Full Waveform Inversion (FWI) is a fundamental inverse problem in geophysics to estimate subsurface properties, such as velocity, from seismic measurements. Recently, self-supervised learning methods remove the need for ground-truth velocity maps during training by leveraging a forward operator to simulate seismic data from predictions and backpropagate the gradients of reconstruction loss through it. In this work, we consider a more challenging setting without accessing such gradients. To this end, we incorporate reinforcement learning into self-supervised FWI and propose a framework named DeepWaveRL. Specifically, we reformulate FWI as a policy learning problem, where the policy network predicts velocity maps as actions, and a forward operator is used solely for simulation. A global scalar reward is computed based on data misfit for each prediction and used to weight policy gradients. Thus, learning proceeds without accessing the gradients from the forward operator. Given the indirect and coarse supervision signals, we further identify key strategies to stabilize learning and mitigate the spatial credit assignment bottleneck. Experimental results show that, despite relying entirely on a single scalar reward, DeepWaveRL achieves competitive performance compared to supervised and self-supervised baselines. These findings suggest that our framework offers a flexible and complementary learning paradigm for the FWI research community.
Artificial Intelligence, and Machine Learning in general, is becoming pervasive across all domains and disciplines. By means of this data-driven approach, algorithms can learns complex task just by learning from samples. However, if the data-set is biased, they might lead to the development of unfair computer systems, that is, characterised by a prejudice towards specific individuals or groups, which might get an unmotivated output from the system. This circumstance lead to the development of multiple metrics designed to quantify the level of unfairness of an algorithm. However, a generalised assumption in the literature is that the sensitive features and the output labels are binary. When the data is not binary, it is forcefully binarized using a one-versus-the-rest simplification. In this work, we discuss the limitations of existing metrics and propose novel, mathematically sound, and generalized multi-group and multi-class formulations of the metrics. Ultimately, our results highlight the empirical assessment distortion caused by binarization strategies, which reveals the limitations of binary fairness metrics in the multi class setting.
We introduce the Poisson Super Learner, a scalable ensemble method for survival and competing risks prediction with right-censored data. The approach exploits the correspondence between the likelihood of piecewise-constant hazard models and the Poisson likelihood. After expanding the data over a chosen time partition, candidate hazard learners are trained under a common Poisson loss, and cross-validated log-hazard predictions are combined by a Poisson meta learner. The resulting ensemble is a geometric combination of candidate hazards, and can be used to estimate survival probabilities and cause-specific absolute risks. The Poisson Super Learner framework is modular and can in principle incorporate any learner that can estimate interval-specific hazards, including parametric, penalised, additive, and other machine-learning hazard models. Training targets the cause-specific hazard directly and does not require explicit modelling of the censoring distribution. For competing risks, separate cause-specific ensembles are fitted and subsequently combined to obtain individualised cumulative incidence predictions. We illustrate the approach using data from the Steno Diabetes Center collected between 2005 and 2022 to predict cardiovascular disease risk in patients with type 1 diabetes from detailed covariate information. The applied analysis is complemented by simulation studies designed to reflect key features of the data and to examine the behaviour of the method under controlled settings. All analyses were implemented using the freely available R software package poissonsuperlearner.
Omni-modal large language models (Omni-LLMs) typically process large numbers of multimodal tokens from audio-visual inputs, incurring substantial computational costs and a large memory footprint for the key-value (KV) cache. Aggressive token pruning can alleviate this problem while maintaining competitive performance on multiple-choice tasks, which require only a single option as output. However, we observe substantial quality degradation when generating long-form, open-ended responses. We attribute this gap to query-guided selection, which can retain sufficient evidence for answer selection while discarding details needed for comprehensive generation. This motivates reducing overlapping information across tokens, termed structural redundancy, while preserving diverse content. However, the retained representations may still contain precision redundancy due to unnecessarily high numerical precision in their cached values. To address these two levels of redundancy in token representations, we introduce a training-free framework named R$^2$-Omni. First, Structure-Aware Token Consolidation (STC) consolidates redundant representations using spatiotemporal similarity in video and local continuity in audio. Then, Semantic-Sensitivity Guided Precision Allocation (SSPA) allocates value precision to the retained representations according to the estimated reduction in attention-output error under quantized keys, with query attention and audio-visual alignment providing a semantic prior. Extensive experiments on long-form generation benchmarks demonstrate an average 5.0$\times$ reduction in the memory footprint while preserving 97.98% of baseline generation quality and achieving a 1.48$\times$ end-to-end speedup, highlighting the benefits of jointly reducing structural and precision redundancy.
Dynamic graphs capture evolving interactions in Web systems, including online communication, user engagement, and recommendation. Graph foundation models aim to reuse predictive capabilities across networks, but extending this capability to dynamic graphs faces two challenges. First, node identities are graph-specific, and similar interaction patterns can appear at different temporal and activity scales, making histories difficult to represent consistently across networks. Second, the predictive roles of historical patterns vary across interaction dynamics, limiting transfer from a narrow range of training processes. In this paper, we show that a common interaction language, combined with learning across diverse interaction processes, offers a basis for dynamic graph foundation modeling. We propose CL-DGFM, a Common Language Dynamic Graph Foundation Model for temporal link prediction. Its Dynamic Interaction Tokenizer encodes interaction memory, historical connectivity, and ordered destination transitions into calibrated tabular inputs, making dynamic histories accessible to existing predictors without an additional trainable graph encoder. Its Dynamic Interaction Prior trains a pretrained tabular foundation predictor across diverse synthetic dynamics. Labeled historical context guides prediction, with a Contextual Value Adapter modulating attention values. The model transfers to real graphs without target-graph parameter updates. Experiments on TGB and TGB-Seq demonstrate competitive performance. Further comparisons show that the common language supports basic machine learning and neural models, and enables direct use of frozen tabular foundation models on dynamic graphs.
Semantic change detection in remote sensing aims to identify and characterize land-cover transitions observed in pairs of satellite or aerial images. Existing methods assume a fixed set of change categories, limiting their applicability in real-world scenarios where changes of interest may vary, for example across geographies or over time. In this work, we tackle the problem of class-incremental semantic change detection, where a model progressively learns to detect new types of semantic change. To do so, we exploit the hierarchical structure of land-cover semantics, where newly introduced categories correspond to fine-grained refinements of existing ones, and represent said semantic categories in hyperbolic space, which is particularly well suited to represent such hierarchical relationships. To capture semantic changes, we introduce HyperFlow, a method relying on Busemann prototypes, modeling change as an anomaly in a learned semantic flow within the hyperbolic embedding space induced by segmentation. HyperFlow couples change detection to class-incremental semantic segmentation, enabling class-incremental semantic change detection. We also present HiChange, a new large-scale, high resolution remote sensing dataset with detailed, fine-grained hierarchical labels. Experimental results demonstrate that HyperFlow achieves both strong change detection and semantic segmentation performance across incremental steps. In particular, we show that progressively refining semantic categories, rather than considering them all directly, yields improved change detection, highlighting the benefits of incremental learning for semantic change detection.
Responses with headers, lists and greater length receive high scores from reward models (RMs), and RLHF policies learn to produce these features even when true quality does not improve. Practitioners therefore edit their RMs, but they learn whether an edit works only after an RL run. We show that the response of a policy to an RM edit can be forecast before training, from samples of the current policy and the scores of the old and new RM. If we read the RM as a set of prices, the first-order theory of KL-regularized RL implies that each feature moves in proportion to its within-prompt covariance with the change in reward. We implant controlled price changes into retrained RMs under a 2^(4-1) fractional factorial design with a self-distilled placebo, run 54 RL runs from six parent policies, and fix every forecast before the runs. With the covariance predictor, we predict the direction of all 20 changes in supply whose forecast has the same sign at every parent and explain 82% of the standardized change. With counterfactual prices, the classical alternative, we explain 50%. The forecast also holds under GRPO and for real RM upgrades, where we explain 66% out of sample with a coefficient frozen from the factorial runs. For rent, the gap between the proxy reward and a gold reward from stronger judges, we rank RM edits correctly with the same covariance over the first 30 steps (Spearman 0.94). This measures the horizon of the forecast: by step 200, the ranking of the factorial edits reverses under both RLOO and GRPO, while the real upgrades keep their order. When we vary the KL coefficient beta, the initial response does not depend on beta, and the response saturates below the KL equilibrium. In one experiment, the long-run response to a price rise is consistent with an additional implicit regularizer beta_0 of about 0.003 to 0.007.
Computer-use agents for scientific software require application-specific expertise. Much of this expertise is documented in human-authored software manuals and web pages. However, turning these resources into executable, reusable skills remains challenging due to mismatches between documented and deployed software versions and differences in scope between documentation coverage and task requirements. We introduce ScienceCUA, a two-stage framework combining resource-to-skill construction with execution-guided skill evolution. First, ScienceCUA identifies tasks supported by the software from its documentation, infers operational trajectories, and generalizes them into reusable Skills. Command catalogs, syntax constraints, and shortcuts are organized separately as References for on-demand lookup. Then, ScienceCUA uses execution feedback to drive skill self-evolution by contrasting successful and failed trajectories for the same task, identifying deficiencies, revising existing skills or adding new ones, and updating the skill library only after repeated validation in real environments. To enable reliable evaluation without benchmark noise or task leakage, we introduce two improved benchmarks. ScienceBoard-Verified provides an audited and corrected basis for measuring task performance, while ScienceBoard-VE separates skill evolution from final testing through disjoint evolution and test tasks and further supports native-agent evaluation. Experiments show that the initial Skills and References improve average task success rate by up to 7.23 percentage points over the base model, while skill self-evolution increases the total gain over the baseline to as much as 17.12 percentage points. ScienceCUA turns software documentation into an evolving set of executable skills and queryable references, offering a practical way for general-purpose agents to use scientific software reliably and at scale.
Multimodal test-time adaptation (TTA) enables pretrained multimodal models to accommodate distribution shifts during inference using only unlabeled test data, without recourse to the original training data. Compared with unimodal adaptation, multimodal adaptation is further complicated by the coexistence and interaction of multiple representation perspectives. This complexity is particularly pronounced under modality reliability shifts, which preserve modality availability but degrade information quality, and modality availability shifts, which render the corresponding modality entirely inaccessible. Category-conditional Gaussian modeling offers a natural starting point for handling such distributional variations, having proven effective in unimodal test-time adaptation for characterizing representation distributions, yet its extension to multimodal settings requires a more elaborate formulation. To this end, we propose Adaptive Probabilistic Gaussian Calibration (AdaPGC), which extends category-conditional Gaussian modeling across modality-specific and multimodal fusion representations for probabilistic calibration under multimodal distribution shifts. AdaPGC continuously updates category-conditional Gaussian models over modality-specific and fusion representations, using the resulting probabilistic predictions to calibrate source model predictions and guide reliability based rectification, while exploiting their joint distributions to recover unavailable fusion representations. Extensive experiments under diverse modality corruption and missingness conditions demonstrate that AdaPGC effectively improves robustness under both modality reliability and availability shifts. Our source code will be released at https://github.com/XuJinglinn/AdaPGC-plus.
Class-incremental learning (CIL) for outdoor 3D object detection remains largely underexplored. Existing outdoor studies are limited to single-benchmark evaluations, with SDCoT++ considering KITTI and CL2OD considering nuScenes under different incremental settings. This makes it difficult to assess whether observed continual-learning behavior generalizes across autonomous-driving benchmarks. To the best of our knowledge, we provide the first unified cross-benchmark analysis of outdoor class-incremental 3D object detection across multiple autonomous-driving benchmarks, covering both two-task and multi-stage settings. We investigate: (1) how severely existing 3D detectors forget previously learned classes, (2) how incremental training exposes the stability-plasticity dilemma between knowledge retention and new-class acquisition, and (3) which forms of supervision are most effective for balancing these objectives. We introduce CL3DOD, a class-incremental 3D detection framework that restores missing supervision for previous classes using predictions from the previous-stage detector, preserves old-class responses through query-wise knowledge distillation, and employs EMA-based temporal consistency regularization. Experiments on KITTI and nuScenes reveal a consistent stability-plasticity dilemma across both benchmarks. CL3DOD generally achieves a more favorable stability-plasticity balance than controlled incremental baselines and improves upon published prior methods in the reported benchmark comparisons. This advantage persists across multiple incremental stages, demonstrating strong retention of previously learned knowledge while maintaining effective new-class acquisition. Our code, evaluation setup, and incremental task splits will be publicly released.
Most image-based 3D reconstructors primarily recover visible surfaces, producing partial point clouds with large holes and leaving unobserved geometry unresolved. To address this limitation, we introduce AnchorComplete, a framework that reconstructs complete scene meshes from partial point clouds while preserving observed geometry and supporting scenes of varying scales. Building on TRELLIS.2, our method combines partial-point conditioning, which integrates global scene tokens with local spatial evidence to anchor generation to the input, and metric-scale coarse-to-fine generation, which progressively refines scene geometry within fixed-size windows while maintaining global structural consistency. Experiments on Replica and SCRREAM demonstrate that AnchorComplete consistently outperforms strong baselines in complete-scene reconstruction, particularly under sparse-view settings.
An aggregate task score does not distinguish failures of scheduling, state retention, and candidate selection. We develop a small executable instrument that varies dependency depth, candidate count, active state width, and access to current-state snapshots while holding the number of required operations fixed. Two dependency-graph families show that increasing depth at fixed work changes both scheduling freedom and the number of reachable completion states, so these controls are coupled, and we quantify the coupling. The primary study contains 49,152 generated tasks and 393,216 reference-controller rollouts. A separately frozen follow-on design adds 295,936 rollouts covering retention strategy, snapshot cadence, hidden state resampling, recovery, query locality, candidate indexing, and relabeling. Residue caches and precomputed plans solve tasks that defeat limited raw-value caches, and indexed selection removes the serial-scan branching deficit. Hidden resampling imposes an exact success ceiling that additional state tracking cannot overcome. The contribution is an executable diagnostic instrument, analytic calibration, and an audit of its assumptions. A small local CPU study with a 0.5B model completed no workflows, and a 16-episode Copilot-routed fractional factorial is descriptive only. Neither supports general capability claims.
Recourse-aware learning trains models so that people who receive an adverse decision can reverse it with an affordable change. For gradient-boosted decision trees (GBDTs), the existing method, RABIT, exposes a single recourse weight whose effect on the recourse rate is abrupt and data-dependent, and its training bound cannot certify recourse for any rejected individual. We present Certified Recourse Boosting (CERBO), which takes a requested recourse rate and returns a GBDT with certified actions: it tracks exact additive effort along tree paths, trains against committed boxes (a family of bounds that contains RABIT's and certifies a within-budget action for each training sample) and adjusts the weight after every tree until the certified rate meets the request. A calibration step guarantees the rate on new individuals, and the certifier runs on any tree ensemble, including XGBoost and LightGBM, in milliseconds per person without a solver. In RABIT's own experimental protocol, at equal accuracy, CERBO raises the recourse ratio from 0.90 to 0.99 on FICO and from 0.64 to 0.77 on Bail. Under additive costs on six data sets, it raises the area under the accuracy-recourse frontier in all 12 settings, and in simulated deployments it guarantees recourse to 70% of rejected individuals in 11 of 12 settings, against 2 for RABIT.
Understanding how stochastic gradient noise affects the convergence of Adam is an important problem in stochastic optimization, particularly when the noise exhibits heavy-tailed behavior. In this paper, we investigate the convergence of vanilla Adam in both the expected risk and the empirical risk minimization, focusing on the effects of heavy-tailed noise and its relationship to index sampling. For the expected risk minimization, we establish convergence guarantees under H\"older smoothness and bounded $\mathfrak{p}$-th moment assumptions, allowing stochastic gradient noise with infinite variance. Our analysis provides explicit upper bounds on the expected squared gradient norm and shows that the iteration-dependent convergence term decays more slowly as $\mathfrak{p}$ decreases. For the empirical risk minimization, we establish asymptotic and non-asymptotic convergence guarantees and demonstrate that heavy-tailed population noise can deteriorate convergence bounds through dataset-dependent constants as the dataset size increases. Furthermore, under Lipschitz smoothness, we formulate multistage Adam with practical learning rate schedules and demonstrate that increasing batch sizes can reduce index-sampling noise and improve convergence guarantees and SFO complexity. Numerical experiments support our theoretical findings on the effects of heavy-tailed noise and the computational benefits of increasing batch sizes.
Long-context LLM inference is costly because attention repeatedly reads a growing key-value (KV) history. We introduce Chunk-Indexed Sparse Attention (CISA), which retrofits pretrained dense attention models without gradient updates. CISA selects fixed-size KV chunks using compact summaries of existing keys, avoiding token-level indexing and a separate full-length index cache. It uses one indexing query per KV group and shares selected chunks across queries in that group. During prefill, QShare reuses selections across adjacent queries to reduce indexing cost and improve KV reuse and GPU utilization. We evaluate CISA on NVIDIA Nemotron 3 Nano and Ultra, and Inkling-Small. In one-layer BF16 benchmarks at batch size 32 using Ultra's per-device attention dimensions, core attention is 11.7x faster for prefill queries near the end of a 1M-token prompt and 13.0x faster during decode. Including chunk scoring, selection, and all other measured GPU work, speedups are 8.6x and 3.0x. On Ultra NVFP4 with a 4,096-chunk budget, aggregate scores stay within 1.5 percentage points of dense attention on MRCR, LongBench v2, and GraphWalks. RULER sweeps on Nano and Ultra BF16 up to one million tokens show scores within 2.05 percentage points of dense attention with 8-16x fewer KV reads. Serving million-token prompts with 32 concurrent requests and 4K output tokens per request, CISA achieves 2.29x the end-to-end throughput of dense attention using Nemotron 3 Ultra NVFP4 on four NVIDIA B300 GPUs. Selecting fewer chunks increases throughput to 2.61x; in separate GraphWalks evaluations, the smaller budget widens the aggregate score gap from 1.49 to 3.59 percentage points. These results expose a practical quality-speed tradeoff without sparse training.
When a language model answers a yes/no question, that answer changes how it answers a different question about the same entity. We study this \emph{self-commitment} in 22 open-weight models from four families (0.5B--72B parameters), fixing the model's first answer and measuring the shift in its second across 3,600 question pairs about 120 entities. Each answer writes a \emph{commitment direction} into the residual stream. The similarity of two questions' directions predicts their coupling on held-out entities ($\rho = 0.43$--$0.85$), and ablating a single direction shared across topics leaves at most 17\% of the effect in every model. How this direction is read is far less stable: the reading circuit spreads across more attention heads as models grow, instruction tuning makes more questions read it in the opposite sense, and in one model changing only the system prompt reverses the circuit's sign. Ablating the direction while sparing the attention sink removes most of the coupling at under 1.3\% perplexity cost in 18 of 19 instruction-tuned models, though it shifts standalone answers in two. Without the sink spared, the direction's alignment with the leading principal component of ordinary text predicts which models pay a high cost. Written once, the commitment is read differently across scale, training and prompt.
Estimating the metric 3D position of people from a monocular camera is important for robot navigation and motion prediction. Lightweight pose-based localizers are attractive for robotic hardware with limited computational resources because they operate on compact 2D keypoints. However, they typically discard the richer visual representations computed by the pose estimator. We introduce MoSoLoco, a per-frame monocular 3D human localizer that combines the pose detector's internal visual features with the detected 2D skeleton, without requiring an additional RGB or depth encoder. MoSoLoco can further incorporate sparse same-frame multi-person context to refine individual position estimates. Experiments on JRDB and NuScenes show that reusing pose-detector representations improves localization, reducing the best JRDB error by approximately 23 percent relative to the strongest evaluated keypoint-based baseline. A sparse social-interaction module incorporates nearby detections from the same image for further refinement. Controlled trajectory-forecasting experiments show downstream benefits, while the complete pipeline remains suitable for efficient Jetson AGX Xavier inference.
Large language agents for literary role-playing are typically modeled around dialogue generation, leaving characters' non-dialogue behaviors and their dynamic evolution over story time underrepresented. In this paper, we propose ACTOR (Action-Centered Temporal Organization for Literary Role-Playing Agents), an automated framework for constructing such behavior trajectories from raw novel texts. ACTOR extracts dialogue and persona-relevant non-dialogue behaviors, represents each behavior using a Thought-Action-Observation structure, and organizes behavior trajectories according to the internal temporal order of the story. We build a behavior dataset from 143 diverse literary works. For each source behavior, we generate counterfactual behaviors that are contextually plausible yet inconsistent with the character's persona. We find that existing models struggle to reliably infer character-consistent actions from historical behaviors and evolving character states. Using these behavior pairs, we train models with Regularized Preference Optimization (RPO), which augments DPO with a negative log-likelihood loss on the source behavior to jointly optimize behavior preference and generation. We further design a four-way next-behavior selection task to evaluate the model's ability to faithfully mimic character actions. Across 17,560 evaluation instances, the model trained on ACTOR's preference pairs achieves substantially improved accuracy. This result demonstrates that structured, temporally organized supervision enables models to leverage character profiles and prior behaviors to infer subsequent actions, greatly advancing literary role-playing beyond mere dialogue imitation. Code and data are available at https://github.com/SilentStar233/ACTOR.
Rapid advances in cowork agents are shifting human-AI collaboration toward real-world automation, delivering software, designs, and other artifacts. However, such productivity work demands both high-quality delivery on complex productivity tasks and continued monitoring and decision-making over long horizons, a combination that remains an open challenge. We introduce CoworkOS, a general-purpose harness for long-horizon work that separates task strategy from execution management: the model decides actions, delegation, and acceptance, while the runtime maintains state and feedback. Its three-layer architecture enables direct execution or delegation to Subagents, synchronously or asynchronously, under stable responsibilities. Through a shared application layer, humans and agents inspect and revise the same evolving documents, presentations, CAD models, and Blender scenes, so human-in-the-loop participation can enter both through conversation and through direct modification of the work itself. To sustain continuity across agents and sessions, CoworkOS further introduces hierarchical workspaces and persistent semantic memory, so working files, key decisions, and long-running Job logs remain shareable, reusable, and resumable after interruption. Across 12 benchmarks and 6 categories of real-world applications, CoworkOS proves highly competitive with frontier agent applications such as Codex and DeepSeek Harness. As a sustained-work case study, CoworkOS uses GLM-5.3 as its backbone to conduct RSI-driven automated post-training on Qwen3.6-35B-A3B. Over five rounds, the model's average score across eight benchmarks rose from 33.19 to 42.95 (+9.76). Looking ahead, we believe CoworkOS opens a meaningful path toward general cowork-agent intelligence.
On July 1, 2025, Vietnam reduced its 63 provinces to 34, abolished its 696 districts, and replaced its 10,035 commune-level units (wards) with 3,321. ViAdminShift, a Vietnamese language resource about this event, consists of 7,995 benchmark items, 2,281 retrieval passages from pre- and post-reform Wikipedia revisions and dated templates, and old-to-new tables that a released script rebuilds from the official conversion table (not redistributed). Changed items store the pre-reform answer (except 121 address items), a $2\times2$ design uses untouched units as controls, and 400 addresses appear without diacritics. The tables agree with an independent lineage on 10,564 of 10,565 official pairs, disagreements were settled against the resolutions, and the probing-track scorer agrees with both authors' blind labels (Cohen's $\kappa=0.91$ and $0.89$). Data, outputs, code, and a datasheet are released. In the probing track, at least 20 of 67 models know the reform. Fourteen of them give the new province alone for at most 11% of plain questions about old addresses (old province only: 14-100%); a bare date recovers up to 100%. Post hoc (80 province-free addresses, single runs), with reasoning off, today's date in the system prompt leaves 68-100% old-province-only answers for 10 of 12 models; with reasoning, an "as of" prefix in the question leaves none for seven of eight. On the 1,262-item benchmark subset, with same-source old and new passages, small models unaware of the reform give 14-30% outdated answers and the five benchmark-track knowers (also differing in size or provider) 0-5%.