Unpaired medical image-to-image translation requires synthesizing target-modality images from source-modality images without aligned supervision. Existing methods learn a direct source-to-target mapping that entangles structure preservation and appearance generation within a single model, often distorting clinically meaningful anatomy in the translated output. We argue that these two objectives should be decoupled entirely: anatomical structure should be captured independently of generation, and the generator should operate only on a frozen structural representation. To this end, we propose a two-stage framework for unpaired medical image translation. In the first stage, we learn a shared structure space by training a structure encoder-decoder on source images, target images, and structural masks jointly, using mask reconstruction to anchor the latent space to modality-agnostic anatomy. For source images without mask annotations, a target-domain memory bank provides soft structural supervision via prototype retrieval. In the second stage, a conditional flow model learns to render target-domain appearance under fixed structural conditions. This ensures the generator inherits target appearance statistics and can focus entirely on structural conditioning rather than learning appearance from scratch. We validate our framework on OCT-to-OCTA synthesis and brain MR-to-CT translation. Our method outperforms all unpaired baselines on both tasks and, despite requiring no paired supervision, also surpasses paired baselines, improving PSNR by up to 11.2% and SSIM by up to 26.4% on OCT-to-OCTA, and achieving consistent gains across whole-image, soft-tissue, and bone regions on MR-to-CT. The project repository is available on GitHub.
Agricultural production faces persistent labor shortages and increasing demand for automation, yet progress in agricultural robotics remains constrained by the limited availability of affordable, flexible, and accessible research platforms. This paper presents Agibot, an open-source, modular, low-cost agricultural robotics platform designed for autonomy research and experimentation, with potential applications in hands-on robotics education. The platform integrates four-wheel independent swerve drive, swappable computing, distributed embedded control, multimodal sensing, and configurable manipulation within a unified sense-plan-act-verify framework supporting perception, control, navigation, manipulation, and autonomous task verification. The platform was evaluated through representative row-crop navigation and manipulation tasks under varying lighting and soil conditions. Experiments demonstrated reliable lane following, autonomous row-to-row navigation, target identification, seed dispensing, plant manipulation, and closed-loop verification of task success. Its modular architecture supports independent testing, modification, and validation of autonomy components for rapid prototyping and experimentation. These proof-of-concept tasks, controller-tuning studies, and integrated trials demonstrate the platform's research capabilities, while its accessible and reconfigurable design supports its use for hands-on robotics education. By combining affordability, modularity, open-source accessibility, configurable manipulation, and end-to-end sense-plan-act-verify autonomy, Agibot provides a practical platform for agricultural robotics research, technology development, and workforce training.
A phone screen can be wrong in ways that need no reference: a label that runs off the screen, letters that touch, a line that breaks one character before its end, one button with sharp corners among round ones. We ask whether a model can box every such fault on a screenshot, given no target design and no reference image. Such a critic could serve as a cheap first pass over generated pages before a designer reviews them, but only if it stays silent on clean pages. ScreenProof poses this target-free critique over three families of faults: general faults, local style faults (one element breaks the style its page sets) and text layout faults (letter spacing, line height, line breaks). A data engine makes each fault in page code and records the changed element's box, and hard negatives change every sibling alike. No frontier judge is reliable: frontier models flag 18-62% of clean real screenshots and 18-82% of clean generated pages, find 47-72% of text layout faults, and rarely find a foreign icon (5-19%). Our small critic ProofJudge (Qwen3.5-9B with LoRA), trained on code-made pages with one to four faults each and keeping a finding only when it is confident, reaches F1 0.48, the highest of all judges (Gemini 4 Argon 0.37; open models at most 0.15; 0.46 even without the confidence cut), and is the only fault-finding judge within 15% false alarms on clean pages (2% real, 9% generated), the property such a first pass needs, at about a hundredth of Argon's cost. It finds the most general faults (59%) but fewer local style faults than Argon (25% against 41%) and fewer text layout faults than the API judges (43%; Gemini 4 Argon 47%). Against a named brand the task is harder still: models know a design language about as well as its public documentation states it.
Mathematical understanding depends on the route from hypotheses to conclusion: how the parts of a proof function and which dependencies carry the argument forward. This structure can persist when wording or notation changes, while a small alteration to a hypothesis or inference can produce a different argument beneath similar prose. Formal verification can establish that a rewrite proves the same theorem, but not whether it follows the same mathematical route. Comparing informal proofs, especially those produced or revised by generative models, therefore requires reconstructing which claims, dependencies, and reasoning steps are preserved, reorganized, or replaced. We present ProofDiff, a system for comparing informal mathematical documents. It is designed to support human review, not to certify mathematical correctness. Given two documents, ProofDiff independently constructs a source-grounded argument map for each and proposes correspondences between the resulting maps. Each map captures the document's claims and reasoning steps and records the typed relations that connect them. An integrated human-review layer allows reviewers to revise the maps and correspondences while preserving an auditable history. Comparing the two argument maps provides a basis for inspecting whether the documents follow the same mathematical route and where their claims or dependencies differ. We demonstrate ProofDiff on four illustrative pairs of AI-generated proofs, ranging from a detailed proof and its concise revision to proofs using different mathematical strategies.
The development of tabular foundation models (TFMs) has accelerated in recent years, outperforming traditional ML methods on many tabular benchmarks. A key finding is that TFMs can be pretrained entirely on synthetic datasets, opening opportunities to design data generators that encourage desirable model properties. Prior work has mainly focused on crafting high-quality priors over generators to improve overall pretraining performance. Our key insight is that parameterizing the generator distribution enables an adversarial, distributional robustness perspective: during training, we can adapt the generator to emphasize datasets that are particularly challenging for the model, which we formalize by introducing an optimality gap measure. Further, we develop a method to align the synthetic data generation with real-world datasets from a given domain, constraining the adversary to generate realistic data. Together, these algorithms comprise the Robust and Domain-Adapted Tabular Foundation Models (RAD-TFM) pipeline, a model-agnostic adversarial training framework. Applied to the TabPFN V2 classifier, RAD-TFM improves performance across 6 diverse tabular benchmarks, with up to a 10.8 percentage-point improvement in mean normalized AUC over the strongest baseline, using fewer than 100k additional synthetic training datasets, less than 0.1% of the original pretraining data. Applying the same pipeline to TabICL V2 also improves performance across three major benchmarks, demonstrating the generality of the framework. These results highlight a promising new direction for targeted adversarial continued training of TFMs using synthetic data alone.
Weak gravitational lensing is a powerful cosmological probe that enables direct mapping of the matter distribution and precise constraints on the fundamental parameters governing structure formation in the universe. In recent years, machine learning has shown strong potential for extracting cosmological information from weak-lensing data beyond traditional summary statistics. However, progress has been limited by the lack of standardized benchmarks that capture the complexity of the dataset, simulator imperfections, and various systematic biases encountered in realistic analyses. To address this gap, we present the FAIR Universe Weak Lensing Machine Learning Uncertainty Challenge, one of the NeurIPS 2025 competitions, as a benchmark dataset and evaluation study. Our dataset reflects realistic weak-lensing observations, incorporating multiple systematic biases, and thereby providing a controlled testbed for robust cosmological inference under uncertainties. We summarize the main methodological advances developed by the winners of our challenge, including data augmentation with conditional optimal transport, an AI-agent-driven approach for parameter inference, ensemble learning, Vision Transformer-based architectures, and physics-informed feature integration. By unifying the benchmark dataset and the winning solutions, this work provides a community resource for developing and evaluating machine learning methods for weak lensing cosmology and, more broadly, for scientific inference under systematic uncertainties.
Bilevel optimization is a fundamental framework for machine learning problems, where an upper-level (UL) objective depends on the solution of a nested lower-level (LL) problem. Existing bilevel algorithms typically assume that the LL objective is available for evaluation, differentiation, or unrolling. However, this assumption can fail when the LL response is produced by a black-box solver, simulator, adversary, or learned predictor. In this paper, we study LL agnostic bilevel optimization, where the UL learner has access only to an approximate LL solution $\bar{\mathbf{y}}(\mathbf{x})$ and cannot query the LL objective, its gradients, Hessian/Jacobian information, or the procedure that generates the LL response. We propose LeGo-BiO, a coordinate-wise pseudo-hypergradient method that estimates the missing Jacobian $\nabla _{\mathbf{x}}\bar{\mathbf{y}}(\mathbf{x})$ using finite differences of $\bar{\mathbf{y}}(\cdot)$. For each UL coordinate, LeGo-BiO constructs a decision vector subject to 1-sparse updates, thereby isolating the corresponding Jacobian column needed in the UL gradient computation. This avoids the directional-projection limitation of full-parameter response differences and, unlike derivative-free/zeroth-order approximations of the entire UL gradient, preserves the available first-order gradient information with respect to the UL variables for more accurate UL gradient evaluation. We establish convergence guarantees for LeGo-BiO in both deterministic and stochastic settings under an LL error-bound/Polyak-Lojasiewicz-type condition, without requiring LL strong convexity. Our bounds yield an $O(T^{-1})$ rate in the deterministic setting and an $O(T^{-1/2})$ rate in the stochastic setting when the LL approximation error decreases geometrically. Experiments demonstrate competitive performance of LeGo-BiO despite being agnostic to the LL objective.
Modern Mixture-of-Experts (MoE) inference is constrained by GPU memory. This constraint is mitigated by expert offloading, which stores part of the model's weight in host memory and executes on demand. Though MoE architecture has computation sparsity, more GPUs are needed for serving increasingly larger frontier models. However, existing expert offloading systems overlook this need and fail to scale performance with additional GPUs. We present AdaOffload, a general and efficient multi-GPU hybrid expert execution framework that effectively handles heterogeneous hardware and workloads. We co-design GPU-CPU parallelism and GPU-CPU co-execution for decode and prefill with online workload observation that adapts to different hardware, workload, and batch size heterogeneity. We evaluated AdaOffload on five different platforms with varying GPU capability (L4, Pro6000, B200), number of GPUs, CPU capability, and NUMA topology on three different frontier open-weight models (284B to 753B). AdaOffload achieves $1.9\times$ to $15.5\times$ throughput speedups across different models over state-of-the-art systems.
Neural operators learn mappings from PDE inputs, such as initial conditions and coefficient fields, to the corresponding solutions. Once trained, they can rapidly predict solutions for new inputs. However, obtaining input-solution pairs for supervised training often requires expensive numerical simulations, whereas PDE inputs and governing equations provide inexpensive physical information. This raises a question: how can cheap physical information reduce the need for labeled solutions? A natural approach is to use physics to select inputs for labeling, but our acquisition experiments show no consistent advantage over random sampling. This observation motivates exploring an alternative: using cheap physics to improve neural operator initialization before supervised training. We propose Physics-Informed Initialization (PI-Init), a simple yet effective method that pretrains neural operators on unlabeled inputs using physics-informed losses and retains the learned parameters for supervised fine-tuning on the labeled dataset. A linearized analysis shows that pretraining can improve downstream prediction accuracy by reducing initialization errors that limited labeled data cannot effectively correct. Experiments across diverse PDE benchmarks and neural operator architectures demonstrate improvements in predictive accuracy, out-of-distribution generalization, and label efficiency. Representative results show error reductions of up to 62.7% at the same label budget and comparable or better accuracy with 90% fewer labels, relative to supervised-only training.
In recent years, singular-spectrum-based optimizers have attracted increasing attention due to their ability to preserve the matrix structure of neural network gradients, thereby enabling geometry-aware optimization. However, most singular-spectrum-based (SSB) approaches remain centralized, limiting their use in distributed learning infrastructures that exploit massive parallelism. Distributed SSB optimization is still at an early stage, with open questions about convergence guarantees and communication efficiency. To fill this gap, we propose the Distributed Momentum-based Spectral Estimator (DisMUSE), an efficient distributed SSB optimization framework for both data-center and federated settings. Using Muon-style exact orthogonalized momentum as the spectral update, DisMUSE integrates a recursive variance-reduced momentum estimator and achieves an $O((KT)^{-1/3})$ convergence rate, where $K$ and $T$ denote the number of nodes and local iterations per node, respectively. This rate matches the best-known rates for variance-reduced nonconvex distributed optimization. Building on this variance-reduced framework, we study two extensions. First, our DisMUSE method with finite Newton-Schulz (NS) iterations preserves the same convergence order under a uniform normalized support-spectrum condition, up to a multiplicative factor. Second, we develop a communication-efficient variant based on low-rank (LR) momentum transmission and quantify the error it introduces to the stationarity bound, characterizing the trade-off between accuracy and communication cost.
We present two compact decoder-only language models trained from scratch exclusively on Greek text: LexisLM-1B (1.09B parameters) and LexisLM-500M (502M parameters). Both use a dedicated 32k byte-level BPE tokenizer and a 22B-token pretraining corpus drawn primarily from Greek Wikipedia and FineWeb2. Both models complete the 22B-token schedule, with best held-out perplexities of 9.77 for LexisLM-1B and 11.31 for LexisLM-500M. Evaluation of both models on Greek-translated ARC, MMLU, and TruthfulQA shows above-chance performance on ARC-Easy, while ARC-Challenge and MMLU remain near chance. We also perform supervised fine-tuning of both models on Greek instruction data, primarily EuroAlpaca, reaching held-out instruction perplexities of 7.12 for LexisLM-500M and 6.37 for LexisLM-1B. The report describes the tokenizer, data preparation, training, and evaluation methods, providing a foundation for further work on compact Greek language models.
We demonstrate that transformer architectures trained with discrete denoising diffusion and decoded via guided test-time beam search can learn to recover the prime factors of semiprime integers directly from data. Our approach establishes a new frontier for learning-based semiprime factorization, successfully factoring 68-bit integers-a substantial advance over the prior 56-bit state of the art. While computationally less efficient than classical factoring algorithms, our model requires no domain-specific number-theoretic priors and learns purely from training examples. In practice, the diffusion model acts as a generative prior, identifying promising factor candidates where exact verification or lattice-based post-processing can complete the factorization. Furthermore, we derive empirical scaling laws relating approximation quality to training data volume, model size, and test-time computation. In contrast to language, problem size is the key driver of scaling behavior. Achieving a fixed accuracy on a held-out test set requires compute that scales exponentially with bit length. Consequently, our findings pose no threat to security, as the projected compute cost required to factorize cryptographically relevant integers remains prohibitive.
Interactive VR/MR applications demand photorealistic 3D representations that support real-time rendering and object-level manipulation. While 3D Gaussian Splatting (3DGS) achieves high rendering quality, it represents scenes as unstructured, monolithic primitives. Existing object-aware 3DGS methods jointly optimize geometry and semantic states but suffer from two key failure modes: noisy foundation-model masks on complex geometry, and label-geometry coupling, in which co-optimized identity gradients mislabel boundary Gaussians. We introduce RigidGS, an end-to-end pipeline that decouples geometric learning from object identity. We propose a depth- and temporally-aware mask refinement module that addresses two complementary SAM2 failures: over-inclusion on thin structures and under-inclusion on multimodal-depth objects. We then determine object identities entirely upstream via COLMAP projection voting and freeze them during training, directing mask supervision solely toward geometric optimization. Refined masks alone substantially improve competing methods, yet a residual gap in label-geometry coupling persists that our fixed-identity design eliminates. RigidGS outperforms prior object-aware 3DGS methods on LLFF/NVOS, Mip-NeRF~360, and iNeRF360, improving segmentation and boundary IoU by 7.7 and 14.5 points on LLFF/NVOS. Finally, RigidGS enables real-time rigid-body manipulation on Meta Quest~2 without mesh extraction. Our code and project page are available at https://gyreyirisansah.github.io/rigidgs.
High-dimensional Bayesian optimisation (HDBO) is an increasingly crowded field with new methods proposed at every major conference, making it ever harder to answer a deceptively simple question: "Which method is currently state-of-the-art?" In this paper, we question the efficacy of benchmarking practices in the field, and find them to be systematically ineffective: multiple published works fail to benchmark against the strongest available baselines, or even include incorrect implementations of key methods. Without a shared reference implementation, such errors are difficult to detect during peer review. We therefore introduce BO-Arena, an actively maintained and extensible package providing canonical implementations of state-of-the-art algorithms, which allows researchers to develop new methods and benchmark them rigorously. Using BO-Arena, we find that what currently drives performance in HDBO is even simpler than previously thought. We use these insights to propose a new algorithm which is competitive with the state of the art on problems with many observations, without resorting to the complex GP approximations used in recent methods. Finally, we question whether there is currently sufficient evidence that any more complex method outperforms a simple Gaussian process baseline for HDBO.
Large language models (LLMs) increasingly support Web-based services that must balance response quality and inference cost across heterogeneous requests. Conventional routing selects one model per prompt, while fixed ensembles cannot adapt their composition and cost to individual requests. We introduce PISA, a prompt-adaptive, cost-aware single-model router, and PISA-E, its extension to ensemble routing. At the core of both methods, PISA combines soft cluster memberships with model-specific predictors to estimate individual model capabilities. Building on these estimates, PISA-E jointly selects ensemble composition and cardinality under an explicit cost--quality preference that can be adjusted at inference time without retraining. A tractable ensemble-success surrogate and efficient candidate selection enable this extension without ensemble-specific training. Across four LLM routing benchmarks under in-distribution (ID) and out-of-distribution (OOD) conditions, PISA achieves the highest macro-average cost--performance hypervolume among the evaluated single-model routers in both settings. Meanwhile, PISA-E achieves the highest macro-average task performance among the evaluated ensemble approaches when prioritizing response quality, surpassing the strongest competitors by 2.09 and 2.87 percentage points under ID and OOD conditions, respectively. At selected lower-cost operating points, it also exceeds the strongest ensemble competitors' macro-average task performance while reducing average inference cost by approximately 36% and 25%, respectively. These results demonstrate the effectiveness of prompt-adaptive capability estimation for both single-model routing and selective ensembling.
In quantum systems, physical quantities of interest can be expressed as expectation values of a probability distribution, which allows their estimation by Markov Chain Monte Carlo (MCMC) methods. This approach, an instance of Quantum Monte Carlo (QMC), plays a central role in obtaining predictions of system properties relevant to condensed matter and high energy physics. The use of Hamiltonian dynamics to design a Markov kernel, known as Hybrid or Hamiltonian Monte Carlo (HMC), is widespread in quantum applications. In this context, it is standard to use an HMC kernel that is ``adjusted'' with the Metropolis-Hastings (MH) criterion, so that estimates of expectations converge to their true value in the limit of infinite sample size. However, recent work in computational statistics suggests that an \emph{unadjusted} HMC kernel can be significantly more efficient, while safely maintaining an asymptotic bias which is small relative to the error arising from the variance of the finite sample size. We adapt this approach to the QMC setting, focusing on a challenging and high-dimensional \emph{spin-fermion} model of a quantum phase transition. We find that unadjusted methods outperform adjusted HMC in terms of effective sample size per gradient call, while retaining accurate results.
Neural additive models (NAMs) offer interpretability but lack expressivity due to their rigid additive structure, whereas multi-layer perceptrons (MLPs) achieve universal approximation at the cost of transparency. We establish the first theoretical foundation for projected neural additive models (PNAMs), which augment NAMs with a learnable linear transformation $\boldsymbol{T}$ to capture cross-variable interactions. While standard NAMs cannot approximate even the simplest interaction term $\chi_1 \chi_2$ because their cross-derivatives are identically zero, we prove that PNAMs are universal approximators and characterize the minimum number of projections required for exact polynomial representation. Furthermore, we derive an approximation rate of $\mathcal{O}(M^{-s / (N - 1)})$ for $\mathcal{C}^s$ functions (i.e., functions with $s$ continuous derivatives), where $M$ denotes the number of projections and $N$ the input dimension; this rate is known to be optimal for ridge function approximation. To recover interpretability from the projected inputs, we introduce structured sparsity regularizers and post-hoc symbolic regression techniques that enable global input ranking, parameter pruning, and the conversion of learned bases into compact analytical expressions. Experiments on knot theory invariants, MNIST, and phase field fracture mechanics demonstrate that PNAMs match or surpass similar architectures in accuracy while discovering salient features and yielding symbolic models with orders of magnitude fewer parameters.
It is often desirable to generate and release synthetic graphs, which resemble real-world data, in a privacy-preserving way. While substantial progress has been made in designing private graph synthesizers (PGS), opportunities remain for more explicit connections between synthesizer models and privacy evaluation, and better accounting for output variation. This paper proposes GSynPEF, a practical framework for empirical evaluation of privacy and utility in settings for which a real-world graph is given as input and output synthetic graph(s) are observed. Our framework links a broad class of graph synthesizers (both statistical and deep-learning based) with privacy metrics, expanding evaluation dimensions to practical utility and risk, through empirical analysis of multiple actual outputs. The framework accommodates both an Exact Solution scenario, where synthesizer behavior is directly observable, and a Monte Carlo Approximation scenario where only synthetic outputs from a black box synthesizer are available. Extensive experiments on four real-world datasets and seven widely used graph synthesizers demonstrate framework's versatility, enabling privacy evaluation for different attack goals (e.g., to infer an edge or node) and capabilities (e.g., observing one or multiple synthetic graphs). We analyze evaluation results by examining the relationship between graph statistics and synthesizer design choices, providing insights for data curators in PGS selection and researchers for improved PGS design.
Recognizing anatomical structures during minimally invasive surgery (MIS) could support intraoperative navigation and improve patient safety, but progress is limited by small, single-procedure datasets and by segmentation methods developed for natural scenes. This work addresses both limitations through a new dataset and model. First, we introduce ATLAS-120k, a clip-level semantic segmentation dataset of 121,018 annotated frames from 100 surgical videos, covering 14 laparoscopic and robot-assisted procedures and 41 classes. Annotations were created with a pipeline that combines manual labeling, tracker-based propagation, and review by surgeons. Second, we propose ATLAS (Anatomy Recognition with Context Learning using Foundation Representations), a video semantic segmentation model for surgical anatomy recognition. ATLAS builds on in-domain foundation-model embeddings and adds context queries for the procedure and surgical phase, together with query propagation between frames for short-term memory. ATLAS outperforms state-of-the-art natural-image and in-domain foundation models in both segmentation accuracy and temporal consistency, achieving a Dice score of 0.70 with ViT-L while running in real time at 76.2 FPS. Ablations show that the context queries improve accuracy and that query propagation improves temporal consistency. ATLAS also generalizes well, matching a dedicated model zero-shot on CholecSeg8k and improving phase recognition on AutoLaparo over SurgeNet initialization. The model weights, dataset annotations, and annotation tool are publicly available at https://github.com/TimJaspers0801/ATLAS.
Medical ultrasound (US) enables real-time, non-invasive diagnosis, yet building US-based AI systems is limited by data scarcity, costly expert annotation, and limited cross-center generalizability. Privacy constraints and instance-level reporting hinder the collection of diverse abnormal cases and frame-level labels, while repeated retraining for new centers increases computational cost. We introduce EvoDiff, a generative active learning framework that integrates diffusion modelling with information-theoretic objectives, evolutionary search, and curriculum-based sampling to select informative samples across heterogeneous US datasets. EvoDiff formulates sample selection as a generate-to-retrieve process, in which diffusion-based optimization in feature space guided by information gain produces synthetic probes, and only the real unlabeled samples retrieved by these probes are annotated and used for downstream training. Across six benchmarks covering fifteen datasets and 16,696 images, we evaluate EvoDiff on convolutional network and transformer-based foundation model. EvoDiff approaches fully supervised performance with substantially fewer labels, outperforms active learning baselines, and improves foundation model generalization under cross-domain external validation on unseen cohorts. A reader study with six radiologists indicates the potential to support clinical annotation workflows, with a model-based estimated 24% to 84% reduction in annotation workload at matched downstream performance.