Domain shift (the difference between source and target domains) poses a significant challenge in clinical applications, e.g., Diabetic Retinopathy (DR) grading. Despite considering certain clinical requirements, like source data privacy, conventional transfer methods are predominantly model-centered and often struggle to prevent model-targeted attacks. In this paper, we address a challenging Online Model-aGnostic Domain Adaptation (OMG-DA) setting, driven by the demands of clinical environments. This setting is characterized by the absence of the model and the flow of target data. To tackle the new challenge, we propose a novel approach, Generative Unadversarial ExampleS (GUES), which enables adaptation from a data-centric perspective. Specifically, we first theoretically reformulate conventional perturbation optimization in a generative way--learning a perturbation generation function with a latent input variable. During model instantiation, we leverage a Variational AutoEncoder to express this function. The encoder with the reparameterization trick predicts the latent input, whilst the decoder is responsible for the generation. Furthermore, the saliency map is selected as pseudo-perturbation labels. Because it not only captures potential lesions but also theoretically provides an upper bound on the function input, enabling the identification of the latent variable. Extensive comparative experiments on DR benchmarks with both frozen pre-trained models and trainable models demonstrate the superiority of GUES, showing robustness even with small batch size.
Stochastic differential equations (SDEs) are a fundamental tool for modelling dynamic processes, including gene regulatory networks (GRNs), contaminant transport, financial markets, and image generation. However, learning the underlying SDE from data is a challenging task, especially if individual trajectories are not observable. Motivated by burgeoning research in single-cell datasets, we present the first comprehensive approach for jointly identifying the drift and diffusion of an SDE from its temporal marginals. Assuming linear drift and additive diffusion, we show that non-identifiability can only arise if the initial distribution possesses generalized rotational symmetries. We further prove that even if this condition holds, the drift and diffusion can almost always be recovered from the marginals. Additionally, we show that the causal graph of any SDE with additive diffusion can be recovered from the identified SDE parameters. To complement this theory, we adapt entropy-regularized optimal transport to handle anisotropic diffusion and introduce APPEX (Alternating Projection Parameter Estimation from $X_0$), an iterative algorithm designed to estimate the drift, diffusion, and causal graph of an additive noise SDE, solely from temporal marginals, and demonstrate its effectiveness on simulated data from linear additive noise SDEs.
Strong data augmentation is a fundamental component of state-of-the-art mean teacher-based Source-Free domain adaptive Object Detection (SFOD) methods, enabling consistency-based self-supervised optimization along weak augmentation. However, our theoretical analysis and empirical observations reveal a critical limitation: strong augmentation can inadvertently erase class-relevant components, leading to artificial inter-category confusion. To address this issue, we introduce Weak-to-strong Semantics Compensation (WSCo), a novel remedy that leverages weakly augmented images, which preserve full semantics, as anchors to enrich the feature space of their strongly augmented counterparts. Essentially, this compensates for the class-relevant semantics that may be lost during strong augmentation on the fly. Notably, WSCo can be implemented as a generic plug-in, easily integrable with any existing SFOD pipelines. Extensive experiments validate the negative impact of strong augmentation on detection performance, and the effectiveness of WSCo in enhancing the performance of previous detection models on standard benchmarks.
Where should reconstructions and new samples lie when a few observations reveal only fragments of a trajectory? We propose hierarchical Bayesian trajectory principal-manifold sampling (HB-TPM), which learns a family of smooth curves from ordered source tasks and adapts it to sparse targets. For unordered vectors, alternating progression assignments and coefficient updates estimate trajectory structure between and beyond the observed shots under the learned source family. Observation-aware extensions connect the trajectory prior to photon arrivals and image-encoded cardiac cycles. Our analysis characterizes the coefficient directions supplied by the prior and bounds sampling error conditional on curve recovery. Across 30 held-out vector tasks, HB-TPM reduces mean sliced-Wasserstein distance by 21.0-41.8% relative to a source-adapted VAE. Prior-strength and sampling ablations identify coefficient shrinkage as the main supported mechanism. On 100 simulated photon records, adaptation lowers mean relative trajectory RMSE by 50.2% versus target-only fitting and 23.6% versus source-mean prediction. Across three training seeds and five folds on 98 cardiac sequences, HB-TPM with latent-MSE training achieves the best mean reconstruction among the tested configurations. Together, the results demonstrate trajectory-family transfer for sparse-target reconstruction and trajectory-centered sampling when source and target tasks share comparable progression structure.
We theoretically and empirically establish a previously unexplored connection between Long-Tailed Recognition (LTR) and Continual Learning (CL). Specifically, we show that training on a long-tailed dataset with imbalance factor $\mathrm{IF}$ drives model parameters into an $O(1/\sqrt{\mathrm{IF}})$ neighborhood of the dominant-class solution, under both uniform and exponentially decaying cardinality distributions. Building on this result, we establish an objective-level reformulation of LTR as a sequential learning problem and prove that the CL objective upper-bounds the balanced LTR loss up to a constant determined by the residual Head loss. Motivated by this connection, we introduce Continual Learning for Long-Tailed Recognition (CLTR), a principled framework that leverages standard off-the-shelf CL methods to sequentially learn Head and Tail classes while mitigating catastrophic forgetting. Extensive experiments on CIFAR100-LT, CIFAR10-LT, ImageNet-LT, and Caltech256 validate our theoretical predictions and demonstrate strong performance across diverse LTR benchmarks. In the foundation-model regime, CLTR with DualPrompt on a CLIP backbone outperforms standard adaptation baselines and is competitive with specialized methods using external semantic supervision. Our work bridges LTR and CL both theoretically and empirically, providing a principled approach for addressing long-tailed learning using standard CL strategies.
Groups can achieve more than individuals in the same amount of time, but only if their members coordinate: they must decide who works on what and what to share. However, this organization in multi-agent systems today usually comes from arbitrary, static harnesses designed a priori without adapting to agent behaviors or task specifications. We ask whether agents can instead learn to organize themselves. We introduce s$^3$, the first multi-agent reinforcement learning method in which agents decide for themselves how long to work alone and when to share findings or read teammates' messages, with no predefined organization or roles. Our s$^3$ agents learn to take different approaches, read each other's work, exhibit emergent roles that complement each other in the swarm, and generalize to bigger teams. We study both coordination, where agents must combine private information, and parallel exploration, where sharing discoveries can make search more efficient. With matched training and inference budgets, s$^3$, swarms evaluated with twice as many agents as in training solve over 25$\times$ more coordination problems than the strongest baseline, and in exploration, communicating agents reduce repeated work by nearly 50% and outperform best-of-N by up to 9.4 percentage points. Rather than being designed in advance, s$^3$ enables organization to emerge from a shared goal. This opens the door to capable systems built from many models that learn to work together.
Calorimeter showers are the most expensive part of detector simulation, and highly granular calorimeters, in which a single hadronic shower leaves thousands of energy deposits, make them more expensive still. We present CaloCHILL (Calorimeter Hits In Learned Latents), a generative model that does not produce these deposits one by one: instead it generates a compact summary of the shower in a learned latent space and expands it to full resolution only at the end. The complete shower, including the arrival time of every hit, is built in three stages: a flow-matching model first sets the number of hits and the earliest hit time in each calorimeter layer, a mean-flow model then generates the latent summary, and the decompressor of a variational autoencoder turns it into the full point cloud, giving every hit its position, energy and time. Sparse attention along the shower axis keeps this compression affordable for showers with thousands of hits. On positive-pion showers in the calorimeter of the International Large Detector, a proposed detector for the FCC-ee, CaloCHILL reproduces the main shower observables, including their time structure. This new model is 600 times faster than its predecessor, while improving on all accuracy metrics considered.
As robots are deployed in the real world, they will inevitably make mistakes. Human interventions provide valuable feedback for improving their behavior, enabling them to adapt online to new scenarios. However, studying human-in-the-loop robot learning is difficult: real-world experiments are slow and expensive, so most algorithm development relies on simulated human models. We show that common human modeling assumptions do not match how real people intervene in practice. Through analyzing intervention data from human participants, we find that humans tend to intervene when agent progress stalls, and that their corrective actions are often suboptimal. Motivated by these findings, we introduce a progress-based model of human intervention that achieves higher rank correlation with real-user algorithm rankings, enabling realistic algorithm benchmarking and development in simulation. Building on these insights, we propose Q-OIL, a method that adds optimistic bonuses to intervention transitions in reinforcement learning and softly guides exploration toward human corrections. We find that decoupled critics are critical for stability because they prevent the bonuses from backing up to pre-intervention states. As a result, Q-OIL learns more efficiently than competitive baselines, achieving 10-50% higher success in simulation and on real robots, and improving pretrained VLAs on two bimanual manipulation tasks. As we ground both evaluation and algorithms in realistic human behavior, this work provides a practical path toward scalable robot learning from interventions.
Personalized health coaching requires moving beyond retrospective wearable summaries toward quantitative, actionable behavioral prescriptions. We introduce Coachable Health Age (CHA), an interpretable health-age metric built from modifiable physiological and behavioral signals passively captured by consumer wearables. Across the WEAR-ME (N=2,780), All of Us (N=22,748), and longitudinal Fitbit (N=1,464) cohorts, CHA aligns with established cardiometabolic health constructs and improves prognostic discrimination for incident cardiovascular and metabolic disease beyond age, sex, and body mass index. To enable prospective behavioral optimization, we developed Sensor World Model (SensorWM), an action-conditioned wearable world model pretrained on over 2.2 million person-weeks of longitudinal sensor data from approximately 100,000 participants. SensorWM accurately forecasts multi-week physiological responses to candidate lifestyle changes ($R^2=0.664$), outperforming generic time-series foundation models while preserving downstream clinical signal for diabetes, hypertension, and other health outcomes. Finally, we integrate CHA and SensorWM into an agentic counterfactual coaching system to synthesize personalized 12-week coaching plans. In blinded paired reviews, certified lifestyle coaches preferred counterfactual-grounded plans over plans grounded in wearable history alone for overall quality (71.4%, P < 0.001) and anticipated adherence (71.0%, P < 0.001), and board-certified clinicians judged 99.2% of counterfactual-grounded plans to meet clinical safety criteria (vs. 97.6% of comparison plans). Automated evaluations across the full cohort (N=2,780) also favored counterfactual-grounded plans. By closing the loop between longitudinal tracking, trajectory prediction, and counterfactual coaching, our framework provides a scalable path to systematically improve cardiometabolic health.
Large language models (LLMs) are attractive rerankers for recommendation. Running an expensive model on every request is costly, and for users with sparse interaction histories it often lowers quality. We present CascadeRec, a two-tier LLM reranking cascade deployed in production at a large financial planning and analysis (FP&A) SaaS provider. It serves content recommendations to enterprise tenants on AWS EKS and GKE. A lightweight Tier-1 model reranks every request. A router escalates only the requests where extra inference-time compute is likely to help, and sends them to a more expensive Tier-2 procedure (self-consistency sampling, multi-path reasoning, or a larger model). The system has run in a production A/B experiment for over a month. Compared with uniform use of the expensive tier, CascadeRec has reduced inference cost by about 10%, with minimal latency impact and no quality regression observed to date. Tenant data cannot be released, so we also evaluate the pipeline on public benchmarks (MovieLens-1M and Amazon-Electronics). Segment-level analysis with paired-bootstrap confidence intervals shows that uniform LLM reranking lowers NDCG@10 for sparse-history users by up to 30%, and its only statistically significant gain appears on the segment with the longest histories. Escalating cold users can make the loss larger instead of repairing it. Which segments benefit from escalation depends on the domain, and this motivates the segment-gated rollout we use in production. We describe the deployment architecture, the measured benefits, and lessons learned, including inter-tenant load imbalance as a driver for cascading and the handling of LLM output-parsing failures at the serving layer.
Natural language offers an intuitive way to direct a virtual crowd while a simulation is running. Handwritten rules execute efficiently but cannot respond flexibly to varied instructions, whereas asking a large language model (LLM) to control the crowd at every update introduces repeated latency, cost, and the risk of unsupported actions. We present CrowdDirector, a framework that uses an LLM only when the user adds or changes an instruction. The LLM translates the request into a structured instruction, and a scene validator checks whether the entities and actions it references exist in the current Unity scene. Accepted instructions are stored and remain continuously active as the simulation is running. At each simulation tick, a graph neural network uses the current scene state and active instruction to choose an available action and target for each agent, while Unity handles movement, interactions, and animation. The network is trained offline on simulated scenarios, with LLM guidance on how agents should respond to instructions. Our evaluations show the scene validator's feasibility decisions matched human judgements more often than the LLM's own assessment, and on a behavioural benchmark CrowdDirector follows user instructions more closely than handwritten rules while avoiding repeated LLM calls. It therefore provides a practical middle ground between fixed rules and continuous LLM control.
Goals are central to how people, teams, and organizations structure their work. Yet, while generative artificial intelligence (AI) systems increasingly operate across a worker's context, information about why a person is pursuing a task, beyond local prompts and interactions, is typically missing. We argue that goals should become an explicit representational layer mediating interactions among people, AI systems, and organizations. Because human goals are hierarchical, partially latent, and dynamically evolving, such a layer must be multidimensional, both grounded in human articulation and inferred from available evidence, and continuously updated. A goal layer could connect immediate requests to broader outcomes, improve the organization and retrieval of relevant context, support reflection and coordination, and provide a foundation for training, aligning, and evaluating AI systems in terms of human-centered outcomes. We highlight these opportunities while identifying open challenges involving goal inference, socio-cognitive frictions, alignment, privacy, and governance.
No-reference image quality assessment (IQA) of low-dose CT predicts the score that radiologists would assign to a slice without a full-dose reference, as needed to compare dose protocols and denoisers. Trained CT IQA networks agree well with radiologists but output only a score, whereas multimodal large language models (MLLMs) can explain a score but perform poorly on CT when only prompted. We propose a measurement-aware MLLM, built on Qwen3.8-27B, that rates a slice from its image together with six image-domain measurements of noise, sharpness, high-frequency content and streaks, given to the language model as percentiles in text. On the LDCT-IQA 2023 test set, it reaches an Overall score of 2.735, on par with the retrained challenge winner (2.735) and MedSigLIP-FiLM (2.713), and substantially above its base MLLM used as a zero-shot agent. Shuffling the measurements at test time reduces the score by 0.048, mostly through noise, but the same measurements also lead the model to assign higher scores to over-smoothed denoised images.
Gemini's Nano Banana models, first introduced in late 2025, led to a shift in the capabilities and use of image generation and editing. We utilize Urania, a Google-developed framework for generating insights about LLM chatbot interactions with rigorous differential privacy (DP) guarantees, to analyze use cases and intents, techniques and ways of prompting, without ever looking at a single image or individual prompt. We select a subsample of 2.1M conversations in early 2026 which utilized Nano Banana models, into roughly 800 clusters in five categories: (1) Social & Viral Sharing, (2) Brand & Marketing Collateral, (3) Educational & Technical Information Media, (4) Product Visualizations & E-commerce, and (5) Architectural & Spatial Designs. We then compare these clusters on 21 features of use, measuring aspects of prompts' (a) editing techniques, (b) artistic style requests, (c) composition, and (d) chatbot interactions. This allows us to describe a baseline of how people are using image generation and editing tools for different use cases---(including growing and emergent use cases which were not possible before Nano Banana)---in a privacy-preserving manner.
Evaluations of large language models (LLMs) for mental health support commonly focus on the clinical appropriateness of individual responses, potentially overlooking failures that emerge across extended interactions. As AI assistants accumulate memory of users' circumstances and prior disclosures, this raises an accountability question: Do AI chatbots provide appropriate support as users' needs evolve? We investigate by analyzing 14,256 longitudinal simulations grounded in donated chatbot histories from 22 users, examining how a memory-equipped assistant responds to suicidal ideation, self-harm, and anxiety crises across recovery trajectories and communication styles. We introduce the Context-Aware and Readiness-Evaluative (CARE) framework, developed with mental health professionals, to assess the timing and quality of advice, including omissions. Advice was less aligned with users' readiness after a crisis than in matched control stretches of the same conversation, largely because of premature recommendations, which were most common when recovery stalled. Revisiting a crisis reduced premature advice but not lagging advice, and users' communication style changed how often and when the assistant gave recommendations. Although explicitly unproductive advice was rare, the assistant frequently omitted clinician-expected support such as follow-up and planning for next steps. These findings show the limits of single response-level safety evaluation and the need to assess timing, omission, and consistency across users when holding AI systems accountable for longitudinal support.
We present FoleyForge, a generative model for instruction-based editing of sound effects. Given an input clip and a free-form instruction describing the desired change, it applies the edit while preserving the input's identity. A single model handles semantic, stylistic, and DSP-like changes. We introduce a data pipeline that guides an audio language model to author both DSP-based acoustic edits and similarity-preserving semantic edits, thus addressing the lack of publicly available audio editing data. Such data is then used to fine-tune a rectified-flow audio generation model, which we extend to expose attribute controls for loudness change, edit strength, and an input-derived contour envelope, each tunable via classifier-free guidance, yielding a more expressive editing model. FoleyForge achieves state-of-the-art instruction adherence at competitive audio quality and source faithfulness.
Analog compute-in-memory (CIM) accelerators compute matrix-vector products directly inside memory arrays, eliminating weight movement and promising order-of-magnitude gains in inference efficiency. However, converter and device errors make pretrained models unusable on these arrays, and hardware-aware retraining is costly. We enable retraining-free, accurate inference of pretrained models on simple analog arrays with low-cost, shared calibration. Our key observation is that analog error depends on how each layer is represented on the array, and that this representation can be changed offline without changing the computation. We center the inputs and apply Hadamard rotations designed for analog arrays, which keep shared reference errors spread and confine input mixing to each array. Across CNNs, vision transformers and LLMs up to 14B parameters on PCM and ReRAM, the method brings vision models and larger LLMs within a few points of floating-point accuracy and matches or approaches published hardware-aware training without training any weights. On ReRAM with 6-bit ADCs, it restores ViT-B/16 from 37.4% to 82.9% and Phi-4 from 27.0% to 72.2%, both within 2.2 points of floating-point accuracy.
Multiplayer Federated Learning (MpFL), introduced by Yoon et al. (2025), extends classical federated learning to settings where clients are rational players in a game with individual objectives and strategies, seeking an equilibrium. A natural baseline is Per-Player Local SGD (PEARL-SGD), where each player performs several local SGD steps before exchanging strategies through a central server. However, the original PEARL-SGD formulation does not extend cleanly to compressed communication. Directly compressing the players' strategies can prevent convergence to the equilibrium, even with deterministic gradients. To address this issue, we reformulate PEARL-SGD so that players communicate local strategy increments rather than the strategies themselves. Without compression, the two formulations are equivalent, but with compression, increment sharing preserves exact convergence, allowing us to design the communication-efficient compressed PEARL-SGD (C-PEARL-SGD). Beyond compression, we also consider variance reduction as another axis of improvement, introducing PEARL-SVRG and its compressed variant C-PEARL-SVRG. Under the original MpFL assumptions, we establish convergence guarantees for all variants through a unified parametric inequality. Numerical experiments, including distributed spectrum sensing, demonstrate the practical benefits of our methods over vanilla PEARL-SGD.
We introduce two fundamental advances to the NanoVDB data structure: dynamic GPU-native topological updates, and a voxel-sequential access paradigm that dramatically accelerates point-wise and stencil-based operators on sparse grids. NanoVDB's linearized memory layout has delivered strong GPU performance, and the Index-Grid concept introduced in the fVDB framework [Williams et al 2024] provided topology/data separation and optimal compactness of storage in sparse machine learning workloads. However, all these prior iterations of NanoVDB assume a static topology that has precluded use of this data structure in simulation and dynamic geometric processes. We eliminate this limitation by introducing a suite of dynamic topology operators -- morphological dilation, grid merging, pruning, and cross-resolution coarsening and refinement. These operators use bit-parallel algorithms to reconstruct grids containing hundreds of millions of voxels in just a few milliseconds. We also introduce a voxel-sequential access paradigm and data structure for GPU kernels operating on sparse data. We stream active voxels as a flat linear sequence, minimizing or eliminating the need for tree traversal in the process. This achieves compute density and regularity that allow even complex stencil-based operators to approach peak memory bandwidth, with minimal storage overhead. Together, these contributions remove the read-only constraint from NanoVDB, establishing it as a high-performance primary representation for dynamic GPU simulation and geometry manipulation. We evaluate the efficacy of our newly introduced operators in benchmarks sourced from applications such as 3D reconstruction, fluid simulation, and large-scale mesh voxelization.
Translating parallel scientific codes between HPC programming frameworks is a critical but labor-intensive task. Existing LLM-based approaches rely on fixed sequential pipelines with undifferentiated repair strategies and correctness criteria that do not match scientific computing practice. We present an agent-skill framework that separates translation, compilation diagnosis, and execution validation into three autonomous agents coordinated by a verdict-driven orchestrator, routing each failure to a targeted repair skill rather than retrying blindly. An automated output check drives the self-correction loop, complemented by an offline LLM-based verification that exposes false positives escaping compiler and runtime checks. We evaluate the framework on 30 HeCBench kernels translated bidirectionally among CUDA, OpenMP, and SYCL with three open-source LLMs. Our framework attains the higher verified correctness rate in all 18 model and direction combinations over a state-of-the-art baseline while consuming up to 52.4% fewer tokens, remaining consistent across models and directions without architectural modification.