Interactive notebook programming is universal in modern ML and AI workflows, with interactive deep learning training (IDLT) emerging as a dominant use case. To ensure responsiveness, platforms like Jupyter and Colab reserve GPUs for long-running notebook sessions, despite their intermittent and sporadic GPU usage, leading to extremely low GPU utilization and prohibitively high costs. In this paper, we introduce NotebookOS, a GPU-efficient notebook platform tailored for the unique requirements of IDLT. NotebookOS employs replicated notebook kernels with Raft-synchronized replicas distributed across GPU servers. To optimize GPU utilization, NotebookOS oversubscribes server resources, leveraging high inter-arrival times in IDLT workloads, and allocates GPUs only during active cell execution. It also supports replica migration and automatic cluster scaling under high load. Altogether, this design enables interactive training with minimal delay. In evaluation on production workloads, NotebookOS saved over 1,187 GPU hours in 17.5 hours of real-world IDLT, while significantly improving interactivity.
We present OWLViz, a challenging benchmark for Open WorLd VISual question answering that evaluates multimodal AI systems on realistic, practical tasks. OWLViz features 248 carefully curated questions requiring the integration of multiple capabilities: common-sense knowledge, visual understanding, web exploration, and specialized tool usage. The benchmark specifically challenges models with visually degraded inputs, complex multi-step reasoning involving counting and measurement operations, and knowledge-intensive queries requiring external information retrieval from minimal visual cues. While humans achieve 69.2% accuracy on these intuitive tasks in under one minute, even state-of-the-art VLMs struggle dramatically, with the best model, Gemini 2.5 Pro, achieving only 27.09% accuracy. Current tool-calling agents and GUI agents, which rely on vision and vision-language models as tools, perform even worse, often failing to engage with available tools effectively. This substantial performance gap reveals critical limitations in multimodal systems' ability to select appropriate tools, coordinate heterogeneous resources, and execute complex reasoning sequences. OWLViz establishes new directions for advancing practical, open-world AI research and agent development.
In many settings, language models struggle with multi-hop question answering. Ideally, humans could help, but how? On which reasoning sub-tasks would humans do well? To answer this, we recruited 40 untrained crowd-workers to perform the subtasks that comprise a multi-hop answering pipeline: recognize a question as complex; break down a complex question; retrieve answers to simple questions and assemble simple facts to answer complex questions. Our tasks were based on the challenging 2WikiMultiHopQA benchmark where human accuracy was 80.2% on direct complex question answering and 84.1% on simple question answering. We found our participants struggled most with recognizing that a question might be complex (67% accuracy) but were better at question decomposition (78.2%) and answer integration (97.3%). This suggests that if a system knew it was struggling with a multi-hop question, it could ask a human to break it down into simpler questions and integrate the answer.
Recognising individual animals from photographs is an important practical challenge in wildlife conservation. Camera traps can capture thousands of images, but researchers still need reliable ways to tell whether two photographs show the same animal, not only the same species. This helps with population monitoring, movement tracking, territory studies, and long-term conservation decisions. The problem is difficult because the same animal may appear from different angles, in different lighting, or partly hidden, while different animals can look very similar. This study presents a practical computer vision workflow for animal re-identification using a small, labelled dataset, with jaguars used as the main case study. The research compares two broad approaches. The first fine-tunes specialised or large vision models for the jaguar task. The second keeps a strong general-purpose vision model frozen, uses it to extract image features once, and trains only a small metric-learning head on top. In the recorded experiments, the frozen DINOv2-Giant approach achieved stronger benchmark results than the earlier fine-tuned MegaDescriptor-L system, reaching 0.958 public identity-balanced mAP, compared with 0.840 for the prior champion. These results suggest that, for small wildlife datasets, fine-tuning a large, specialised model is not always necessary; a powerful frozen foundation model can sometimes deliver better inference quality with much lower training overhead. The broader contribution is not a claim that one model will always win, but a reproducible model-selection workflow that ML engineers and conservation technologists can adapt to other pattern-bearing species such as tigers, leopards, zebras, giraffes, and whale sharks. Because AI and machine learning are evolving quickly, practitioners should treat this study as a decision framework and benchmark example, while also evaluating newer models, training methods, and deployment practices as they become available.
Typography is an essential component of visual communication. Editing and animating text appearance for graphic designs, title sequences, commercials, and logos remain highly skilled tasks requiring detailed, hands-on efforts from pro artists. Automating these challenging manual workflows requires balancing the competing goals of maintaining a text's legibility and aesthetics, while enabling creative expression. Variable fonts, recent parametric extensions to traditional fonts, offer the promise of new ways to ease and automate typographic design and animation. Variable fonts provide custom-constructed parameters along which fonts can be smoothly varied. These parameterizations could then potentially serve as high-value continuous design spaces, opening the door to modern automated design-optimization tools. However, currently variable fonts are underutilized in creative applications, exactly because artists so far still need to manually tune font parameters. Our work provides intuitive and automated font design and animation workflows with differentiable variable fonts. To do so we distill the current variable font specification to a compact mathematical formulation that differentiably connects the highly non-linear, non-invertible mapping of variable font parameters to the underlying vector graphics representing the text. In turn, this enables us to construct a differentiable framework, with respect to variable font parameters, that allows us to perform gradient-based optimization of energies defined on vector graphics control points, and likewise, via differentiable SVG rasterization, on target rasterized images. We demonstrate the utility of this framework with a range of applications, including direct shape manipulation, overlap aware modeling, physics-based text animation, and automated font-design optimization. Our work now enables leveraging the carefully designed affordances of variable fonts with differentiability to use modern design-optimization technologies, and so opens new possibilities for easy, intuitive and expressive typographic design workflows.