Chest radiography (CXR) plays a crucial role in the diagnosis of various diseases. However, the inherent class imbalance in the distribution of clinical findings presents a significant challenge for current self-supervised deep learning models. These models often fail to accurately classify long-tailed classes. Current Vision-Language models such as Contrastive Language Image Pre-training (CLIP) models effectively model the manifold distribution of the latent space, enabling high zero-shot classification accuracies. Although CLIP performs well on most of the primary classes in the dataset, our work reveals that its effectiveness decreases significantly for classes with a long-tailed distribution. Our approach employs a class-weighting mechanism that directly aligns with the distribution of classes within the latent space. This method ensures a substantial improvement in overall classification performance, with particular emphasis on enhancing the recognition and accuracy of rarely observed classes. We accomplish this by applying Gaussian Mixture Model (GMM) clustering to the latent space. The subsequent clusters are further refined by Student t-distribution, followed by a metric loss that utilizes the altered embeddings. Our approach facilitates stable and adaptive clustering of the features. This results in a notable average improvement of 7\% points in zero-shot AUC scores across 40 classes in the MIMIC-CXR-JPG dataset from previous SOTA models.

Large Language Models (LLMs) for code generation evolve rapidly through fine-tuning, merging, or new model releases. However, such updates can introduce regressions, not only in correctness but also in code quality and performance. To address this, we present ReCatcher, a regression testing framework for Python code generation. ReCatcher systematically compares two LLMs, typically a current model and a candidate update, across three dimensions: logical correctness, static code quality, and execution performance. We apply ReCatcher to assess regressions across three update scenarios, fine-tuning, merging, and model release, using CodeLlama, DeepSeek-Coder, and GPT-4o. Our evaluation shows that fine-tuning with cross-language datasets increases syntax errors by up to 12%. Merging with general-purpose models like Llama2 leads to regressions in correctness by up to 18%. GPT-4o introduces regressions of up to 50% in handling missing imports compared to GPT-3.5-turbo, while GPT-4o-mini suffers up to 80% performance degradation in execution time versus GPT-4o. Overall, logical correctness, performance, and error handling (e.g., syntax errors and missing imports) are the most regression-prone areas. Comparing ReCatcher with baseline solutions, it presents better and consistent accuracy across logical and performance aspects. ReCatcher highlights the importance of systematic regression evaluation before adopting new models, while assisting researchers and practitioners in making more informed update decisions.

We introduce MMBench-GUI, a hierarchical benchmark for evaluating GUI automation agents across Windows, macOS, Linux, iOS, Android, and Web platforms. It comprises four levels: GUI Content Understanding, Element Grounding, Task Automation, and Task Collaboration, covering essential skills for GUI agents. In addition, we propose a novel Efficiency-Quality Area (EQA) metric to assess GUI agent execution efficiency in online automation scenarios. Through MMBench-GUI, we identify accurate visual grounding as a critical determinant of overall task success, emphasizing the substantial benefits of modular frameworks that integrate specialized grounding modules. Furthermore, to achieve reliable GUI automation, an agent requires strong task planning and cross-platform generalization abilities, with long-context memory, a broad action space, and long-term reasoning playing a critical role. More important, task efficiency remains a critically underexplored dimension, and all models suffer from substantial inefficiencies, with excessive redundant steps even when tasks are ultimately completed. The integration of precise localization, effective planning, and early stopping strategies is indispensable to enable truly efficient and scalable GUI automation. Our benchmark code, evaluation data, and running environment will be publicly available at https://github.com/open-compass/MMBench-GUI.

Many recent papers have studied the development of superforecaster-level event forecasting LLMs. While methodological problems with early studies cast doubt on the use of LLMs for event forecasting, recent studies with improved evaluation methods have shown that state-of-the-art LLMs are gradually reaching superforecaster-level performance, and reinforcement learning has also been reported to improve future forecasting. Additionally, the unprecedented success of recent reasoning models and Deep Research-style models suggests that technology capable of greatly improving forecasting performance has been developed. Therefore, based on these positive recent trends, we argue that the time is ripe for research on large-scale training of superforecaster-level event forecasting LLMs. We discuss two key research directions: training methods and data acquisition. For training, we first introduce three difficulties of LLM-based event forecasting training: noisiness-sparsity, knowledge cut-off, and simple reward structure problems. Then, we present related ideas to mitigate these problems: hypothetical event Bayesian networks, utilizing poorly-recalled and counterfactual events, and auxiliary reward signals. For data, we propose aggressive use of market, public, and crawling datasets to enable large-scale training and evaluation. Finally, we explain how these technical advances could enable AI to provide predictive intelligence to society in broader areas. This position paper presents promising specific paths and considerations for getting closer to superforecaster-level AI technology, aiming to call for researchers' interest in these directions.

Large language models (LLMs) are increasingly adapted to downstream tasks via reinforcement learning (RL) methods like Group Relative Policy Optimization (GRPO), which often require thousands of rollouts to learn new tasks. We argue that the interpretable nature of language can often provide a much richer learning medium for LLMs, compared with policy gradients derived from sparse, scalar rewards. To test this, we introduce GEPA (Genetic-Pareto), a prompt optimizer that thoroughly incorporates natural language reflection to learn high-level rules from trial and error. Given any AI system containing one or more LLM prompts, GEPA samples system-level trajectories (e.g., reasoning, tool calls, and tool outputs) and reflects on them in natural language to diagnose problems, propose and test prompt updates, and combine complementary lessons from the Pareto frontier of its own attempts. As a result of GEPA's design, it can often turn even just a few rollouts into a large quality gain. Across four tasks, GEPA outperforms GRPO by 10% on average and by up to 20%, while using up to 35x fewer rollouts. GEPA also outperforms the leading prompt optimizer, MIPROv2, by over 10% across two LLMs, and demonstrates promising results as an inference-time search strategy for code optimization.

Large language models (LLMs) face low hardware efficiency during decoding, especially for long-context reasoning tasks. This paper introduces Step-3, a 321B-parameter VLM with hardware-aware model-system co-design optimized for minimizing decoding costs. Step-3 innovates in two key dimensions: (1) A novel Multi-Matrix Factorization Attention (MFA) mechanism that significantly reduces both KV cache size and computation while maintaining high attention expressiveness, and (2) Attention-FFN Disaggregation (AFD), a distributed inference system that decouples attention and Feed-Forward Network (FFN) layers into specialized subsystems. This co-design achieves unprecedented cost efficiency: Step-3 significantly reduces theoretical decoding costs compared with models like DeepSeek-V3 and Qwen3 MoE 235B, with the gains widening at longer context. Step-3 achieves low cost while activating 38B parameters per token (more than DeepSeek-V3 and Qwen3 MoE 235B), demonstrating that hardware-aligned attention arithmetic intensity, MoE sparsity, and AFD are critical to cost-effectiveness. We perform a head-to-head comparison with DeepSeek-V3 in its favorable scenarios. Our implementation on Hopper GPUs achieves a decoding throughput of up to 4,039 tokens per second per GPU under 50ms TPOT SLA (4K context, FP8, no MTP). It is higher than DeepSeek-V3's 2,324 in the same setup and sets a new Pareto frontier for LLM decoding.

The processing mechanisms underlying language and image understanding in large vision-language models (LVLMs) have been extensively studied. However, the internal reasoning mechanisms of LVLMs for spatiotemporal understanding remain poorly understood. In this work, we introduce a systematic, circuit-based framework designed to investigate how spatiotemporal visual semantics are represented and processed within these LVLMs. Specifically, our framework comprises three circuits: visual auditing circuit, semantic tracing circuit, and attention flow circuit. Through the lens of these circuits, we discover that visual semantics are highly localized to specific object tokens--removing these tokens can degrade model performance by up to 92.6%. Furthermore, we identify that interpretable concepts of objects and actions emerge and become progressively refined in the middle-to-late layers of LVLMs. In contrary to the current works that solely focus on objects in one image, we reveal that the middle-to-late layers of LVLMs exhibit specialized functional localization for spatiotemporal semantics. Our findings offer significant mechanistic insights into spatiotemporal semantics analysis of LVLMs, laying a foundation for designing more robust and interpretable models.

Building neural systems that can learn to execute symbolic algorithms is a challenging open problem in artificial intelligence, especially when aiming for strong generalization and out-of-distribution performance. In this work, we introduce a general framework for learning convergent term rewriting systems using a neuro-symbolic architecture inspired by the rewriting algorithm itself. We present two modular implementations of such architecture: the Neural Rewriting System (NRS) and the Fast Neural Rewriting System (FastNRS). As a result of algorithmic-inspired design and key architectural elements, both models can generalize to out-of-distribution instances, with FastNRS offering significant improvements in terms of memory efficiency, training speed, and inference time. We evaluate both architectures on four tasks involving the simplification of mathematical formulas and further demonstrate their versatility in a multi-domain learning scenario, where a single model is trained to solve multiple types of problems simultaneously. The proposed system significantly outperforms two strong neural baselines: the Neural Data Router, a recent transformer variant specifically designed to solve algorithmic problems, and GPT-4o, one of the most powerful general-purpose large-language models. Moreover, our system matches or outperforms the latest o1-preview model from OpenAI that excels in reasoning benchmarks.

This position paper examines the use of Large Language Models (LLMs) in social simulation, analyzing both their potential and their limitations from a computational social science perspective. The first part reviews recent findings on the ability of LLMs to replicate key aspects of human cognition, including Theory of Mind reasoning and social inference, while also highlighting significant limitations such as cognitive biases, lack of true understanding, and inconsistencies in behavior. The second part surveys emerging applications of LLMs in multi-agent simulation frameworks, focusing on system architectures, scale, and validation strategies. Notable projects such as Generative Agents (Smallville) and AgentSociety are discussed in terms of their design choices, empirical grounding, and methodological innovations. Particular attention is given to the challenges of behavioral fidelity, calibration, and reproducibility in large-scale LLM-driven simulations. The final section distinguishes between contexts where LLMs, like other black-box systems, offer direct value-such as interactive simulations and serious games-and those where their use is more problematic, notably in explanatory or predictive modeling. The paper concludes by advocating for hybrid approaches that integrate LLMs into traditional agent-based modeling platforms (GAMA, Netlogo, etc), enabling modelers to combine the expressive flexibility of language-based reasoning with the transparency and analytical rigor of classical rule-based systems.

Recently, recurrent large language models (Recurrent LLMs) with linear computational complexity have re-emerged as efficient alternatives to self-attention-based LLMs (Self-Attention LLMs), which have quadratic complexity. However, Recurrent LLMs often underperform on long-context tasks due to their limited fixed-size memory. Previous research has primarily focused on enhancing the memory capacity of Recurrent LLMs through architectural innovations, but these approaches have not yet enabled Recurrent LLMs to match the performance of Self-Attention LLMs on long-context tasks. We argue that this limitation arises because processing the entire context at once is not well-suited for Recurrent LLMs. In this paper, we propose Smooth Reading, a chunk-wise inference method inspired by human reading strategies. Smooth Reading processes context in chunks and iteratively summarizes the contextual information, thereby reducing memory demands and making the approach more compatible with Recurrent LLMs. Our experimental results show that this method substantially narrows the performance gap between Recurrent and Self-Attention LLMs on long-context tasks, while preserving the efficiency advantages of Recurrent LLMs. Our Smooth Reading boosts SWA-3B-4k (a Recurrent LLM) from 5.68% lower to 3.61% higher performance than Self-Attention LLMs on LongBench. Besides, our method maintains the high efficiency, training 3x faster and inferring 2x faster at 64k context compared to Self-Attention LLMs. To our knowledge, this is the first work to achieve comparable performance using Recurrent LLMs compared with Self-Attention LLMs on long-context tasks. We hope our method will inspire future research in this area. To facilitate further progress, we will release code and dataset.