A model that takes over another model's conversation must normally re-read the whole history; a text hand-off shortens the re-read only by dropping content. KV-cache translation instead maps the source's cache into the target's own format, so the target continues from the full context without re-reading it. It works for grouped-query attention (GQA), but open-weights frontier models now cache compressed latents (MLA), sparse-attention indexer keys and fixed-size linear-attention states. We replicate KV-Lingo on a GQA pair and then give, for each of these objects, where to capture and inject it and a mechanism that reaches native quality. We study two hybrid pairs at 16-48B scale in both directions, Kimi-VL-A3B - Kimi-Linear-48B (MLA - MLA + Kimi Delta Attention) and Qwen3.5-27B - Qwen3.8-27B (Gated DeltaNet hybrids), plus a sparse-attention target (Qwen3.8-27B - Qwen3.8-Flash-Next). MLA latents translate with token-wise linear maps; recurrent states are copied with a closed-form correction or synthesised by replaying the target's own recurrence. Exact long-range recall needs token identity: decoding token ids from the source's layer-0 MLA latent raises needle recall from 0.09 to 0.99 at 2-32K and to 0.83-1.00 at 64K, eight times the longest training prompt. The hybrid full-MLA direction is within the paired confidence interval of native quality on every task metric, and same-shape GDN hybrids reach the native level in both directions (apart from 6% refusals inherited from the source), with time to first token 18.7 and 46.8 times shorter than a re-read at 64K tokens. Segment-wise multi-turn transfer stays at the re-prefill level; naive re-translation collapses. On the sparse-attention target, predicted indexer keys recall the needle as well as the target's own keys (one seed). We state as a hypothesis how the recipes transfer to 300B+ hybrid flagships.
We present BARS (Beat-Adaptive Rap Synthesis), a non-autoregressive rapping voice generation model based on Flow Matching, conditioned on a beat track (instrumental backing), lyrics, and a short reference voice for zero-shot timbre control. Unlike prior autoregressive rap synthesis, where rap pace and output duration are strictly determined by the beat, BARS enables independent control over the rap tempo, allowing flexible-length generation without explicit duration prediction or phoneme-level alignment. We encode lyrics using a pretrained ByT5 encoder with a lightweight Transformer adapter trained jointly with the model, without relying on speech-pretrained text encoders, enabling natural handling of non-standard orthography (e.g., slang, abbreviations) and numerals common in rap lyrics. To address alignment instabilities, such as token skipping or repetition, under strong beat conditioning, we propose Asymmetric Scaled RoPE for cross-attention, which introduces a positional structure to facilitate sequential lyric-audio alignment learning. Furthermore, the model supports partial re-generation via span masking, enabling lyric-level editing while preserving the surrounding vocal style. Experiments demonstrate that the proposed model achieves high lyric fidelity, natural rap flow that is rhythmically coherent with the beat, and robust tempo control across a wide range of rapping styles and speeds.
Large language models offer a promising technique for translating natural-language optimization problems into solver code, but their reliability remains constrained by the quality of template retrieval. While recent structured methods improve over flat prompting through hierarchical taxonomy search, their traversal rules remain fixed or LLM-driven at inference time, leaving retrieval unable to improve from past successes and failures. This motivates treating template retrieval as a policy that can be learned from modeling experience, rather than as a static inference-time rule. We propose COMPASS, a policy-amortized framework for retrieving optimization templates in LLM-based solver modeling, which represents the template library as a directed acyclic graph and formulates retrieval as a sequential decision process over template nodes. A deep Q-network learns a reusable traversal policy from template-level judgments and execution feedback, reducing reliance on fixed heuristic traversal through feedback-driven graph navigation. To support extensible retrieval, COMPASS embeds problem descriptions and template nodes in a shared semantic space, allowing newly inserted templates to be scored from their text descriptions without redefining the action space. Experiments across multiple LLM backbones and optimization-modeling benchmarks demonstrate that COMPASS consistently improves over heuristic structured retrieval, with the largest gains on complex and mixed-category problems.
Online Conformal Prediction (OCP) sequentially constructs prediction sets whose average empirical coverage converges to a target level over time, for arbitrary, potentially adversarial datasets. In the multi-level setting, prediction sets are produced at $d$ coverage levels at once. Prior work argued for the use of online gradient descent with isotonic projections, which produces sets nested across levels, but with a $\sqrt{d}$ dependence in the per-level coverage convergence rate. Here, we both generalize and improve prior work. First, we show that isotonic projections can be used for the construction of multi-level online conformal prediction sets from \emph{any} online learning algorithm (not just gradient descent). Then, through a black-box reduction, we show that the linearized regret of the online algorithm itself automatically gives control over both per-level coverage and distributional consistency of multi-level predictions via Fenchel duality. Finally, by using the Weighted Interval Score, a standard scoring rule for assessing multi-level interval forecasts, we prove that these projections reduce both the size of the sets and their under-coverage. We study two instantiations of our theory, $p$-norm Online Mirror Descent and coordinate-wise Universal Portfolio methods, both improving the dependence on the number of levels to $\sqrt{\log d}$ for the multi-level OCP problem. Experiments on several datasets show favorable performance compared to baselines.
The Natarajan dimension is a fundamental measure of the capacity of multiclass hypothesis classes and plays an important role in analyzing their generalization ability. In this paper, by establishing lower and upper bounds, we show that the exact Natarajan dimension of homogeneous multiclass linear predictors with pairwise distinct class weights is \(n(k-1)-1\), where $n$ is the dimension of the feature space and $k$ is the number of classes, for \(n,k\ge3\), \(n,k\in\mathbb N\).
Multi-objective evolutionary optimization increasingly involves large populations, expensive problem evaluations, and diverse algorithmic structures. These workloads expose the limited parallelism of loop-based implementations, while rewrites for accelerator execution can alter selection order, tie handling, and population updates. AscendMOEA is presented as a scalable platform that executes multi-objective evolutionary algorithms (MOEAs) on Huawei Ascend neural processing units (NPUs) and conventional central processing units (CPUs). The platform represents the decision vectors, objectives, constraints, and auxiliary states as tensors and provides reusable interfaces for algorithms, problems, population operators, quality indicators, monitoring, and repeated experiments. Dense computations are executed in batches, such as dominance, distance, variation, and problem-evaluation calculations. Short recurrences with causal dependencies retain their prescribed order and consume precomputed tensor results, which preserves the behavior of the original algorithms. AscendMOEA contains 18 algorithms, 155 benchmark and application problems, 19 public operator interfaces, and 11 quality indicators. Experiments on eight Ascend NPUs and 192 physical CPU cores evaluate operator throughput, end-to-end runtime, memory consumption, numerical agreement, and implementation ablation. Tensorized operators obtain a median acceleration of 1.58 times on the NPU. At large population scales, complete algorithms achieve a median acceleration of 5.76 times while retaining the recorded quality indicators. Moreover, the results identify the workload scale and ordered computation that determine whether accelerator execution is beneficial. AscendMOEA is open-sourced at https://github.com/dqlme/AscendMOEA.
Training Gaussian processes (GPs) at large scale is computationally challenging because evaluating the log-marginal likelihood and its gradients involves large, dense kernel matrices. We develop a scalable low-rank GP training approach that combines AnchorNet landmark selection with a fully independent training conditional (FITC) correction. The resulting objective and its analytic gradients can be evaluated using low-rank factors, diagonal quantities, and small dense matrices, without forming the full kernel matrix or its derivatives. The complete training procedure is implemented in single precision and evaluated on nine large-scale regression datasets. Across these problems, the method reaches the reference predictive accuracy while substantially reducing training time, although the required approximation rank varies across datasets. We also study the numerical behavior of the full-kernel predictive solve and show that single precision can become limiting for ill-conditioned problems. Finally, we investigate further reducing selected matrix operations to fp16 or TF32. These lower-precision variants provide only modest additional speedups and can become unreliable when the reduced system is poorly conditioned.
Constrained preference-conditioned MORL typically assumes full task observability, treating preferences and safety budgets as stationary task signals. In real-world deployment, however, these signals are often unreliable proxies: intent may drift, safety tolerances may shift with context, and exposed descriptors may be noisy, delayed, or incomplete. We therefore formulate constrained preference-conditioned MORL as a partially observed task-specification problem, where the agent must navigate a fundamental conflict between unreliable external signals and latent interaction dynamics. We propose \textbf{BECON}, a belief-conditioned framework for constrained MORL with drifting preferences and budgets, which infers latent task context from recent histories to estimate preferences and safety budgets. These estimates are fused with the exposed task signal via uncertainty-dependent gates, enabling the policy to defer to observations under high epistemic uncertainty while correcting them when history is informative. The fused context jointly conditions preference optimization and context-adaptive constraint enforcement. Experiments on diverse multi-objective control tasks demonstrate that BECON consistently improves over state-of-the-art MORL baselines, achieving stronger robustness and constraint satisfaction under task drift.
Cross-view image sequence geo-localization aims to determine the geographic location of a sequence of ground images by matching it against geo-tagged satellite imagery. Although sequential observations provide complementary scene information, substantial overlap between adjacent images introduces feature redundancy, while conventional fusion strategies may weaken spatial relationships essential for cross-view matching. Therefore, we propose a cross-view image sequence geo-localization method via Attention-guided Sequence fusion and Metric Learning (ASML). Specifically, a Sequential Content Fusion (SCF) module models cross-image feature correlations to suppress redundant content and adaptively preserve geographically informative non-overlapping regions. Next, a Spatial Feature Aggregation (SFA) module is used to capture and aggregate the spatial and semantic information of local features, so as to maintain the spatial layout of the image sequence. Finally, a Geo-InfoNCE loss function is proposed, which comprehensively considers the feature distance and geographic distance for mixed-space metric learning, thus improving the ability of the model to distinguish hard negative samples. A new cross-view image sequence dataset CrossRoad-66K is established to provide more complex scenarios. Extensive experiments and visualizations on the SeqGeo and CrossRoad-66K datasets demonstrate that compared to state-of-the-art (SOTA) methods, the proposed method ASML achieves an improvement of 3.37% and 2.10% at R@1, respectively.
Floorplanning is an early step in turning a logical circuit into a physical chip: it positions large circuit modules and constrains the space available for subsequent logic placement and wiring. These decisions affect congestion and timing, but their physical consequences emerge through downstream implementation. For AI agents, the task requires translating engineering constraints into spatial decisions whose quality cannot be judged by successful tool execution alone. We present an evaluation framework that connects submitted layouts to a prescribed downstream implementation flow, checks geometric validity, measures physical quality, and reports resource use. On one industrial design whose 99 macros occupy approximately 71\% of a fixed L-shaped layout region, we compare our specialized system Nova-SI with general-purpose engineering agents and agents restricted to file access, obtaining complete physical measurements for 39 layout outputs. Under our 0--100 composite quality score, Nova-SI achieves a mean of 57.97 without online feedback from the prescribed power-grid and placement evaluation, compared with 31.17 for the strongest measured general-purpose configuration, GPT-6 Astra. The difference is supported by raw congestion and timing measurements and persists across scoring-sensitivity checks. Layout-level analysis further shows that similar global utilization can conceal substantially different local congestion. Together, the protocol and case study assess agents through the physical consequences of their delivered layouts and provide evidence about system performance on this industrial instance.
Background: Diagnostic I-131 whole-body planar imaging in differentiated thyroid cancer (DTC) is count-limited by low administered activity, and quality worsens under rapid acquisition. Models trained on such data struggle because their targets lack photon statistics. Purpose: To test whether high-quality post-therapeutic (PT) I-131 scans can serve as superior training references for enhancing sparse diagnostic planar images, using transfer learning and input-image addition. Methods: In this single-center prospective study, 111 patients were enrolled in PT training (n = 19), diagnostic follow-up (FU) training (n = 58), and FU paired test (n = 34) cohorts. Five Restormer-based models were trained: FU_DLx2 on FU labels; PT_DLx2 transfer-learned on PT data; PT_DLx1+ and PT_DLx2+ with input-image addition; PT_DLx6 as a PT-only reference. Test-time processing used two-step denormalization and Poisson re-scaling to preserve noise statistics. Quality was assessed by PSNR, SSIM, and a blinded dual-reader evaluation of all 34 test-cohort patients. Results: PT_DLx2 achieved the highest PSNR (34.4 +/- 3.6 dB) and SSIM (0.81 +/- 0.07). PT_DLx1+ and PT_DLx2+ were comparable (33.8 and 33.3 dB; SSIM 0.80). FU_DLx2 ranked below the unprocessed fast scan (30.3 vs. 32.0 dB; 0.76 vs. 0.79). Subjectively, PT_DLx2+ scored highest in image quality (4.54 +/- 0.60) and diagnostic confidence (4.54 +/- 0.60), surpassing the reference standard (4.19 +/- 0.83 and 4.31 +/- 0.81; P = 0.005 and P = 0.008). FU_DLx2 and PT_DLx6 scored lowest. Conclusion: Count-rich PT scans provide effective training priors for reconstructing extremely sparse images. Input-image addition and cross-domain transfer learning from PT data outperformed direct training on count-limited diagnostic data, supporting 2x scan acceleration with quality matching or exceeding the unaccelerated standard scan for selected PT-derived configurations.
Generating text-bearing images requires coherent scenes and accurate characters. Large text-to-image models improve both capabilities through repeated denoising, at substantial sampling cost. Few-step distillation reduces this cost but requires the student to organize scenes and refine glyphs within a short trajectory. We present a four-step distillation framework combining local response regularization with teacher-guided text training. Built on distribution matching, the framework distills the GLM-Image DiT while retaining its autoregressive visual prior, glyph encoder, and VAE. A path-length penalty promotes consistent local responses of the conditional denoising field. Progressive training combines synthetic strings with teacher-generated imagery to preserve character accuracy and scene-integrated typography. The student achieves GenEval Overall 0.84 and CVTG-2K word accuracy 0.5045, compared with 0.81 and 0.1690 for four-step TDM. A coefficient study favors $10^{-5}$, and Qwen-Image experiments show additional gains from response regularization and text supervision. Single-image inference takes 1.742 s, compared with 5.818 s for the teacher, yielding a $3.34\times$ speedup at the same 40 GB GPU memory requirement.
Large language model agents increasingly rely on reusable natural-language skills, yet growing skill collections remain accumulations of isolated text artifacts that are difficult to compare, deduplicate, and compose. We argue that the reusable content of a skill lies not in its surface wording but in its procedural transitions: which operation to execute under a given state and which observable outcome to produce. Based on this observation, we propose State-Move-Outcome (SMO), a compilation-like framework that decomposes raw skills into reusable procedural bricks and organizes admitted bricks as an externalized symbolic structure. When compiling a skill, SMO first retrieves and composes existing bricks, dynamically connects them through Outcome-to-State conditions, and creates a candidate brick only for an uncovered procedure; the candidate becomes reusable only after validation and admission. Experiments on SkillsBench v1.1 demonstrate that SMO represents 232 rawSkills from 87 underlying tasks with 95 shared bricks, reducing the corresponding skill representations from 366,472 to 149,154 tokens--a $\mathbf{59.3}\times$ reduction and $\mathbf{2.46}\times$ compression. Across 232 skill-conditioned evaluation instances, SMO raises the average task score from $\mathbf{0.6379}$ to $\mathbf{0.7478}$ and increases 5/5 Pass instances from $\mathbf{138}$ to $\mathbf{164}$, while matching or outperforming rawSkill on $\mathbf{208}$ of $\mathbf{232}$ instances. It also reduces average runtime token consumption by $\mathbf{38.38\%}$. These results show that compiling natural-language skills into shared procedural units can jointly improve skill reuse, representation efficiency, and downstream execution.
In this paper, we propose CM3AE 2.0, a unified RGB Frame and Event-Voxel/-Frame pre-training framework. By jointly learning multimodal spatiotemporal representations and exploiting latent cross-modal dependencies through cross-modal generation, the framework enhances multimodal information interaction and collaborative modeling capabilities. Specifically, we first propose an Event-aware Spatiotemporal Masked Reconstruction (ESMR) strategy to learn unified multimodal spatiotemporal representations. It leverages event responses to guide masking toward dynamically salient regions. Through spatiotemporal masked reconstruction, it learns spatial structures and their temporal variations from continuous RGB-Event observations. In addition, it incorporates multimodal fusion reconstruction and cross-modal contrastive learning to facilitate the modeling of cross-modal complementary information and the learning of shared representations. Building upon this, we further propose a Hierarchical Cross-modal Generation (HCG) strategy to learn latent cross-modal dependencies through visual and language generation. At the visual level, Event-conditioned RGB generation is performed using a pre-trained diffusion model to capture visual associations between the two modalities. At the language level, video-level text description generation based on joint RGB-Event representations establishes connections between multimodal visual features and linguistic semantics, further enriching the learning of cross-modal dependencies. Extensive experiments on four diverse downstream tasks demonstrate that CM3AE 2.0 consistently improves performance, validating the effectiveness and generalization capability of the proposed pre-training framework.
This paper presents DPGS-SLAM (Diffusion-Prior and Provenance-Aware Monocular Gaussian Splatting SLAM) for dense mapping in environments containing dynamic objects. Dynamic observations can contaminate camera tracking and scene reconstruction in visual SLAM (simultaneous localization and mapping). Although 3DGS (3D Gaussian splatting) enables efficient dense mapping, existing 3DGS-SLAM systems may exhibit instability when dynamic objects violate their underlying static-scene assumption. Excluding dynamic observations can reduce their influence on tracking and mapping, but may leave occluded static backgrounds missing or weakly constrained. Our proposed DPGS-SLAM is a monocular RGB SLAM framework that supports provisional representations of insufficiently observed static backgrounds through asynchronous diffusion-based background completion and provenance-aware Gaussian mapping. Dynamic regions are identified through optical-flow-based motion discovery and temporally consistent mask propagation by SAM2 (segment anything model 2). For selected keyframes, a fine-tuned diffusion inpainting model generates plausible background appearance in occluded regions, while Metric3D v2 predicts metric depth from the completed images to initialize provisional Gaussian primitives. Provenance labels distinguish generated primitives from observation-supported primitives, excluding generated content from camera-pose estimation and observed-map optimization. When previously occluded backgrounds become visible, eligible generated primitives are removed, and observation-supported primitives are independently initialized from real RGB observations and predicted depth. Experimental results demonstrated competitive camera tracking, with average ATE RMSE values of 1.2 cm on TUM RGB-D and 2.7 cm on BONN RGB-D. On StaDy4D, DPGS-SLAM improved average PSNR by 4.60 dB over WildGS-SLAM using only monocular RGB input.
Single-view 3D hair reconstruction has progressed rapidly, but many methods optimize for visual plausibility rather than producing strand assets that can be directly used in grooming, editing, or simulation pipelines. We propose SAFE-Hair, a framework that reconstructs explicit strand grooms from a single portrait by decoding hair from persistent follicles in a canonical scalp-UV domain. Given a fitted template head, SAFE-Hair projects self-supervised image features onto the scalp, uses a conditional Rectified Flow model to generate latent scalp fields, and decodes occupancy, segment directions, length allocation, and total length at a fixed set of scalp follicles. This representation enforces scalp-rooted strand identities by construction and enables direct export with follicle IDs and guide-family metadata. We further introduce attachment, collision, and bending priors to improve geometric validity without per-subject test-time optimization. To evaluate both reconstruction fidelity and asset usability, we introduce HairBench-3D, a benchmark built from public 3D hair assets with standardized alignment, rendering, resampling, and metrics for geometry, direction consistency, invisible-region completion, silhouette agreement, and collision. SAFE-Hair improves Chamfer distance from 3.74 mm to 3.28 mm, F-score from 77.9 to 81.7, invisible-region F-score from 66.1 to 69.8, and penetration ratio from 4.8% to 3.9% over the strongest baseline.
ReLU networks are well-known to have continuous scale symmetries; infinitely many reparameterized weight vectors that represent the same function. Dinh et al. (2017) have shown that, by reparameterizing, the geometry of the loss landscape (e.g., curvature) of these networks can be changed arbitrarily without affecting the underlying function. Thus, any quantity that depends on the network's geometry should be defined with care. In this paper, we show that the Empirical Bayes (EB) Laplace's approximation is not well-defined for ReLU networks, mainly due to Hessian dependence in computing the marginal likelihood, which is not invariant under ReLU scale symmetries. Thus, we can get different prior precisions, posterior covariances, and predictive uncertainties for a network having identical predictions under ReLU scaling. We define family of prior precision fields indexed by $\gamma$ and introduce $\gamma$-canonicalization, which first converts the network to its canonical representation by balancing the ratio of incoming to outgoing weights, and then fits the Empirical Bayes (EB) Laplace to the canonical representation. This procedure is equivalent to transporting the isotropic Gaussian prior from the canonical representation to the original coordinates, by producing a ReLU-scale-compatible precision field. This makes the marginal likelihood invariant under ReLU scaling, while allowing $\gamma$ to be selected by maximizing the marginal likelihood. The resulting EB Laplace approximation is well-defined on ReLU scale-equivalence classes and yields consistent prior precision selection and predictive uncertainties across functionally identical parameterizations.
Typical diagnosis of Parkinson's disease (PD) occurs only after significant dopaminergic loss, while resting-state EEG abnormalities like alpha slowing and reduced coherence occur earlier. While many published EEG-based PD detectors claim accuracy greater than 95%, many use epoch-level splits which leak subject identity. I introduce MambaPD, a bidirectional selective state space model for resting-state EEG-based PD detection, blending the channel mixing, patch tokenization and bidirectional Mamba encoder. It is compared with four baselines (TCN, BiLSTM with attention, EEGNet and a patch Transformer) in two public datasets (UNM, SanDiego) using subject-independent 5-fold cross-validation (checkpoints were selected on the held-out fold, so absolute estimates are optimistic). On UNM the accuracy is between 0.67 and 0.71, and MambaPD (0.714) is not statistically distinguishable from the Transformer (0.715), and TCN (0.708), with the highest sensitivity (0.82), while the lowest specificity (0.52). For the smaller SanDiego cohort, accuracy is in the range of 0.60 to 0.70, with the highest scores (0.703) from TCN and the lowest (0.601) from EEGNet, while MambaPD scores 0.631, with the lowest specificity (0.46) among the five models. The differences among the models on UNM are minor compared to the differences among folds. I report the whole pipeline, per-fold results, and the shortcomings of the protocol.
Enhancing image quality through super-resolution is crucial in image and video processing. While contemporary deep learning models, such as CNNs and Transformers, excel in pixel-level enhancements, they often overlook the rich semantic context within images. To address this, we introduce a novel approach that synergizes pixel intensity with semantic information for superior detail recovery. Our method involves extracting semantic data, including textual descriptors and spatial positioning, and encoding this information. We also present a Cross-Modal Feature Fusion module for generating semantic feature maps via multimodal training, and a Prompt Traction Head to augment the backbone network with semantic insights, thereby refining image details. Our empirical results across various datasets validate the superiority of our approach.
Verifying the local robustness of neural networks under $\ell_0$ perturbations is inherently combinatorial, as it must reason over all sparse subsets of $\epsilon$ perturbed pixels. This paper presents FocusBranch, a combinatorial branch-and-bound framework for $\ell_0$ robustness verification. The key insight is to make the $\ell_0$-ball decomposable and branchable using a node abstraction, which consists of a focus pixel set $\mathcal{S}$ and a level $l$, permitting at most $\epsilon-l$ pixels outside $\mathcal{S}$. This abstraction unifies two branching mechanisms: concrete nodes branch through upper shadows to fix additional perturbed locations, while super nodes branch through covering designs to cover many perturbation cases with fewer child nodes. FocusBranch proves node safety by concretizing linear relaxations over node-constrained $\ell_0$-balls and recursively branches on unknown nodes. Rather than relying on a fixed verification plan, it samples nodes across levels to estimate safe rates and costs, then selects a starting level and branching strategy predicted to minimize verification time. Experiments on fully connected and convolutional classifiers for MNIST, Fashion-MNIST, and CIFAR-10 show that FocusBranch verifies $\ell_0$ robustness effectively across architectures and outperforms state-of-the-art verifiers. The code is available at: https://github.com/simpl-group/FocusBranch.