Speculative decoding accelerates rollout generation, which dominates the cost of reinforcement learning (RL) post-training. Online co-training keeps the draft aligned with the evolving policy, further increasing acceptance length and speedup. However, training large models on long contexts requires context parallelism (CP) and pipeline parallelism (PP), and co-training drafts under this layout poses two obstacles: (1) the draft's branch attention is unsupported by causal CP, and (2) the target features that the draft consumes are spread across PP stages. In this paper, we present the design and implementation of online draft co-training in NeMo-RL. For CP, we extend packed, load-balanced zigzag ring attention by merging rank-local branch attention with causal main-sequence attention. For PP, TapChannel transports intermediate target features across stages via a separate path, leaving the pipeline schedule unaffected. Experiments show that co-training keeps the policy's reward and accuracy close to the no-draft baseline while delivering $1.48$--$2.25\times$ rollout speedups and $1.19$--$1.88\times$ end-to-end training speedups across model scales up to 122B. Our CP design achieves strong scaling at 256K sequence length with significant memory savings over prior work, and our PP transport incurs modest overhead. Code can be found at https://github.com/NVIDIA-NeMo/RL/issues/3698.
A query-oblivious coreset for a softmax-attention head is a subset of the key--value pairs whose attention output is within $\varepsilon$ of the full one for every query in a ball. With $\rho$ the query radius times the centred key radius, we build such coresets of size $O(e^{\rho}(\sqrt{d_v}+\sqrt{d_k\log(1+\rho)})/\varepsilon)$ in randomised polynomial time, improving the existential $O(\sqrt d\,e^{\rho+\frac12\log\rho+o(\log\log\rho)}/\varepsilon)$ of Liberty, Andoni and Kleiner to within $\sqrt{\log(1+\rho)}$ of their $\Omega(\sqrt d\,e^{\rho}/\varepsilon)$ lower bound: a spherical lift of both balls into one exponential-kernel instance admits the Bozzai--Rothvoss chaining bound, and Chevet's inequality splits key from value dimension. A sampling cap $O(e^{2\rho}/\varepsilon^{2})$ completes the envelope, and in fixed dimension Tai's diameter-free bound removes the logarithm. A census of every head of Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct finds $\rho\ge24.100$, so every whole-ball bound with a factor $e^{\rho}/\varepsilon$ prescribes a coreset larger than the cache: the algorithmic contribution is asymptotic on these models.
Scalable video coding (SVC) encodes a video into a layered bitstream of a base layer and one or more enhancement layers, enabling decoding at different bitrate/quality/resolution operating points for diverse devices and network conditions. SVC has been incorporated into major video coding standards and has recently attracted growing interest for both scene-agnostic and scene-adaptive neural video codecs. Among the latter, implicit neural representation (INR) based codecs overfit a compact neural network to an individual video, offering fast decoding and competitive coding efficiency. However, research on scalable INR-based compression remains in its infancy: existing methods introduce additional network layers, which couple the bitrate with the decoding complexity and cannot match strong scalable/non-scalable codecs. This paper proposes S-NVRC, a scalable INR-based video codec that jointly supports fine-grained bitrate and decoding complexity scalability from a single embedded bitstream. It adopts a coarse-to-fine prefix for feature grids and a nested prefix for network layers, which scale bitrate and decoding complexity, respectively. With a single encoding, S-NVRC spans a wide range of bitrate and decoding complexity and outperforms SHM-12.4 and the multi-layer VTM-20.0 by 43.7% and 5.6% in BD-rate on the UVG dataset.
Large language models (LLMs) are increasingly used to assess depression in counseling conversations, where psychology evidence suggests that tracking change across sessions matters more than any single score. However, existing dialogue benchmarks score each session on its own, and an accurate score for every session does not guarantee that a method follows the changes between sessions. We introduce LongCounsel-8, a benchmark for tracking depression across counseling sessions. Its systematic pipeline takes client profiles from real counseling cases, depression scores from real longitudinal data, and turn length and form from real counseling dialogues, yielding 7,749 five-session trajectories from three LLMs. Every session is generated to express a preset score on the PHQ-8, a standard eight-symptom depression questionnaire. Validation shows that the labels are self-consistent and the dialogues close to real counseling. Following clinical research on treatment monitoring, we propose a trend-based evaluation protocol reflecting real-world tracking needs: changes between sessions, crossings of the screening cutoff, trajectory types, and symptom-level changes. Extensive experiments show all five existing methods tested fall short in tracking trends: none estimates changes better than a no-change baseline, at most 40% of cutoff crossings are detected, and adding earlier sessions lowers change error but causes more missed improvements.
This work studies hybrid quantum-classical neural networks for learning routing heuristics. Specifically, this paper asks whether small quantum neural networks can replace parameter-heavy modules inside a competitive attention-based routing model while maintaining solution quality. For the capacitated vehicle routing problem, encoder feed-forward replacement emerges as the most promising design: it reduces the number of model parameters by 56.6% while keeping the hybrid model close to the classical neural baseline at small and medium instance sizes, although the gap grows for larger instances. This work also compares to classical routing algorithms, which remain highly competitive and often superior on the fixed Euclidean test sets. Our results therefore do not indicate quantum advantage or solver dominance, but identify encoder feed-forward replacement as a viable hybrid-module compression strategy for neural combinatorial optimization.
Two of the most fundamental questions in statistical learning theory are the following: which prediction problems are learnable, and how should they be learned? For the former, elegant answers often take the form of combinatorial dimensions. The latter question, however, has proved considerably more elusive: all known general-purpose multiclass learners rely on intricate orientations of exponentially large one-inclusion structures, and familiar algorithmic principles such as proper learning and regularization remain poorly understood. Motivated by prior work, we ask whether learning reduces to proper learning---possibly over a larger hypothesis class---and whether proper or improper multiclass learning can ultimately be captured by suitable regularizers. Our primary results answer both questions negatively, resolving three open problems from prior work. First, we exhibit a learnable multiclass problem that cannot be embedded in any properly learnable class, meaning learning cannot be reduced to proper learning by enlarging the hypothesis class. Second, we demonstrate that proper learning can require training error and characterize this phenomenon precisely: every properly learnable class admits a proper learner making $o(m)$ errors on samples of size $m$, but every prescribed sublinear scale $a_m=o(m)$ is necessary for some properly learnable problem. Third, regularization is not a general learner: we exhibit a properly learnable class that cannot be learned by any Structural Risk Minimization (SRM) learner, and a learnable class that cannot be learned by any local regularizer. We complement these impossibility results with a positive theory that gives two sufficient conditions for SRM learnability and characterizes SRM representability through integrability of revealed preferences.
Facial attractiveness has been linked to statistical regularities such as symmetry and averageness, suggesting that beauty may depend on the ease with which a face is perceived. We empirically test this hypothesis by training variational autoencoders on four face datasets without attractiveness supervision and evaluating their representations on the 597 faces from the Chicago Face Database. We report three main novel findings. First, we identify that human attractiveness ratings closely aligns with the direction defined by the VAE evidence lower bound (ELBO) in rate-distortion space. Second, we find that independently learned latent spaces contain an attractiveness direction that transfers strongly across random initializations and training data. Finally, we observe that attractive faces are more prototypical in both shape and latent space. Our work provides the first explicit connection between classic accounts of aesthetics with learned generative models and empirical support for a variational interpretation of the processing fluency theory of aesthetic pleasure.
Despite remarkable progress in instruction-based video editing, textual descriptions remain limited in expressing fine-grained appearance, spatial transformations, and temporal dynamics. We introduce Visual In-context Editing, a new paradigm that shifts video editing from text-only instructions to visual demonstrations, where users specify desired edits through heterogeneous visual references including single images, image pairs, and video pairs. To establish this paradigm, we curate VicEdit-400K, the first large-scale dataset for visual in-context video editing, containing 400K high-quality samples across ten editing tasks with comprehensive semantic and visual consistency filtering. Building upon this foundation, we propose VicEdit, a unified framework that enables effective reasoning over diverse visual contexts and textual instructions. Specifically, we introduce Modality-Adaptive Semantic Distillation to extract modality-aware visual editing semantics and Dual-Context Injection to align visual contexts with textual editing intents. Extensive experiments on VicEdit-Bench demonstrate that VicEdit achieves state-of-the-art performance on both conventional instruction-based editing and visual in-context editing. Further controlled studies and human evaluations verify that VicEdit effectively leverages heterogeneous visual references, establishing visual in-context editing as a promising paradigm for controllable video editing.
Memory is becoming core infrastructure for long-horizon LLM agents, yet existing evaluations offer limited guidance on which memory substrate, namely the underlying medium in which memory is represented and stored, should be used under different operating regimes. We introduce a unified suite of 26 performance and efficiency metrics, including auxiliary-LLM costs attributed to the write, read, and management stages, and apply it identically to 11 memory methods in a controlled harness that varies only the memory substrate. The methods cover dense and sparse indices, text records, structural stores, hierarchical stores, refinement-based memories, parametric updates, and activation-compatible context mechanisms, and we evaluate them with three backbone models on four benchmark suites spanning user-centric question answering and agent-centric decision-making. Our results show that no single substrate consistently dominates: broad retrieval benefits long-context factual QA, while excessive retrieval can harm sequential decision-making by shifting attention away from action-critical context. Scalability introduces a further routing axis, as substrates that perform well at moderate history lengths can become costly or brittle at longer horizons. These regime-dependent reversals make substrate selection a design axis. Choosing a substrate per observable operating regime improves over the best fixed substrate on our benchmarks, while query-level routing leaves most of the oracle headroom open. The findings provide empirical guidance for designing efficient, reliable, and regime-aware long-term memory for LLM agents.
Perception from a single unmanned surface vehicle (USV) is limited by sensing range, field of view, and occlusions caused by islands, port infrastructure, and nearby vessels. Multi-USV cooperative perception can alleviate these limitations by integrating complementary observations from distributed viewpoints. However, progress is constrained by the high cost of coordinated sea trials, the scarcity of multi-USV datasets, and the lack of integrated maritime simulation and evaluation tools. We present MMUSV-Sim, a perception-oriented maritime simulation and data-generation platform built on Unreal Engine 5 and Project AirSim. It supports configurable maritime scenarios, synchronized multimodal acquisition across multiple USVs, and consistent 3D annotations in a shared world frame. Using MMUSV-Sim, we construct MarU2U-200m, a synchronized multi-USV RGB--LiDAR dataset with 341 sequences and 14,385 scene frames. We establish a cooperative bird's-eye-view (BEV) vessel detection benchmark with LiDAR and RGB--LiDAR inputs. Experimental results show that Early Fusion improves LiDAR detection AP@0.5 from 64.26 to 81.46 compared with the single-USV baseline. RGB features further improve intermediate fusion, increasing AP@0.5 by up to 4.55 percentage points. LiDAR and RGB transfer experiments also show the value of the generated data as a pretraining source for real maritime detection. The project are available at https://github.com/lizaowo/MMUSV-Sim
As large language models (LLMs) are increasingly applied for emotional companionship, evaluating their behavior and capabilities in intimate relationships has become a pressing issue. However, existing assessments primarily characterize general personality traits, providing limited insight into model behavior within intimate and emotionally sensitive contexts. Therefore, we introduce adult attachment theory into LLM evaluation and use the Experiences in Close Relationships-Revised (ECR-R) scale to characterize attachment anxiety and avoidance. To evaluate emotional companionship capabilities of LLMs in realistic interaction scenarios, we present an emotional companionship benchmark, ECBench, spanning four scenarios including emotional support, collaborative tasks, conflict resolution, and social guidance, across friendship and romantic relationships. ECBench is utilized to assess model behavior using 11 dialogue-quality metrics and three evaluation methods. We evaluate the attachment tendencies of 32 LLMs and select representative models to investigate how these tendencies manifest in contextualized multi-turn interactions and whether they can be shaped through prompting. Our study provides a theoretical lens from psychology, along with practical tools to understand and select LLMs for emotional companionship.
Generative models can produce images nearly indistinguishable from real data, yet rigorous and interpretable evaluation remains challenging. Conventional metrics such as FID provide only scalar scores with limited diagnostic insight. Widely adopted CLIP-based metrics enable semantic evaluation beyond simple training class labels, but inherit limitations from CLIP's training paradigm that restrict attribute-wise analysis. We propose RA-CLIPScore, a novel metric that mitigates these issues and extends CLIP-based evaluation to spatial distribution alignment, measuring whether generated objects adhere to the positional priors found in the training data. RA-CLIPScore introduces dual prompts to decouple competing attributes and leverages local patch tokens to capture fine-grained regional semantics. We evaluate image generative models on their ability to match both attribute and spatial distributions of the training data. Extensive experiments show that RA-CLIPScore provides more robust and interpretable evaluations than prior methods, particularly under distribution misalignment or partially irrelevant textual attributes. We further demonstrate how it reveals spatial biases in generative models. User evaluations confirm that Regional Single Attribute Divergence based on our RA-CLIPScore aligns more closely with human perception of visual diversity than existing semantic metrics.
Modern systems are increasingly expected to transfer across tasks not specified during training. What data facilitates generalization to unanticipated settings? One hypothesis is that data with more structural information could contain shared circuits and subprograms that may be recycled in a wider array of downstream settings. Epiplexity, a recently proposed measure of the structural information a compute-bounded learner can extract from data, offers a mechanism to reason about this relationship. As existing estimators for epiplexity only operate on data already seen during training, we show how to operationalize epiplexity as an online training signal by greedily maximizing its stepwise gain in two settings. For data selection, we fit scaling laws to the training loss curves of natural data domains to predict the expected epiplexity gain as a function of training tokens, and use this signal to adaptively determine the sampling weights over domains during training. For synthetic data generation, where no domain structure exists, we reward a generator with the change in epiplexity of a learner's synthetic training data and use REINFORCE policy gradients to guide the generator toward an epiplexity-maximizing distribution. In both cases, we find higher epiplexity predicts improved downstream performance on zero-shot and fine-tuning tasks, supporting the hypothesis that data rich in structural information yield representations that better transfer across domains.
Human-artificial intelligence (AI) collaboration takes many forms and occur in both professional and educational scenarios. It is important to identify how human users interact with AI and how their behavior affects task success and learning outcomes. Existing approaches often rely on predefined behavioral categories or derive patterns without establishing their relationship to outcomes. In this work, we propose Principal Trait Analysis (PTA), a data-driven method for identifying interpretable behavioral traits from human-AI collaboration dialogues. Our method analyzes large corpora of human-AI collaborative session traces, derive common traits across the dataset, scores each human collaborator on how well their interaction aligns with these traits, and selects a set that covers the most variance across users. While unsupervised and data-driven, our method also enables domain expertise to be injected during trait identification. We evaluate PTA on two human-AI collaborative coding datasets and investigate how the traits predict future task success, learning outcomes, and evolve over time.
Wavelet convolution (WTConv) has emerged as an increasingly popular drop-in replacement for standard convolutions, expanding a network's receptive field exponentially with the number of decomposition levels while keeping the parameter count linear. However, its reference implementation is severely memory-bound due to excessive data movement through high-bandwidth memory (HBM). We develop an I/O model of WTConv to characterize this bottleneck and use it to guide three algebraic reformulations: (1) recomputing the inexpensive Haar analysis butterfly on chip, (2) collapsing the multi-level synthesis cascade into a single closed-form pass indexed by output-coordinate bits, and (3) folding learned per-channel scales into the convolution weights. Together, these reformulations enable an I/O-aware fused implementation that substantially reduces HBM traffic. We evaluate the WTConvNeXt configuration across decomposition levels and a broad range of tensor shapes. Despite performing comparable arithmetic, the reference WTConv is substantially slower than the depthwise convolution it replaces. Our reformulation reduces modeled HBM traffic by approximately $2.55\times$, yielding up to a $4.35\times$ training speedup over the reference while roughly halving peak memory usage. Thus, our reformulation preserves the benefits of WTConv while substantially reducing its execution time and memory footprint, removing the systems overhead that previously limited its practical efficiency.
As LLM-generated content becomes more sophisticated, detection systems for distinguishing those texts from human-written text must operate at scale while handling diverse writing styles, domains, languages, and generator models. Input-only encoder detectors are suitable for practical deployment setting, but standard binary classification supplies only the class label and does not explicitly organize the substantial variation within either class. We propose MD-ProTector, which represents each class with multiple trainable reference vectors in the encoder embedding space, referred to as prototypes. These prototypes provide separate decision boundaries for different groups of texts within the same class. However, adding multiple prototypes alone does not determine which variation each prototype should represent. MD-ProTector addresses this problem with Prototype Positioning loss, which separates class-level structure from the within-class variation that differentiates individual prototypes. Evaluated across five settings from three large-scale benchmarks covering domain, generator, language, and adversarial variation, MD-ProTector achieves the highest AvgRec on MAGE CDCM and RAID and the highest AUROC and lowest FPR95 on RAID among the compared encoder-based methods.
Tabular Foundation Models (TFMs) are the state of the art for many tabular prediction tasks. They are transformers that approximate the Bayesian posterior predictive distribution under a pretraining prior. TFMs become multivariate predictors autoregressively, by sampling one target and appending it to the features. This paper is the first to evaluate whether this setup induces a valid stochastic process. Informally, we ask: *Do we get the same answer to the same question asked in two different ways?* The standard Kolmogorov consistency conditions cannot be checked on TFMs directly, since TFMs are unable to model certain conditional distributions. We therefore propose two adapted, testable requirements. Marginalization consistency demands that marginalized conditionals equal directly predicted marginals, and factorization consistency that both factorization orders yield the same joint. Since the data-generating distribution satisfies both, a perfect TFM would as well, and every violation lower-bounds how far the model's predictions are from the ground truth. Every TFM we evaluate violates both requirements, for classification and regression, on all datasets.
An LLM agent can acquire persistent improvements by evolving its harness---the executable program governing model calls, tool use, verification, and recovery---without updating model weights. When the same harness serves users with different workloads, isolated evolution limits the sharing of useful discoveries, while centralized adaptation can introduce conflicting task-specific changes. We introduce EvolveNet, a framework for collaborative harness evolution that connects local specialization with shared capability accumulation. Clients evolve the same broadcast harness on their local workloads in parallel, and a server composes their adaptations into an updated shared program. Redistribution makes accepted discoveries available to all clients, allowing subsequent local evolution to build on improvements accumulated across the collaboration. EvolveNet supports this process through evidence-guided, scope-typed program aggregation and behavioral validation. Our theoretical analysis provides sufficient conditions for retaining and combining complementary improvements despite program interactions. Experiments across five settings show improved shared-harness performance, with ablation and behavioral analyses supporting the contribution of composition beyond specialist selection alone.
Accurate detection of sedentary behavior is important for studying health risks related to prolonged sitting, but posture-based classification remains challenging with wearable sensors, especially at the wrist. We study whether a deep learning model trained on hip-worn accelerometer data can transfer to wrist-worn accelerometer data for sitting versus non-sitting classification. We use CHAP, a CNN-BiLSTM model originally developed for hip accelerometers, and evaluate its zero-shot performance on wrist data as well as its adaptation through finetuning with varying amounts of labeled wrist data. Experiments are conducted on the iWatch dataset with ground-truth posture labels derived from wearable cameras. The hip-trained model performs strongly on hip data without retraining, but accuracy drops on wrist data due to sensor placement shift. Finetuning CHAP provides consistent advantages over transformer models trained from scratch. These findings suggest that hip-based pretraining provides a useful starting point for wrist deployment, while highlighting the need for wrist-specific adaptation to handle higher signal variability.
Language remains an outlier in generative modeling: while images, video, and audio are increasingly modeled in continuous latent spaces, text generation still relies predominantly on discrete tokens. Existing continuous language models either inherit embedding spaces not designed for joint generation and decoding, or compress autoencoded latents to ease diffusion, sacrificing token-level fidelity. Instead of simplifying the representation to suit the generative model, we preserve a high-capacity, decodable text latent and design the diffusion model to learn its distribution directly. We introduce AURORA-LM, a continuous-latent diffusion language model that separates the construction of a decodable text representation from the modeling of its distribution. A Query-based Encoder-Decoder organizes text into a high-capacity, prefix-aligned latent sequence, and a Block-causal Diffusion Transformer learns its distribution through flow matching, generating blocks left to right while denoising positions within each block in parallel. Because such a latent is harder for diffusion to model, AURORA-LM restricts only the noisy-input pathway while retaining the full clean-latent prediction target, accommodating full-width latents without reducing decoder-facing capacity. We further calibrate the noise-level distribution to the latent width, and introduce self-trajectory consistency to bridge independently sampled training noise and iterative denoising at inference. AURORA-LM achieves the strongest performance among evaluated continuous and diffusion-based language models on OpenWebText free generation and XSum summarization. Scaling to 1B parameters with about 1500 EFLOPs of total compute yields further gains, surpassing a larger publicly released latent-diffusion language model under a matched evaluation protocol. All experiments are conducted on Ascend NPUs.