The exponential growth in demand for GPU computing resources has created an urgent need for automated CUDA optimization strategies. While recent advances in LLMs show promise for code generation, current SOTA models achieve low success rates in improving CUDA speed. In this paper, we introduce CUDA-L1, an automated reinforcement learning framework for CUDA optimization that employs a novel contrastive RL algorithm. CUDA-L1 achieves significant performance improvements on the CUDA optimization task: trained on A100, it delivers an average speedup of x3.12 with a median speedup of x1.42 against default baselines over across all 250 CUDA kernels of KernelBench, with peak speedups reaching x120. In addition to the default baseline provided by KernelBench, CUDA-L1 demonstrates x2.77 over Torch Compile, x2.88 over Torch Compile with reduce overhead, x2.81 over CUDA Graph implementations, and remarkably x7.72 over cuDNN libraries. Furthermore, the model also demonstrates portability across different GPU architectures. Beyond these benchmark results, CUDA-L1 demonstrates several properties: it 1) discovers a variety of CUDA optimization techniques and learns to combine them strategically to achieve optimal performance; 2) uncovers fundamental principles of CUDA optimization, such as the multiplicative nature of optimizations; 3) identifies non-obvious performance bottlenecks and rejects seemingly beneficial optimizations that actually harm performance. The capabilities demonstrate that, RL can transform an initially poor-performing LLM into an effective CUDA optimizer through speedup-based reward signals alone, without human expertise or domain knowledge. This paradigm opens possibilities for automated optimization of CUDA operations, and holds promise to substantially promote GPU efficiency and alleviate the rising pressure on GPU computing resources.

As valuable digital assets, deep neural networks necessitate robust ownership protection, positioning neural network watermarking (NNW) as a promising solution. Among various NNW approaches, weight-based methods are favored for their simplicity and practicality; however, they remain vulnerable to forging and overwriting attacks. To address those challenges, we propose NeuralMark, a robust method built around a hashed watermark filter. Specifically, we utilize a hash function to generate an irreversible binary watermark from a secret key, which is then used as a filter to select the model parameters for embedding. This design cleverly intertwines the embedding parameters with the hashed watermark, providing a robust defense against both forging and overwriting attacks. An average pooling is also incorporated to resist fine-tuning and pruning attacks. Furthermore, it can be seamlessly integrated into various neural network architectures, ensuring broad applicability. Theoretically, we analyze its security boundary. Empirically, we verify its effectiveness and robustness across 13 distinct Convolutional and Transformer architectures, covering five image classification tasks and one text generation task. The source codes are available at https://github.com/AIResearch-Group/NeuralMark.

Large language models (LLMs) make it possible to generate synthetic behavioural data at scale, offering an ethical and low-cost alternative to human experiments. Whether such data can faithfully capture psychological differences driven by personality traits, however, remains an open question. We evaluate the capacity of LLM agents, conditioned on Big-Five profiles, to reproduce personality-based variation in susceptibility to misinformation, focusing on news discernment, the ability to judge true headlines as true and false headlines as false. Leveraging published datasets in which human participants with known personality profiles rated headline accuracy, we create matching LLM agents and compare their responses to the original human patterns. Certain trait-misinformation associations, notably those involving Agreeableness and Conscientiousness, are reliably replicated, whereas others diverge, revealing systematic biases in how LLMs internalize and express personality. The results underscore both the promise and the limits of personality-aligned LLMs for behavioral simulation, and offer new insight into modeling cognitive diversity in artificial agents.

Large language model (LLM) agents show promise in automating machine learning (ML) engineering. However, existing agents typically operate in isolation on a given research problem, without engaging with the broader research community, where human researchers often gain insights and contribute by sharing knowledge. To bridge this gap, we introduce MLE-Live, a live evaluation framework designed to assess an agent's ability to communicate with and leverage collective knowledge from a simulated Kaggle research community. Building on this framework, we propose CoMind, an multi-agent system designed to actively integrate external knowledge. CoMind employs an iterative parallel exploration mechanism, developing multiple solutions simultaneously to balance exploratory breadth with implementation depth. On 75 past Kaggle competitions within our MLE-Live framework, CoMind achieves a 36% medal rate, establishing a new state of the art. Critically, when deployed in eight live, ongoing competitions, CoMind outperforms 92.6% of human competitors on average, placing in the top 5% on three official leaderboards and the top 1% on one.

Large Language Models (LLMs) increasingly rely on knowledge graphs for factual reasoning, yet how retrieval design shapes their performance remains unclear. We examine how question decomposition changes the retrieved subgraph's content and structure. Using a hybrid retrieval function that controls the importance of initial question and sub-questions, we show that subquestion-based retrieval improves content precision, but yields disjoint subgraphs, while question-based retrieval maintains structure at the cost of relevance. Optimal performance arises between these extremes, revealing that balancing retrieval content and structure is key to effective LLM reasoning over structured knowledge.

Deep actor-critic algorithms have reached a level where they influence everyday life. They are a driving force behind continual improvement of large language models through user feedback. However, their deployment in physical systems is not yet widely adopted, mainly because no validation scheme fully quantifies their risk of malfunction. We demonstrate that it is possible to develop tight risk certificates for deep actor-critic algorithms that predict generalization performance from validation-time observations. Our key insight centers on the effectiveness of minimal evaluation data. A small feasible set of evaluation roll-outs collected from a pretrained policy suffices to produce accurate risk certificates when combined with a simple adaptation of PAC-Bayes theory. Specifically, we adopt a recently introduced recursive PAC-Bayes approach, which splits validation data into portions and recursively builds PAC-Bayes bounds on the excess loss of each portion's predictor, using the predictor from the previous portion as a data-informed prior. Our empirical results across multiple locomotion tasks, actor-critic methods, and policy expertise levels demonstrate risk certificates tight enough to be considered for practical use.

Motivated by the success of general-purpose large language models (LLMs) in software patching, recent works started to train specialized patching models. Most works trained one model to handle the end-to-end patching pipeline (including issue localization, patch generation, and patch validation). However, it is hard for a small model to handle all tasks, as different sub-tasks have different workflows and require different expertise. As such, by using a 70 billion model, SOTA methods can only reach up to 41% resolved rate on SWE-bench-Verified. Motivated by the collaborative nature, we propose Co-PatcheR, the first collaborative patching system with small and specialized reasoning models for individual components. Our key technique novelties are the specific task designs and training recipes. First, we train a model for localization and patch generation. Our localization pinpoints the suspicious lines through a two-step procedure, and our generation combines patch generation and critique. We then propose a hybrid patch validation that includes two models for crafting issue-reproducing test cases with and without assertions and judging patch correctness, followed by a majority vote-based patch selection. Through extensive evaluation, we show that Co-PatcheR achieves 46% resolved rate on SWE-bench-Verified with only 3 x 14B models. This makes Co-PatcheR the best patcher with specialized models, requiring the least training resources and the smallest models. We conduct a comprehensive ablation study to validate our recipes, as well as our choice of training data number, model size, and testing-phase scaling strategy.

The go-to strategy to apply deep networks in settings where uncertainty informs decisions--ensembling multiple training runs with random initializations--is ill-suited for the extremely large-scale models and practical fine-tuning workflows of today. We introduce a new cost-effective strategy for improving the uncertainty quantification and downstream decisions of a large model (e.g. a fine-tuned ViT-B): coupling it with a less accurate but much smaller "sidekick" (e.g. a fine-tuned ResNet-34) with a fraction of the computational cost. We propose aggregating the predictions of this Asymmetric Duo by simple learned weighted averaging. Surprisingly, despite their inherent asymmetry, the sidekick model almost never harms the performance of the larger model. In fact, across five image classification benchmarks and a variety of model architectures and training schemes (including soups), Asymmetric Duos significantly improve accuracy, uncertainty quantification, and selective classification metrics with only ${\sim}10-20\%$ more computation.

Training data attribution (TDA) methods aim to measure how training data impacts a model's predictions. While gradient-based attribution methods, such as influence functions, offer theoretical grounding, their computational costs make them impractical for large-scale applications. Representation-based approaches are far more scalable, but typically rely on heuristic embeddings that are not optimized for attribution, limiting their fidelity. To address these challenges, we propose AirRep, a scalable, representation-based approach that closes this gap by learning task-specific and model-aligned representations optimized explicitly for TDA. AirRep introduces two key innovations: a trainable encoder tuned for attribution quality, and an attention-based pooling mechanism that enables accurate estimation of group-wise influence. We train AirRep using a ranking objective over automatically constructed training subsets labeled by their empirical effect on target predictions. Experiments on instruction-tuned LLMs demonstrate that AirRep achieves performance on par with state-of-the-art gradient-based approaches while being nearly two orders of magnitude more efficient at inference time. Further analysis highlights its robustness and generalization across tasks and models. Our code is available at https://github.com/sunnweiwei/AirRep

Web navigation is a unique domain that can automate many repetitive real-life tasks and is challenging as it requires long-horizon sequential decision making beyond typical multimodal large language model (MLLM) tasks. Yet, specialized reward models for web navigation that can be utilized during both training and test-time have been absent until now. Despite the importance of speed and cost-effectiveness, prior works have utilized MLLMs as reward models, which poses significant constraints for real-world deployment. To address this, in this work, we propose the first process reward model (PRM) called Web-Shepherd which could assess web navigation trajectories in a step-level. To achieve this, we first construct the WebPRM Collection, a large-scale dataset with 40K step-level preference pairs and annotated checklists spanning diverse domains and difficulty levels. Next, we also introduce the WebRewardBench, the first meta-evaluation benchmark for evaluating PRMs. In our experiments, we observe that our Web-Shepherd achieves about 30 points better accuracy compared to using GPT-4o on WebRewardBench. Furthermore, when testing on WebArena-lite by using GPT-4o-mini as the policy and Web-Shepherd as the verifier, we achieve 10.9 points better performance, in 10 less cost compared to using GPT-4o-mini as the verifier. Our model, dataset, and code are publicly available at LINK.