extensive amounts of training data but also struggles to effectively understand GUI screenshots and generalize to unseen interfaces. The issue significantly limits its application in real-world scenarios, especially for high-level tasks. Inspired by reinforcement fine-tuning (RFT) in large reasoning models (e.g., DeepSeek-R1), which efficiently enhances the problem-solving capabilities of large language models in real-world settings, we propose GUI-R1, the first reinforcement learning framework designed to enhance the GUI capabilities of L VLMs in high-level real-world task scenarios, through unified action space rule modeling. By leveraging a small amount of carefully curated high-quality data across multiple platforms (including Windows, Linux, MacOS, Android, and Web) and employing policy optimization algorithms such as group relative policy optimization (GRPO) to update the model, GUI-R1 achieves superior performance using only 0.02% of the data (3K vs. 13M) compared to previous state-of-the-art methods like OS-Atlas across eight benchmarks spanning three different platforms (mobile, desktop, and web). These results demonstrate the immense potential of reinforcement learning based on unified action space rule modeling in improving the execution capabilities of L VLMs for real-world GUI agent tasks.

Top-down images play an important role in safety-critical settings such as autonomous navigation and aerial surveillance, where they provide holistic spatial information that front-view images cannot capture. Despite this, Vision Language Models (VLMs) are mostly trained and evaluated on front-view benchmarks, leaving their performance in the top-down setting poorly understood. Existing evaluations also overlook a unique property of top-down images: their physical meaning is preserved under rotation. In addition, conventional accuracy metrics can be misleading, since they are often inflated by hallucinations or "lucky guesses", which obscures a model's true reliability and its grounding in visual evidence. To address these issues, we introduce TDBench, a benchmark for top-down image understanding that includes 2000 curated questions for each rotation. We further propose RotationalEval (RE), which measures whether models provide consistent answers across four rotated views of the same scene, and we develop a reliability framework that separates genuine knowledge from chance. Finally, we conduct four case studies targeting underexplored real-world challenges. By combining rigorous evaluation with reliability metrics, TDBench not only benchmarks VLMs in top-down perception but also provides a new perspective on trustworthiness, guiding the development of more robust and grounded AI systems. Project homepage: https://github.com/Columbia-ICSL/TDBench

Upcycling pre-trained dense models into sparse Mixture-of-Experts (MoEs) efficiently increases model capacity but often suffers from poor expert specialization due to naive weight replication. Our analysis reveals that upcycled MoEs, even with conventional regularization, exhibit low-confidence, weakly differentiated routing, hindering performance. We introduce Dirichlet-Prior Shaping Loss (DPSL), a novel router regularization technique that directly shapes routing probability distributions by matching expert assignments to a target Dirichlet prior. DPSL offers fine-grained control over expert balance and specialization, and enables encoding of inductive biases such as encouraging experts to focus on specific modalities or tasks, without requiring manual intervention; notably, DPSL is a general tool applicable to any module that outputs categorical probability distributions, extending its utility beyond MoE training. Experiments on upcycled MoE vision-language models (with Qwen2, Phi3, Llama3.2 LLM backbones) show DPSL consistently outperforms upcycling strategies and regularization techniques across standard vision-language benchmarks, addressing the critical issue of poor specialization and fostering more adaptive, higher-performing models.