Papers

Coupling a pretrained vision-language model (VLM) with a diffusion or flow-matching action expert has become the mainstream vision-language-action (VLA) paradigm for embodied policies, yet they remain unreliable once a task demands precise manipulation, e.g., submillimeter cable insertion. We attribute this weakness to a training objective that averages flow-matching errors across contexts and flow times, allocating little supervision to the few decisive moments determining task success. Along the generative path, inference integrates a locally supervised velocity field, so equally penalized errors at different flow times can affect the executed action unequally. Consequently, precision-task performance deteriorates as physical tolerance tightens by orders of magnitude. We call this supervision imbalance across contexts and flow times Flow-Time Starvation. The remedy is to jointly align supervision allocation and representation learning with context-dependent precision demand, enabling the policy to recognize when greater precision is needed. We instantiate this precision-forcing strategy as Precision-aware Reweighting and Internalization for Skilled Manipulation (PRISM), which scores precision demand offline, strengthens flow supervision where errors are consequential and action signals are visible, and internalizes that demand in the VLM condition representation. Simulation and real-robot experiments show that PRISM preserves or improves general-task performance while raising mean end-to-end success from 38.3% to 86.7% across drive connector insertion, bimanual connector mating, and screw tightening, substantially extending the precision boundary of existing embodied policies without new sensors or observation modalities.

On-device inference imposes constraints on model storage and operand traffic and thus benefits from compressed operand representations, which mandate the use of encoding and decoding operators. However, naïve placement of codec operators can cancel the operand traffic gains sought from compression. In this work, I present an algorithm to place codec operators into a source computational graph to make it compatible with target hardware. The edges of the graph are colored so that each induced subgraph is weakly connected and has a maximal node that both postdominates all subgraph inputs and dominates all subgraph outputs; this unique node - dubbed the $\textit{absolute dominator}$ - is the natural location to observe the traffic flowing through the subgraph and learn parametric codec operators. The algorithm's time complexity is linear in the source graph's size. With respect to naïve local codec placement, the algorithm reduces modeled operand traffic by $24\%$-$44\%$ across four models, including $24\%$ on SigLIP and $34\%$ on SmolLM2.

Despite recent advances in Aerial Vision-and-Language Navigation (VLN), single Unmanned Aerial Vehicle (UAV) search-and-rescue in outdoor scenarios remains challenging due to partial observability, making cross-altitude collaboration a promising direction. However, existing collaborative navigation methods rely on precise target localization and rigid goal-dispatching paradigms, ignoring high-altitude observation uncertainty and suppressing low-altitude exploration autonomy. To address these issues, we introduce BGC-Nav, a training-free cross-altitude collaboration framework that enables robust navigation via Belief-Guided Coupling (BGC). Specifically, the high-altitude UAV maintains a Global Belief Memory encoding task-conditioned priors, which the low-altitude agent incorporates into decision-making via Local Exhaustion Memory (LEM), Local Attraction Augmentation (LAA), and Directional Belief Prior (DBP) as an advisory bias rather than a rigid command. Extensive experiments on ESARBench show that BGC-Nav consistently outperforms state-of-the-art single-agent and collaborative methods. Code is available at https://github.com/uestchjw/BGC-Nav.

Model merging has become widely adopted for enhancing the task-specific performance of Large Language Models (LLMs) without additional training. While the specialized model from third-party platforms may undergo safety audits (e.g., backdoor detection), we reveal a novel and critical threat: One-sided Merging-Conditional Backdoor (OMCB). OMCB remains dormant and undetectable within one standalone model but activates upon merging, triggering safety unalignment. We first theoretically analyze the existence and geometric location of OMCB, revealing that the parameter shift crossing safety decision boundaries is the main cause of backdoor activation. Building upon this insight, we propose SMOKE, a meta-learning-based attack framework. SMOKE implants OMCB via a bi-level optimization strategy: the inner loop synthesizes surrogate meta-vectors through random low-rank initialization and finite-step maximization, while the outer loop optimizes the target model by simulating the merging process. Furthermore, preliminary analysis shows that the injected OMCB has bounded generalization loss under unknown merging configurations. Extensive experiments across diverse model architectures and merging techniques validate the effectiveness of SMOKE and its resistance to existing backdoor detections, highlighting a significant yet previously overlooked vulnerability in the model merging paradigm.

Most text-to-SQL benchmarks provide the target database. In a large data catalog, however, choosing the sources is part of the problem: an agent must find every database needed for a request, determine how their results fit together, and execute the resulting workflow. We present DSABench to evaluate these capabilities jointly. Its 815 human-reviewed tasks draw on Spider, BIRD, and LiveSQLBench; 618 require several databases to contribute, and 105 allow more than one interpretation of the required sources. Executable contracts define source participation, cross-source dependencies, and source-local execution, while RouteEnv lets agents inspect public catalog evidence without revealing the gold annotations. In our routing-only study across seven model backends, the highest RA-Set score is 60.55%, compared with a highest RA-Graph score of 34.66% when coordination requirements are included. The gap becomes more pronounced when execution is required. Across four backends, median non-privileged Route+Exec-Graph success ranges from 6.69% to 7.98% without clarification and from 11.44% to 13.25% with one clarification; the best observed clarified configuration reaches 16.99%. On ambiguous tasks, clarification improves exact branch-source recovery by 39.8-43.2 percentage points. These findings illustrate why retrieving plausible databases or producing correct local SQL is not enough: an agent must carry the right information through the entire multi-database workflow.

The design of high-affinity, target-specific protein binders is central to the development of modern therapeutics. However, existing protein-generation methods remain limited in this setting, as many generated designs fail to simultaneously achieve geometric and physicochemical complementarity at binding interfaces. To address this challenge, we introduce PIDiff, a physicochemically enhanced multimodal diffusion framework for protein-protein complex design. PIDiff jointly designs target-conditioned protein sequences and structures by coupling diffusion in discrete sequence space with diffusion on the continuous (SE(3)) structural manifold. At its core, PIDiff employs PIFormer, a novel architecture that interleaves geometry-aware invariant point attention with global sequence self-attention. We further curate PIBench, a protein-protein complex dataset with diverse target scaffolds derived from the Protein Data Bank (PDB). We first pretrain PIDiff on PIBench and subsequently fine-tune it for two therapeutically relevant applications: minibinder design and antibody design. PPFrame consistently outperforms baseline methods across sequence design success rate, structure binding energy, and sequence-structure co-designability on all tasks.

Feed-forward 3D reconstruction models have achieved remarkable progress through large-scale training and strong learned priors. However, they remain fragile when applied to dynamic in-the-wild videos, where non-rigid motion, occlusions, and illumination changes are prevalent. A key limitation lies in their reliance on expensive depth and camera pose annotations, which hinders scaling to diverse real-world dynamic data and restricts generalization. To address this issue, we propose PURE, an unlabeled post-training refinement framework that improves depth and camera estimation without requiring ground-truth supervision. Instead of relying on annotations, PURE elegantly distills robust structural knowledge from state-of-the-art 2D priors into 4D feed-forward architectures. PURE leverages relative depth constraints, affine-invariant geometry distillation, and pose refinement based on feature correspondences to stabilize geometric predictions under challenging dynamic conditions. Extensive experiments on both synthetic and real-world dynamic benchmarks demonstrate that PURE delivers consistent improvements across multiple state-of-the-art architectures. Furthermore, our data-efficient paradigm effectively mitigates domain shifts, successfully unlocking the architectural limits of these base models in challenging, in-the-wild dynamic scenes.

Speech-conditioned generation maps speech signals to target audio outputs. Prior end-to-end speech-conditioned generative models achieve high fidelity in clean acoustic conditions, but their performance degrades substantially under environmental noise. While prior noise-robust approaches have achieved significant success in feature-level purification, maintaining consistent generation under clean and noisy conditioning remains a key challenge. To address this, we propose AlignAudio, a noise-robust generative framework based on dual alignment: (i) aligning the representations of clean and noisy speech to preserve semantic cues, and (ii) enforcing generation-level consistency within a flow matching paradigm to improve conditioning fidelity. Experiments on speech-to-audio (STA) and voice conversion (VC) tasks demonstrate that AlignAudio maintains performance comparable to baselines on clean speech while significantly outperforming them across various noisy conditions. Furthermore, the proposed dual-alignment strategy is applicable to both single- and multi-stream speech conditioning and can be integrated into different flow matching-based generation frameworks and post-training settings. Audio demonstrations and qualitative comparisons: \href{https://alignaudio.github.io/AlignAudio}{\textcolor{cyan}{https://alignaudio.github.io/AlignAudio}}.

Bayesian optimization uses posterior uncertainty to decide where to explore, but this uncertainty is not spatially neutral. On bounded domains, Gaussian-process posterior variance is typically larger near faces, edges, and corners, giving variance-based acquisition functions a systematic boundary bias. This spatial preference is usually inherited from the surrogate rather than chosen as part of the exploration strategy. We separate these roles by keeping the predictive model fixed and modifying the spatial preference of the exploration term. We control the preference with spatial weight fields and amortize their construction with neural surrogates. Controlled experiments show that the Gaussian process' default boundary bias changes which parts of the domain are explored and which optima are reached. We compare three alternative exploration principles and find, across standard benchmark functions and stochastic robot-pushing tasks, that they can often lower regret, particularly early in the optimization. The spatial structure of posterior uncertainty is therefore not a neutral default: it shapes the exploration strategy and can materially affect optimization performance.

Instruction-following text-to-speech (TTS) typically leaves voice design implicit, as conventional speaker embeddings encode reference audio rather than text-driven targets. We propose VoTSpeech, which decouples voice design from speech generation via an explicit intermediate voice latent. Built on a shared language model backbone, VoTSpeech first predicts a continuous voice latent from text instructions using a flow-based module. This latent serves as a shared condition for both the language model and the Acoustic DiT, guiding autoregressive sequence modeling and acoustic latent generation, respectively, before an audio VAE decoder converts the acoustic latents into waveforms. To capture speaker identity and complementary acoustic nuances, the voice latent is jointly supervised by speaker embeddings and temporally pooled audio VAE features. With task-specific adaptation, VoTSpeech performs strongly on InstructTTSEval-ZH, demonstrating competitive capabilities against strong open-source models while maintaining high naturalness. These results confirm that explicit voice latent design and its conditioning interface effectively empower text-driven voice design for controllable speech synthesis. Code, model, and audio samples are available at https://github.com/ywbn/VoTSpeech .

Explicit point correspondence is attractive for adapting video generators: it describes how visible surfaces move, but can overwrite the appearance prior learned by a large pretrained model. We study this tension in a 5.01B-parameter text-to-video diffusion transformer. ACROSSVAM2.0 combines a confidence-weighted track-and-visibility head, zero-initialized residual feedback, and asymmetric PCGrad that removes only the auxiliary component opposing the primary flow-matching gradient on shared LoRA parameters. The design preserves the generator exactly at initialization and protects its primary update during adaptation. In a four-way ablation trained for 1,000 steps with three independent seeds, the full method lowers KLT forward-backward cycle error in all three seeds relative to a matched flow-only continuation (mean seed-level difference -0.00711, SD 0.00505), while robust sharpness remains positive in all three. An expanded 24-prompt evaluation (216 videos) yields 9.7% lower aggregate CoTracker3 cycle error than the matched control and near-matched performance on six VBench dimensions; however, the seed-level tracking effect is heterogeneous and I3D-FVD regresses from 0.1824 to 0.2022. These results support a bounded claim: conflict-aware adaptation improves correspondence consistency without broad perceptual degradation, but does not uniformly improve video quality. We release equations, pseudocode, matched outputs, source tables, and artifact provenance.

Recent advances in 3D hand-object interaction (HOI) perception have opened new opportunities for learning dexterous robotic skills from human videos. However, existing dexterous imitation learning methods often rely on costly and manually teleoperated robot demonstrations, requiring substantial human effort. Therefore, we build DEXCAST, a framework that learns dexterous robotic hand control directly from monocular RGB-D videos by leveraging joint HOI reconstructions, without requiring any teleoperation. Furthermore, how to model high-dimensional robotic hand motion and extract temporal context to inform future action generation remains a key challenge. To this end, we propose a DexHand Motion Perception Module (DMPM) to capture inter-joint coordination and temporal dependencies in compact latent representations that condition our diffusion policy. Extensive real-world experiments demonstrate that DEXCAST achieves exceptional performance across diverse dexterous manipulation tasks.

LLM-augmented data workflows often involve multiple heterogeneous models operating over substantial shared context, such as schemas, tool specifications, interaction history, and intermediate results. Although prefix caching can reduce prefilling overhead, existing approaches are largely limited to reuse within the same model because KV states are tightly coupled to model-specific architectures. We present MapFlow, a system for efficient heterogeneous LLM serving via cross-model attention map reuse. Instead of reusing KV states, MapFlow reuses attention maps, which capture token dependency patterns and are more transferable across models. To enable efficient and accurate reuse, we combine partial recomputation with lightweight linear translation across layers and heads. To make attention reuse practical for long contexts, we further introduce block-level sparse compression and a fused retrieval kernel. Experiments show that MapFlow reduces prefilling latency and improves end-to-end serving efficiency while preserving generation quality across heterogeneous model pairs and multi-turn routing workloads. These results suggest that attention maps provide an effective reusable state for heterogeneous LLM serving, especially in prefix-heavy data workflows. Our code is available: https://github.com/PKU-DAIR/MapFlow.

Many physical systems exhibit one-way coupling: exogenous fields such as medium density or obstacle geometry influence the dynamics without being affected by them. We investigate how explicitly representing these fields can improve the accuracy and computational efficiency of transformer-based PDE surrogates. Building on spatially adaptive normalization, we introduce Patchwise Adaptive Normalization (PAdaNorm), which partitions exogenous fields into patches aligned with the state tokenization. We evaluate PAdaNorm on datasets from three different domains, using a Walrus-inspired transformer architecture. Across the three PDE families, PAdaNorm is on par with or up to 16% better than parameter-matched baselines in mean long-horizon variance-scaled RMSE on 60-step autoregressive rollouts. We also find that PAdaNorm better maintains hard physical constraints over long rollouts, such as preventing flow into impenetrable solid regions in porous media. When the conditioning is static, PAdaNorm caches the conditioning embeddings once and reuses them across the entire rollout, reducing inference FLOPs per step by up to 36% compared with parameter-matched baselines. An ablation over eight conditioning variants shows that modulating a single attention axis (temporal or spatial) suffices and that a lightweight shared MLP across blocks remains competitive with per-block projections.

Agentic applications generate multi-turn sequences of LLM requests, with each new turn extending the context accumulated from prior interactions and tool calls. Prefix caching reuses the KV cache of earlier turns and avoids redundant prefill, but routing every turn to its cache-affine instance can create severe local queues while capacity remains idle elsewhere. Remote KV cache access relaxes this tension, yet migrating execution adds transfer work and is harmful when the cluster is already congested. Thus, deciding where to execute a request requires reasoning jointly about cache locality, queueing, and the current operating point of the cluster. We present Vesta, a queueing-aware scheduler for distributed agentic LLM serving. Vesta assigns each session a stable holder for its authoritative KV cache while independently selecting an executor for every request. Grounded in classical queueing theory, its objective accounts for backlog, reusable prefixes, KV cache transfer, and the queueing externality of additional service demand; it migrates execution to recover idle capacity when beneficial and favors placements that add the least prefill and transfer work as congestion rises. Vesta also migrates reusable prefixes across length buckets to avoid cold starts. On real-world agentic workloads, Vesta improves throughput by 14.0% and reduces mean TTFT by 57.8% over the state-of-the-art scheduler in a PD-colocated deployment. In PD-disaggregated deployments, it reduces mean TTFT by up to 18.1% and P99 TTFT by up to 29.0% at matched throughput.

Recent advances in video diffusion enable novel-view video synthesis from a single input video, supporting world modeling, embodied intelligence, and robotic simulation. However, existing methods process camera trajectories independently, lacking the cross-trajectory consistency needed for coherent, navigable 4D worlds. We introduce Demiurge, an autoregressive video rerendering pipeline built on a dynamic 4D point-cloud memory cache. It promotes cross-trajectory consistency through four components: (1) an appearance video diffusion model that synthesizes photorealistic novel-view videos conditioned on the 4D memory; (2) a memory-consistent video depth completion model that aligns known-region depths with cached geometry and completes missing regions, mitigating scale ambiguity and discontinuities during memory updates; (3) memory-aware dual-stream warping that provides high-fidelity visual guidance and occlusion cues for novel-view synthesis; and (4) geometry-aware video refinement that enhances perceptual quality. Extensive experiments demonstrate state-of-the-art performance and three key capabilities: coherent spatiotemporal lattice construction over dense multi-view trajectories; large-parallax video rerendering under challenging camera motions, including 180-degree rotations with no input-view overlap; and geometry-faithful outputs that can be directly optimized into high-quality 3D Gaussian Splatting representations with minimal artifacts. Project page: https://hanxinzhu-lab.github.io/demiurge_project_page/.

Large language model (LLM)-based AI Scientists advance scientific discovery, yet they suffer from inefficiencies due to noisy observations. Existing principle-evolving agents act on every anomaly, leaving its source unresolved: a new mechanism or an evaluation artifact, so invalid revisions bias evolution. To address this, we formulate revision as evidence admission: anomalies steer search immediately, but only independently verified evidence may rewrite persistent principles, and contradicted principles are retracted. We realize this in TRIAGE-PIEVO , which pairs an anomaly-responsive search portfolio with a staged verification gate that alone controls what enters principle state. Across four scientific tasks with three backbones, TRIAGE-PIEVO (1) attains the highest average solution quality with +7.5%~11.8% over the strongest baseline per backbone, matching the strongest baseline's full-budget quality with one-third of the evaluation budget on Qwen3-235B; and additional stress tests and cross-harness deployment demonstrate that (2) in constructed stress tests, it admits none of the corrupted, hidden-source, or exploit evidence that ungated variants admit, while missing no planted mechanism, and (3) plugged unchanged into three external research harnesses, it cuts false evidence admissions by over 60% and raises verified scores by 31%~43%.

Vision-and-language navigation (VLN) agents are typically evaluated by whether they reach a target and how efficiently they follow an instructed route, yet these metrics provide little information about collision behavior. Moreover, continuous VLN evaluation commonly uses a sliding regime, where an agent can slide along obstacles after a collision, potentially masking the effect of unsafe actions on navigation performance. We introduce VLN-Safe, a benchmark for analyzing collision safety under two complementary evaluation regimes: sliding, reflecting common VLN evaluation practice, and blocking, where collision-inducing actions prevent forward motion and more directly expose their consequences. Across six published agents on R2R-CE, 38.5-56.8% of successful episodes contain contact under sliding, while switching to blocking decreases success rate by 11.7-36.1 percentage points. We further analyze signals for anticipating imminent collisions and find that depth geometry and recent movement history are more informative than policy-derived signals, while safe alternatives near the policy's preferred action are rarely available. Motivated by these findings, we introduce a training-free collision guard that vetoes risky forward actions, executes an escape maneuver, and returns control to the original policy. Under blocking, the guard reduces collision steps by 62-98% and improves success rate by 6.9-29.3 percentage points across five recurrent, topological, and vision-language-action agents. Under sliding, it leaves success rate unchanged while raising the share of collision-free successes by 4.1-12.4 points. These results reveal a substantial gap between task success and collision-safe navigation and show that lightweight collision-aware intervention can improve both safety and navigation performance.

In this paper we study pairwise compatibility graphs whose witness tree is required to be an unweighted caterpillar. This is a natural restriction of PCGs, in the same line of research as the study of star witnesses, which led to the class of double-threshold graphs. In the caterpillar case, we show that this restriction is equivalent to a one-dimensional difference-labelling problem: a graph admits such a representation if and only if its vertices can be assigned integer labels so that adjacency is determined exactly by whether the absolute difference of two labels belongs to a given interval. Using this characterization, we derive structural properties of the class, including bounded twin-width, and study its relation with familiar graph families. We show that the class contains cycles and two-dimensional grids, while it excludes natural obstructions such as the subdivided star $S_{2,2,2,2,2}$, the cube graph $Q_3$, the wheel $W_6$, and the complement of the $3$-sunlet. We also consider restricted representations, including centipede witnesses and representations with a given interval and we give a polynomial-time recognition algorithm when the interval is fixed.

Sign Language Recognition (SLR) presents complex challenges due to variable signing speeds and spatial intricacies. Specifically, Indian Sign Language (ISL) requires precise modeling of its inherently bimanual nature and is bottlenecked by the scarcity of large-vocabulary benchmarks and the inability of standard models to handle irregular signing velocities. To address this, a structurally-aware hybrid framework for efficient, gloss ISL recognition is proposed. To minimize computational redundancy, an entropy-driven SVD sampler that dynamically extracts informative keyframes is used. Because this subsampling creates irregular physical time gaps, the proposed architecture employs a Continuous Dual Positional Encoding paired with a distance-aware Elastic Transformer to preserve natural signing rhythm. Spatially, a Kinematically-Conditioned Adaptive GCN explicitly models bimanual geometry using unified bone-joint features. To validate our framework, a 120-class, 9,453-video native ISL dataset is created. Augmented by Physical Node Dropout and ArcFace optimization to mitigate subject-specific memorization, our framework achieves a competitive 89.8% accuracy, establishing a strong baseline for complex sign language modeling.

We introduce HumanPhys, a human-centric benchmark for real-world physical perception, and a training corpus emitted by the same pipeline. Each question targets a physically grounded quantity related to people, including depth, inter-person distance, gaze angle, displacement, velocity, heading-step trajectory, blink count, and contact-deformation strength. The rendered probes additionally test illumination inference from body and face cues and camera ego-motion, both with and without a person in the frame. All labels are derived programmatically from sensor calibration, source metadata, or renderer state, rather than written by annotators or produced by estimators on the benchmark frames. HumanPhys systematically probes six dimensions of human-centric physical perception: spatial state, kinematics, blink events, deformation, illumination, and camera pose. To the best of our knowledge, no existing MLLM benchmark-and-training corpus jointly supervises this set of human-centered physical quantities under direct metadata- or renderer-derived labels. The six dimensions appear only partially dissociable across the tested MLLMs. We benchmark a range of open- and closed-source MLLMs on HumanPhys and find that all six capability dimensions remain challenging for current models. We hope HumanPhys will support future research on multimodal systems that perceive and reason about people in the physical world.

Moving operands to a processor can cost more than the arithmetic performed on them, and for small binary similarity and search workloads data movement can dominate. We present HADES, a programmable processing-near-memory engine built around one shared 32x32-bit 2R1W register-file macro that holds both data and program. Twelve 16-bit opcodes, three working registers and a 32-slot program express ten search workloads in 8 to 19 slots, and SCANP and SCANSEL shorten dual-row and selected scans. A post-layout 2x2 Tiny Tapeout implementation closes timing at 40 MHz with 26,968 square microns of standard cells beside a 15,744 square microns macro. All HADES results come from RTL or gate-level simulation of this implementation, and HADES energy is estimated. Against ESP8266 and ESP32 microcontrollers running plain C, and with an ideal host, HADES answers a resident query 10x to 32x and 2.2x to 11x faster, respectively. End to end, it wins at the evaluated batch sizes of B = 16 and B = 256 against both boards and loses the single-query (B = 1) runs to the ESP32. An Arduino Nano runs a restricted subset. Kernel power stays within about 2% of a spin loop on all three boards. These results hold for plain-C baselines under the evaluated conditions. At this scale the control and instruction storage fit around a shared memory macro, while the byte-wide interface and the loading path remain significant costs.

We present CAST-ASD (Causal Attention over Speakers and Time), a causal joint model for active speaker detection (ASD). Recent ASD models improve performance by jointly processing all faces in a scene, but operate offline, with the full clip available before any frame is labelled. Low-latency causal detectors instead process each face independently. CAST-ASD combines joint face reasoning with fully causal inference. It applies scene-level attention across faces within the current frame and causal temporal attention over each face's history, enabling information sharing across speakers with zero lookahead. Its parameter count does not grow with the number of faces. On MSDWild, where most negative face-frames contain a visible competing speaker, CAST-ASD reduces competing-speaker false alarms by 32% relative to the causal per-face detector it extends, with larger gains as scene occupancy increases. It outperforms all causal baselines on both validation splits while using only one-tenth the inter-speaker parameters of a three-block fixed-roster convolution. Code and models are available at https://github.com/prathamg007/CAST-ASD.

User-generated audio collections frequently contain textual metadata such as titles, tags, and descriptions. Effectively leveraging these complementary fields requires accounting for both field het erogeneity and varying semantic relevance relative to the acoustic content. We propose FAM-Gate, a two-stage framework for metadata-enriched audio classification. In the first stage, field-aware metadata-view attention dynamically aggregates multiple metadata views using audio-guided weighting, explicit field-identity embeddings, a shared identity-initialized projection, and mean-anchored regularized aggregation. In the second stage, a prior-constrained vector gate performs bounded, dimension-wise fusion of audio and metadata representations centered around an acoustic prior. Ex periments on BSD10K, BSD35K, and FSD50K show consistent improvements in both hierarchical single-label and multi-label classification. Compared with matched random-view scalar-fusion base lines, FAM-Gate improves the hierarchical F-score by 2.53 and 5.72 percentage points on BSD10K and BSD35K, respectively, while increasing macro mAP and macro F1 by 5.07 and 10.25 percentage points on FSD50K.

Training-free sparse attention is a practical route to efficient video diffusion transformer inference, yet existing methods typically fix the competition domain at each query row for every attention head. Prior work primarily addresses how to estimate token-level importance within this fixed domain; whether the domain itself should adapt to head heterogeneity remains under-specified. We introduce EG-Sparse, a training-free framework that routes low-entropy heads to a grouped cross-head pool while retaining per-row local selection for high-entropy heads, using only scores already materialized by the sparse pipeline. A controlled diagnostic fixes the ranking score and varies only the competition rule: grouped cross-head selection reduces low-entropy-head reconstruction error by 9.8%, whereas forcing high-entropy heads into an all-head global pool increases it by 7.9%. Across four configurations, EG-Sparse achieves higher video quality while reducing attention density and GPU denoising latency compared with strong training-free sparse-attention baselines. Mechanism ablations further show that selective domain adaptation outperforms uniform, random, and fixed domain-assignment policies. These results support competition-domain adaptation as a distinct design axis within centroid-guided training-free sparse attention.

Streaming video understanding requires answering questions from the observed video, with evidence drawn from recent observations or earlier events. Historical context can improve recall but may interfere with judgments based on recent frames. Its usefulness depends on both when it is introduced and which clips are selected, while relevance alone does not guarantee sufficient evidence for an answer. We propose SPARE, a framework that combines adaptive historical retrieval with visual encoder sharing. A lightweight gate predicts whether recent observations provide sufficient evidence to answer the question. When additional evidence is predicted to be necessary, a multimodal embedding model retrieves relevant historical clips to supplement the current context. LoRA instruction tuning trains the understanding model to use recent observations and retrieved history jointly and abstain when the evidence remains insufficient. To reduce redundant visual parameters, the understanding and retrieval branches share a visual encoder while retaining separate language backbones. Four linear adapters align the shared features with the retrieval model through distillation at the visual token and final retrieval embedding levels. With visual-only inputs, SPARE achieves 77.78% on the evaluated Real-time and Backward task groups of OVO-Bench and 73.31% on StreamingBench. Compared with Qwen3-VL-8B-Instruct evaluated at 2 fps, SPARE improves accuracy by 20.68 percentage points on OVO-Bench and 14.15 percentage points on StreamingBench.

Dividing each trajectory's score contribution by its realized length can change the direction of a reward-centered policy update. In a finite trajectory model, we derive an exact ascent threshold for a three-outcome softmax policy and exhibit a binary-reward example in which the expected Euclidean update decreases task success. Natural-gradient ascent holds for fully controllable outcome distributions with deterministic binary labels, but can fail under a one-parameter restriction. A third-order U-statistic estimates the first-order return change without bias from independent current-policy trajectories, one per request. In the example, a same-sample plug-in estimator instead has the opposite sign in expectation for batch sizes two through forty-nine. Exact policy-grid enumeration confirms the local sign reversal. In a branching process, length-normalized learning initially loses return, then recovers and exceeds the unnormalized comparator. The analysis concerns the expected update field at the current policy.

We consider stable and effective restarts of the SS–RR and SS–RRR methods for computing the eigenpairs of a large Hermitian matrix $A\in\mathbb{C}^{n\times n}$ with the $n_{ev}$ eigenvalues in a given interval. Taking the CJ–SS–RR and CJ–SS–RRR methods for instance, we prove that the commonly used restart scheme and its orthonormalized variant, named the natural restart schemes, are not numerically stable, thereby lowering attainable accuracy of approximate eigenpairs and missing some of the $n_{ev}$ desired eigenpairs. To cure these deficiencies, we propose a new restart scheme, called the combination restart, that takes the $n\times \ell$ starting matrix to be the product of the matrix, which consists of the selected $n_{ev}$ approximate eigenvectors, and an $n_{ev}\times \ell$ Gaussian random matrix. We prove the remarkable mathematical and numerical superiority of the combination restart over the natural restarts, then present a hybrid restart scheme that fully exploits the advantages of the natural restart and combination restart, and finally develop CJ–SS–RR and CJ–SS–RRR algorithms with the hybrid restart. Numerical experiments confirm that the new algorithms are numerically backward stable to the working precision and more efficient than the CJ–SS–RRR algorithm with the natural restart. The combination restart and hybrid restart schemes can be straightforwardly extended to the SS–RR and SS–RRR methods for general standard and generalized eigenvalue problems.

4D Gaussian Splatting (4DGS) has emerged as a powerful dynamic scene representation, supporting real-time high quality rendering but suffers from prohibitive storage overhead. Existing methods typically attempt to solve this by treating compression as a per-scene optimization problem, limiting generalizability and applicability where low-latency compression is required. In this work, we leverage existing image and video codecs and propose a generalized post-training compression framework for 4DGS to address these issues. Specifically, we apply importance-based adaptive quantization to Gaussians according to their importance levels, effectively refining bitrate allocation while remaining fully compatible with existing codecs. To further improve compression efficiency, we utilize the statistical properties of 4DGS and propose motion- and importance-aware packing to reorganize attributes into codec-friendly 2D maps. Our approach can achieve over 40x compression rate in seconds while maintaining high quality, facilitating the practical deployment of 4DGS.

Self-improving agents use failure feedback to revise their code or instructions, but gains after selection do not establish target-specific repair or generalization. We measure directedness by comparing repair of a randomly selected target failure with repair of the parent's other failures. Across 145 eligible modifications in public Darwin Gödel Machine logs, we find no detectable target-specific advantage. Controlled instruction-revision experiments with small open models show that relevant feedback can improve target repair. In a second study with a pre-specified analysis plan, target feedback raised repair by 20 percentage points for Qwen3-4B relative to matched same-category feedback; the 7-point increase for Phi-4-mini was not significant. Target repair did not reliably imply held-out improvement. Held-out gains were concentrated in settings with identifiable, repairable failure mechanisms, including planted errors and natural tool-interface failures. Benefits of edit localization and regression screening varied across settings, while eight-step revision loops and GEPA ended at similar mean held-out accuracy. Evaluations should report target repair, transfer to related failures, and held-out net change separately.