We present Adjacent Possible Exploration (APE), a selective fine-tuning method for adapting large language models that systematically explores parameter modifications while maintaining model stability. Inspired by evolutionary optimization principles, APE evaluates multiple candidate parameter updates through fine-tuning on small data subsets and accepts only those exceeding a performance threshold. Unlike standard fine-tuning that follows single gradient directions, APE implements a filtered selection process that prevents destabilizing parameter changes while enabling systematic improvement. Our method achieves 33.9\% BLEU improvement and 36.2\% perplexity reduction on news summarization tasks while using minimal computational resources. The approach provides a practical framework for controlled model adaptation that balances performance gains with representational stability.
The exponential growth in demand for GPU computing resources, driven by the rapid advancement of Large Language Models, has created an urgent need for automated CUDA optimization strategies. While recent advances in LLMs show promise for code generation, current SOTA models (e.g. R1, o1) achieve low success rates in improving CUDA speed. In this paper, we introduce CUDA-L1, an automated reinforcement learning framework for CUDA optimization. CUDA-L1 achieves performance improvements on the CUDA optimization task: trained on NVIDIA A100, it delivers an average speedup of x17.7 across all 250 CUDA kernels of KernelBench, with peak speedups reaching x449. Furthermore, the model also demonstrates excellent portability across GPU architectures, achieving average speedups of x17.8 on H100, x19.0 on RTX 3090, x16.5 on L40, x14.7 on H800, and x13.9 on H20 despite being optimized specifically for A100. Beyond these benchmark results, CUDA-L1 demonstrates several remarkable properties: 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; 3) Identifies non-obvious performance bottlenecks and rejects seemingly beneficial optimizations that harm performance. The capabilities of CUDA-L1 demonstrate that reinforcement learning can transform an initially poor-performing LLM into an effective CUDA optimizer through speedup-based reward signals alone, without human expertise or domain knowledge. More importantly, the trained RL model extend the acquired reasoning abilities to new kernels. 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.
This paper addresses the challenge of high-fidelity view synthesis of humans with sparse-view videos as input. Previous methods solve the issue of insufficient observation by leveraging 4D diffusion models to generate videos at novel viewpoints. However, the generated videos from these models often lack spatio-temporal consistency, thus degrading view synthesis quality. In this paper, we propose a novel sliding iterative denoising process to enhance the spatio-temporal consistency of the 4D diffusion model. Specifically, we define a latent grid in which each latent encodes the image, camera pose, and human pose for a certain viewpoint and timestamp, then alternately denoising the latent grid along spatial and temporal dimensions with a sliding window, and finally decode the videos at target viewpoints from the corresponding denoised latents. Through the iterative sliding, information flows sufficiently across the latent grid, allowing the diffusion model to obtain a large receptive field and thus enhance the 4D consistency of the output, while making the GPU memory consumption affordable. The experiments on the DNA-Rendering and ActorsHQ datasets demonstrate that our method is able to synthesize high-quality and consistent novel-view videos and significantly outperforms the existing approaches. See our project page for interactive demos and video results: https://diffuman4d.github.io/ .
Learning-based monocular visual odometry (VO) poses robustness, generalization, and efficiency challenges in robotics. Recent advances in visual foundation models, such as DINOv2, have improved robustness and generalization in various vision tasks, yet their integration in VO remains limited due to coarse feature granularity. In this paper, we present DINO-VO, a feature-based VO system leveraging DINOv2 visual foundation model for its sparse feature matching. To address the integration challenge, we propose a salient keypoints detector tailored to DINOv2's coarse features. Furthermore, we complement DINOv2's robust-semantic features with fine-grained geometric features, resulting in more localizable representations. Finally, a transformer-based matcher and differentiable pose estimation layer enable precise camera motion estimation by learning good matches. Against prior detector-descriptor networks like SuperPoint, DINO-VO demonstrates greater robustness in challenging environments. Furthermore, we show superior accuracy and generalization of the proposed feature descriptors against standalone DINOv2 coarse features. DINO-VO outperforms prior frame-to-frame VO methods on the TartanAir and KITTI datasets and is competitive on EuRoC dataset, while running efficiently at 72 FPS with less than 1GB of memory usage on a single GPU. Moreover, it performs competitively against Visual SLAM systems on outdoor driving scenarios, showcasing its generalization capabilities.
FourCastNet 3 advances global weather modeling by implementing a scalable, geometric machine learning (ML) approach to probabilistic ensemble forecasting. The approach is designed to respect spherical geometry and to accurately model the spatially correlated probabilistic nature of the problem, resulting in stable spectra and realistic dynamics across multiple scales. FourCastNet 3 delivers forecasting accuracy that surpasses leading conventional ensemble models and rivals the best diffusion-based methods, while producing forecasts 8 to 60 times faster than these approaches. In contrast to other ML approaches, FourCastNet 3 demonstrates excellent probabilistic calibration and retains realistic spectra, even at extended lead times of up to 60 days. All of these advances are realized using a purely convolutional neural network architecture tailored for spherical geometry. Scalable and efficient large-scale training on 1024 GPUs and more is enabled by a novel training paradigm for combined model- and data-parallelism, inspired by domain decomposition methods in classical numerical models. Additionally, FourCastNet 3 enables rapid inference on a single GPU, producing a 60-day global forecast at 0.25{\deg}, 6-hourly resolution in under 4 minutes. Its computational efficiency, medium-range probabilistic skill, spectral fidelity, and rollout stability at subseasonal timescales make it a strong candidate for improving meteorological forecasting and early warning systems through large ensemble predictions.
Writing GPU kernels is a challenging task and critical for AI systems' efficiency. It is also highly iterative: domain experts write code and improve performance through execution feedback. Moreover, it presents verifiable rewards like correctness and speedup, making it a natural environment to apply Reinforcement Learning (RL). To explicitly incorporate the iterative nature of this process into training, we develop a flexible multi-turn RL recipe that addresses unique challenges encountered in real-world settings, such as learning from long trajectories and effective reward attribution across turns. We present Kevin - K(ernel D)evin, the first model trained with multi-turn RL for CUDA kernel generation and optimization. In our evaluation setup, Kevin shows significant gains over its base model (QwQ-32B), improving correctness of generated kernels (in pure CUDA) from 56% to 82% and mean speedup from 0.53x to 1.10x of baseline (PyTorch Eager), and surpassing frontier models like o4-mini (0.78x). Finally, we study its behavior across test-time scaling axes: we found scaling serial refinement more beneficial than parallel sampling. In particular, when given more refinement turns, Kevin shows a higher rate of improvement.
Inference is now the dominant AI workload, yet existing systems force trade-offs between latency, throughput, and cost. Arctic Inference, an open-source vLLM plugin from Snowflake AI Research, introduces Shift Parallelism, a dynamic parallelism strategy that adapts to real-world traffic while integrating speculative decoding, SwiftKV compute reduction, and optimized embedding inference. It achieves up to 3.4 times faster request completion, 1.75 times faster generation, and 1.6M tokens/sec per GPU for embeddings, outperforming both latency- and throughput-optimized deployments. Already powering Snowflake Cortex AI, Arctic Inference delivers state-of-the-art, cost-effective inference for enterprise AI and is now available to the community.
Transformer models are computationally costly on long sequences since regular attention has quadratic $O(n^2)$ time complexity. We introduce Wavelet-Enhanced Random Spectral Attention (WERSA), a novel mechanism of linear $O(n)$ time complexity that is pivotal to enable successful long-sequence processing without the performance trade-off. WERSA merges content-adaptive random spectral features together with multi-resolution Haar wavelets and learnable parameters to selectively attend to informative scales of data while preserving linear efficiency. Large-scale comparisons \textbf{on single GPU} and across various benchmarks (vision, NLP, hierarchical reasoning) and various attention mechanisms (like Multiheaded Attention, Flash-Attention-2, FNet, Linformer, Performer, Waveformer), reveal uniform advantages of WERSA. It achieves best accuracy in all tests. On ArXiv classification, WERSA improves accuracy over vanilla attention by 1.2\% (86.2\% vs 85.0\%) while cutting training time by 81\% (296s vs 1554s) and FLOPS by 73.4\% (26.2G vs 98.4G). Significantly, WERSA excels where vanilla and FlashAttention-2 fail: on ArXiv-128k's extremely lengthy sequences, it achieves best accuracy (79.1\%) and AUC (0.979) among viable methods, operating on data that gives Out-Of-Memory errors to quadratic methods while being \textbf{twice as fast} as Waveformer, its next-best competitor. By significantly reducing computational loads without compromising accuracy, WERSA makes possible more practical, more affordable, long-context models, in particular on low-resource hardware, for more sustainable and more scalable AI development.
Graph-based Point Cloud Networks (PCNs) are powerful tools for processing sparse sensor data with irregular geometries, as found in high-energy physics detectors. However, deploying models in such environments remains challenging due to stringent real-time requirements for both latency, and throughput. In this work, we present a deeply pipelined dataflow architecture for executing graph-based PCNs on FPGAs. Our method supports efficient processing of dynamic, sparse point clouds while meeting hard real-time constraints. We introduce specialized processing elements for core graph operations, such as GraVNet convolution and condensation point clustering, and demonstrate our design on the AMD Versal VCK190. Compared to a GPU baseline, our FPGA implementation achieves up to 5.25x speedup in throughput while maintaining latencies below 10 {\mu}s, satisfying the demands of real-time trigger systems in particle physics experiments. An open-source reference implementation is provided.
Score-function policy gradients have delivered strong results in game-playing, robotics and language-model fine-tuning. Yet its high-variance often undermines training stability. On the other hand, pathwise policy gradients alleviate the training variance, but are reliable only when driven by an accurate action-conditioned value function which is notoriously hard to train without relying on past off-policy data. In this paper, we discuss how to construct a value-gradient driven, on-policy algorithm that allow training Q-value models purely from on-policy data, unlocking the possibility of using pathwise policy updates in the context of on-policy learning. We show how to balance stochastic policies for exploration with constrained policy updates for stable training, and evaluate important architectural components that facilitate accurate value function learning. Building on these insights, we propose Relative Entropy Pathwise Policy Optimization (REPPO), an efficient on-policy algorithm that combines the sample-efficiency of pathwise policy gradients with the simplicity and minimal memory footprint of standard on-policy learning. We demonstrate that REPPO provides strong empirical performance at decreased sample requirements, wall-clock time, memory footprint as well as high hyperparameter robustness in a set of experiments on two standard GPU-parallelized benchmarks.
Poultry farming is a vital component of the global food supply chain, yet it remains highly vulnerable to infectious diseases such as coccidiosis, salmonellosis, and Newcastle disease. This study proposes a lightweight machine learning-based approach to detect these diseases by analyzing poultry fecal images. We utilize multi-color space feature extraction (RGB, HSV, LAB) and explore a wide range of color, texture, and shape-based descriptors, including color histograms, local binary patterns (LBP), wavelet transforms, and edge detectors. Through a systematic ablation study and dimensionality reduction using PCA and XGBoost feature selection, we identify a compact global feature set that balances accuracy and computational efficiency. An artificial neural network (ANN) classifier trained on these features achieved 95.85% accuracy while requiring no GPU and only 638 seconds of execution time in Google Colab. Compared to deep learning models such as Xception and MobileNetV3, our proposed model offers comparable accuracy with drastically lower resource usage. This work demonstrates a cost-effective, interpretable, and scalable alternative to deep learning for real-time poultry disease detection in low-resource agricultural settings.
Pixel-level annotation is expensive and time-consuming. Semi-supervised segmentation methods address this challenge by learning models on few labeled images alongside a large corpus of unlabeled images. Although foundation models could further account for label scarcity, effective mechanisms for their exploitation remain underexplored. We address this by devising a novel semi-supervised panoptic approach fueled by two dedicated foundation models. We enhance recognition by complementing unsupervised mask-transformer consistency with zero-shot classification of CLIP features. We enhance localization by class-agnostic decoder warm-up with respect to SAM pseudo-labels. The resulting decoupled enhancement of recognition and localization (DEARLi) particularly excels in the most challenging semi-supervised scenarios with large taxonomies and limited labeled data. Moreover, DEARLi outperforms the state of the art in semi-supervised semantic segmentation by a large margin while requiring 8x less GPU memory, in spite of being trained only for the panoptic objective. We observe 29.9 PQ and 38.9 mIoU on ADE20K with only 158 labeled images. The source code is available at https://github.com/helen1c/DEARLi.
In medical image segmentation, convolutional neural networks (CNNs) and transformers are dominant. For CNNs, given the local receptive fields of convolutional layers, long-range spatial correlations are captured through consecutive convolutions and pooling. However, as the computational cost and memory footprint can be prohibitively large, 3D models can only afford fewer layers than 2D models with reduced receptive fields and abstract levels. For transformers, although long-range correlations can be captured by multi-head attention, its quadratic complexity with respect to input size is computationally demanding. Therefore, either model may require input size reduction to allow more filters and layers for better segmentation. Nevertheless, given their discrete nature, models trained with patch-wise training or image downsampling may produce suboptimal results when applied on higher resolutions. To address this issue, here we propose the resolution-robust HNOSeg-XS architecture. We model image segmentation by learnable partial differential equations through the Fourier neural operator which has the zero-shot super-resolution property. By replacing the Fourier transform by the Hartley transform and reformulating the problem in the frequency domain, we created the HNOSeg-XS model, which is resolution robust, fast, memory efficient, and extremely parameter efficient. When tested on the BraTS'23, KiTS'23, and MVSeg'23 datasets with a Tesla V100 GPU, HNOSeg-XS showed its superior resolution robustness with fewer than 34.7k model parameters. It also achieved the overall best inference time (< 0.24 s) and memory efficiency (< 1.8 GiB) compared to the tested CNN and transformer models.
We investigate reinforcement learning (RL) for privileged planning in autonomous driving. State-of-the-art approaches for this task are rule-based, but these methods do not scale to the long tail. RL, on the other hand, is scalable and does not suffer from compounding errors like imitation learning. Contemporary RL approaches for driving use complex shaped rewards that sum multiple individual rewards, \eg~progress, position, or orientation rewards. We show that PPO fails to optimize a popular version of these rewards when the mini-batch size is increased, which limits the scalability of these approaches. Instead, we propose a new reward design based primarily on optimizing a single intuitive reward term: route completion. Infractions are penalized by terminating the episode or multiplicatively reducing route completion. We find that PPO scales well with higher mini-batch sizes when trained with our simple reward, even improving performance. Training with large mini-batch sizes enables efficient scaling via distributed data parallelism. We scale PPO to 300M samples in CARLA and 500M samples in nuPlan with a single 8-GPU node. The resulting model achieves 64 DS on the CARLA longest6 v2 benchmark, outperforming other RL methods with more complex rewards by a large margin. Requiring only minimal adaptations from its use in CARLA, the same method is the best learning-based approach on nuPlan. It scores 91.3 in non-reactive and 90.6 in reactive traffic on the Val14 benchmark while being an order of magnitude faster than prior work.
We present any4, a learned 4-bit weight quantization solution for large language models (LLMs) providing arbitrary numeric representations without requiring pre-processing of weights or activations. any4 yields higher accuracy compared to other related 4-bit numeric representation types: int4, fp4 and nf4, as evaluated on a range of model sizes, generations and families (Llama 2, Llama 3, Mistral and Mixtral). While any4 does not require preprocessing of weights or activations, it is also competitive with orthogonal techniques that require such preprocessing (e.g., AWQ and GPTQ). We also experiment with any3 and any2 and show competitiveness at lower bits. Additionally, we show that we can calibrate using a single curated diverse sample rather than hundreds of samples from a dataset as done in most quantization approaches. We also open source tinygemm, a latency optimized GPU matrix multiplication library for LLMs, that implements any4 using a GPU-efficient lookup table strategy along with other common quantization methods. We open source our code at https://github.com/facebookresearch/any4 .
The production of high-quality 2D animation is highly labor-intensive process, as animators are currently required to draw and color a large number of frames by hand. We present SketchColour, the first sketch-to-colour pipeline for 2D animation built on a diffusion transformer (DiT) backbone. By replacing the conventional U-Net denoiser with a DiT-style architecture and injecting sketch information via lightweight channel-concatenation adapters accompanied with LoRA finetuning, our method natively integrates conditioning without the parameter and memory bloat of a duplicated ControlNet, greatly reducing parameter count and GPU memory usage. Evaluated on the SAKUGA dataset, SketchColour outperforms previous state-of-the-art video colourization methods across all metrics, despite using only half the training data of competing models. Our approach produces temporally coherent animations with minimal artifacts such as colour bleeding or object deformation. Our code is available at: https://bconstantine.github.io/SketchColour .
Kernel development in deep learning requires optimizing computational units across hardware while balancing memory management, parallelism, and hardware-specific optimizations through extensive empirical tuning. Although domain-specific languages like Triton simplify GPU programming by abstracting low-level details, developers must still manually tune critical parameters such as tile sizes and memory access patterns through iterative experimentation, creating substantial barriers to optimal performance and wider adoption. In this work, we introduce AutoTriton, the first model dedicated to Triton programming powered by reinforcement learning (RL). AutoTriton performs supervised fine-tuning (SFT) to be equipped with essential Triton programming expertise using a high-quality data gathering pipeline, and conducts RL with Group Relative Policy Optimization (GRPO) algorithm, combining a rule-based reward and an execution-based reward to further improve Triton programming ability, sequentially. Experiments across five evaluation channels of TritonBench and KernelBench illustrate that our 8B model AutoTriton achieves performance comparable to mainstream large models, including Claude-4-Sonnet and DeepSeek-R1-0528. Further experimental analysis demonstrates the crucial role of each module within AutoTriton, including the SFT stage, the RL stage, and the reward design strategy. These findings underscore the promise of RL for automatically generating high-performance kernels, and since high-performance kernels are core components of AI systems, this breakthrough establishes an important foundation for building more efficient AI systems. The model and code will be available at https://github.com/AI9Stars/AutoTriton.
Diffusion models have achieved remarkable success in various generative tasks, but training them remains highly resource-intensive, often requiring millions of images and many days of GPU computation. From a data-centric perspective addressing this limitation, we study diffusion dataset condensation as a new and challenging problem setting. The goal is to construct a "synthetic" sub-dataset with significantly fewer samples than the original dataset, enabling high-quality diffusion model training with greatly reduced cost. To the best of our knowledge, we are the first to formally investigate dataset condensation for diffusion models, whereas prior work focused on training discriminative models. To tackle this new challenge, we propose a novel Diffusion Dataset Condensation (D2C) framework, which consists of two phases: Select and Attach. The Select phase identifies a compact and diverse subset using a diffusion difficulty score and interval sampling. The Attach phase enhances the selected subset by attaching rich semantic and visual representations to strengthen the conditional signals. Extensive experiments across various dataset sizes, model architectures, and resolutions show that our D2C framework enables significantly faster diffusion model training with dramatically fewer data, while preserving high visual quality. Notably, for the SiT-XL/2 architecture, D2C achieves a 100x training speed-up, reaching a FID score of 4.3 in just 40k steps using only 0.8% of the training data.
Low-Rank Adapters (LoRAs) have transformed the fine-tuning of Large Language Models (LLMs) by enabling parameter-efficient updates. However, their widespread adoption remains limited by the reliance on GPU-based training. In this work, we propose a theoretically grounded approach to LoRA fine-tuning designed specifically for users with limited computational resources, particularly those restricted to standard laptop CPUs. Our method learns a meta-operator that maps any input dataset, represented as a probability distribution, to a set of LoRA weights by leveraging a large bank of pre-trained adapters for the Mistral-7B-Instruct-v0.2 model. Instead of performing new gradient-based updates, our pipeline constructs adapters via lightweight combinations of existing LoRAs directly on CPU. While the resulting adapters do not match the performance of GPU-trained counterparts, they consistently outperform the base Mistral model on downstream tasks, offering a practical and accessible alternative to traditional GPU-based fine-tuning.
Large Language Models (LLMs) have achieved remarkable performance on complex mathematical benchmarks, yet often struggle with simple arithmetic tasks and exhibit a tendency toward over-explaining or "overthinking" answers. To systematically assess this phenomenon, we introduce LLMThinkBench, a modular benchmarking framework that enables researchers to evaluate basic math reasoning and overthinking in LLMs. The framework provides 14 configurable math tasks with randomized test data generation and robust parsing strategies. Researchers can quantify overthinking using our Overthinking Score metric, which captures accuracy-verbosity tradeoffs through harmonic mean formulation. The tool offers flexible evaluation with a scalable vLLM/Transformers backend, multi-GPU support, and full configurability. Users can extend the tool with custom tasks, reproduce experiments with seeding, and generate detailed efficiency reports. Distributed as a pip-installable package with CLI and API access, LLMThinkBench provides researchers and practitioners an accessible, cost-effective alternative to expensive LLM-as-a-judge methods for diagnosing basic reasoning capabilities and efficiency analysis. Package can be installed as: pip install llmthinkbench