论文库

Dual-batch overlap (DBO) hides mixture-of-experts (MoE) communication behind another microbatch's computation, but only after dispatch send releases the GPU. The time dispatch send holds the GPU therefore stays exposed, while network progress after release can be hidden. We evaluate GPUDirect Async - Kernel Initiated (GDA-KI) Free Flow, an NCCL-EP research configuration that leaves WQE construction on the GPU but moves send progression to a host thread. On a GB300 NVL72 system, GDA-KI Free Flow reduces dispatch-send time relative to default GDA-KI at every operating point in the evaluated parameter sweep. An experiment that inserts a GPU delay between dispatch send and receive shows that offload increases the in-flight interval. Depending on the load, 32-96 microseconds of computation or other independent work between send and receive hides the deferred progress, leaving less exposed time than default GDA-KI. In disaggregated vLLM serving with DBO, GDA-KI Free Flow lowers median TPOT by 15-25% relative to NCCL-EP default GDA-KI and by 16-42% relative to DeepEP v1.

We study multipoint-to-multipoint multicasting in synchronous radio networks with unknown participant locations and unrestricted message aggregation. Let $n$ be the network size, $k$ the number of participants, and $d$ their hop diameter. A supplied distance-two coloring with $C$ colors yields completion within $(C-1)d+1$ slots and at most $nk$ transmissions. In unit disk graphs, coordinates, one-hop neighbor data, and a common degree bound $Δ_0$ suffice for a $15(Δ_0+1)$-color schedule. A known fixed separation bound permits optimal deterministic $Θ(d)$ completion from coordinates alone. Without geometric data, strong selectivity gives $O(d\min\{n,(Δ_0+1)^2\log n\})$ slots when $n$ is known. Independent random transmission gives $O((Δ_0+1)(d+\log(k/η)))$ slots with failure probability at most $η$; for fixed $Δ_0$, this simplifies to $O(d+\log(1/η))$. A star construction shows that the logarithmic term is necessary for this random rule when the degree grows. We extend color schedules to fixed deletions and subgraphs connected over bounded windows of cycles. Separate conditional reductions preserve additive costs under diameter guessing and control parameter-table and cluster preprocessing. Finally, a counterexample shows that pruning and reranking a breadth-first tree can destroy interference-free delivery. For one-hop designated delivery with receiver degree at most $b$, we prove that the optimal schedule length equals the chromatic number of an associated conflict graph and is at most $2b-1$, with matching examples.

Pre-trained encoders fine-tuned on natural language inference (NLI) and other language understanding tasks keep much of their accuracy after the input is perturbed, e.g. by removing all nouns or shuffling the words. Does the same hold for large language models (LLMs), and if so, is it because they have seen the benchmarks during pre-training? In this paper we revisit these perturbation experiments with eight LLMs alongside fine-tuned encoders on GLUE and the three rounds of ANLI. Regardless of model size, every system that answers directly retains 36-72% of its above-chance performance without nouns on MNLI compared to only 14-29% on ANLI. Only a model that reasons before answering departs from this pattern. We also check contamination indicators (including exact $n$-gram counts in the public pre-training data of two of the models) and find no link between exposure and robustness. Our strongest evidence comes from 368 new, human-validated items that none of the models can have seen. All eight LLMs keep more of their performance without nouns on items written in MNLI's style than on items written without its cues. When the perturbation instead removes the function words carrying such cues, this advantage reverses. In light of these results, we argue that robustness to perturbation on familiar NLI benchmarks reflects dataset shortcuts rather than memorisation, and tells us more about the benchmark than about the model.

Hyperbolic vision-language models are designed to encode abstraction geometrically: general concepts near the origin, specific ones farther out, and entailment cones representing directed order. We ask whether trained MERU, HyCoCLIP, and PHyCLIP models actually use these mechanisms. We audit seven released checkpoints and matched from-scratch interventions. All audited converged checkpoints remain near-Euclidean in the dimensionless radius $u=\sqrt{c}ρ$: the largest observed image-side value is $0.37$ -- well below $u\approx0.84$, where local metric distortion reaches $10\%$. Releasing the curvature floor changes curvature and norms but not this regime, with mixed, generally modest downstream shifts. Trained entailment cones are saturated or nearly saturated, so low violation rates can arise from trivially wide cones rather than learned order. Preregistered semantic traversal detects weak within-branch order but no operative full-hierarchy readout. Shuffle-controlled tests detect no pair-specific radial ordering in released checkpoints, and no positive result is consistent across all three matched ViT-B seeds. This non-detection is sensitivity-bounded: the diagnostic reaches $80\%$ power only at $\approx13.5$ pp of shuffle-controlled excess, while the largest released gap is $\approx5$ pp. We trace this to a low-curvature shortcut: lowering curvature widens entailment cones and suppresses violations without learning order. In the probed trajectories, gradient decomposition identifies entailment as the dominant curvature-lowering pressure during collapse. Yet curvature contracts even when entailment is removed, so the shortcut is not the sole cause. Under our diagnostics, the audited formulations do not demonstrate an operative radial or cone-based hierarchy. We distill the audit into a five-number geometry report for evaluating future hierarchy claims.

We present Kosmo-GFM, a scene reconstruction system that integrates a geometry foundation model (GFM), a pose optimization engine (PoseEngine), and reconstruction-derived supervision from KosmoWorld. GFM predicts camera parameters, depth, cross-view correspondences, and dynamic 3D points from images and videos. PoseEngine combines these predictions with geometric optimization to recover scene-level camera poses and geometry for downstream 3D and 4D Gaussian Splatting. KosmoWorld generates images and geometric supervision from reconstructed real-world scenes to improve GFM's geometric predictions. On the Sparse and Medium subsets of SpatialBench, GFM achieves the lowest absolute relative depth error and the highest relative-pose AUC at 5 degrees among the compared methods. It also achieves the lowest mean endpoint errors for dynamic points and scene flow across the five evaluated datasets. Across 34 scenes completed by all compared reconstruction systems, PoseEngine achieves the highest scene-averaged pose AUC at thresholds of 1, 3, and 5 degrees, with a cumulative runtime of 14.07 hours, approximately 20% of GLUEMAP's 69.67 hours. PoseEngine Flash provides a further accuracy-efficiency trade-off, with a measured mixed-precision runtime of 2.62 hours, approximately 4% of GLUEMAP's runtime. Separate scaling experiments show generally favorable trends with increased data, model computation, and training budgets. These results support the use of learned geometric priors within a scene-level reconstruction pipeline.

We establish the first nontrivial upper bound on the number of edges of a Gomory corner polyhedron, resolving a long-standing open question. For an $n$-dimensional corner polyhedron associated with an integer lattice of determinant $δ$, the number of edges is at most $$ O\left(n6^n \cdot \bigl(1+\log_2(2δ)\bigr)^{2n-3}\right). $$ This improves the standard bound obtained by counting pairs of vertices, reducing the logarithmic exponent from $(2n-2)$ to $(2n-3)$.

The use of accurate and reliable open-source human activity recognition (HAR) models on passively collected wrist-accelerometer data is essential in large-scale epidemiological studies that investigate the association between physical activity and health outcomes. While self-supervised learning has generated considerable excitement in improving HAR, the extent to which these models, coupled with hidden Markov models (HMMs), would make a tangible improvement to classification performance and the effect this may have on the predicted daily activity intensity compositions is unknown. Using up to 24 hours of wrist-worn accelerometer data from 151 CAPTURE-24 participants (aged 18-91, mean age 42, 66% female), we trained the ActiNet model, consisting of HARNet, a self-supervised, 18-layer, modified ResNet-V2 model, followed by hidden Markov model (HMM) smoothing, to classify labels of activity intensity. The performance of this model, evaluated using 5-fold stratified group cross-validation, was then compared to a baseline random forest (RF) + HMM, established in existing literature. Differences in performance and classification outputs were compared with different subgroups of age and sex within the CAPTURE-24 population. The ActiNet model was able to distinguish labels of activity intensity with a mean macro F1 score of 0.80 and a mean Cohen's kappa score of 0.85. This exceeded the performance of the RF + HMM, trained and validated on the same dataset, with mean scores of 0.76 and 0.80, respectively. The improvements in performance were consistent across subgroups of age and sex. These findings encourage the use of ActiNet for the extraction of activity intensity labels from wrist-accelerometer data in future epidemiological studies.

We explore the use of simulated data for training a model for protein annotation in crowded cryo-electron tomography volumes reconstructed from images collected at limited tilt angles and severely corrupted by the measurement operator. Firstly, we leverage the corruptions imposed by the forward model to generate domain-specific augmented paired views of the exact same scene for an invariance objective integrated into the LeJEPA self-supervised training framework. Secondly, we use additional information from the simulation pipeline such as the positions and identity of proteins in the simulated volumes to inform the architecture of the model and the loss function, so that semantic information is localised at protein positions in the resulting dense feature volume. The resulting model, CARNIVAL, is evaluated without finetuning on classification and detection tasks in real tomograms, using a benchmark dataset containing multiple protein types and two tomogram processing types. We show that CARNIVAL outperforms a state-of-the-art model trained using a contrastive objective on simulated data but without forward model-based paired views or privileged information.

For any fixed integers $q\ge2$ and $t\ge1$, we construct explicit $q$-ary codes of length $n$ that correct any combination of at most $t$ insertions, deletions, and substitutions, with redundancy at most $2t\log_q n+O_{q,t}(\log\log n)$ $q$-ary symbols. This matches the leading redundancy term of Levenshtein's classical existential bound. Building on the standard graph-coloring approach, we use a prime-based multicoloring of the edit-confusion graph to assign multiple colors to each $q$-ary word of length $m<n$. Appending a short, edit-protected representation of an assigned color to the length-$m$ word yields a length-$n$ codeword. The choice among the assigned colors carries additional message information, thereby reducing the redundancy. The construction admits deterministic encoding and decoding in $\widetilde O_{q,t}(n^{6t+2})$ and $\widetilde O_{q,t}(n^{7t+2})$ time, respectively.

As artificial intelligence (AI) becomes more common in scientific writing, some authors show fewer signs of AI assistance than expected across successive periods. We analysed OpenAlex publication histories for 71,133 authors with at least one eligible first-authored English-language publication each year from 2023 through September 2026. Using abstract-based writing scores, we defined resistance as the predicted minus observed share of abstracts classified as AI-assisted. Positive scores indicate fewer such abstracts than expected. Early resistance was positively associated with later resistance, although the share with positive scores fell. Environmental characteristics contributed more than individual characteristics to predicting later scores, although overall predictive power was low. Having a collaborator with a non-positive early score was associated with lower subsequent resistance. Authors with positive mean scores in both 2023-2024 and 2025-2026 cited older work, drew on fewer subfields and more often cited sources they had used before than authors with non-positive mean scores in both periods. These differences were evident before 2023. The adjusted gap in reference reuse rose from 3.24 percentage points before 2023 to 7.95 in 2025-2026. The findings link variation in signs of AI assistance to earlier differences in how scientists used prior work.

The Solvability Complexity Index (SCI) provides an extensional limit-height formalism for recovering a target map $Ξ$ from finite samples of an evaluation interface $Λ\subseteq\mathbb C^Ω$ by finite-height towers of pointwise limits. We first give a foundational analysis of what this extensional framework does and does not determine. In second place, we give a concrete non-trivial bridge result from a comparable SCI object to the well known limit rank of Weihrauch theory.

Environmental, Social and Governance (ESG) scores are increasingly used to evaluate corporate sustainability, supporting investment decisions, ensuring regulatory compliance and assessing risk. However, current ESG evaluations largely rely on proprietary methodologies that often produce inconsistent ratings. This motivates the development of data-driven prediction frameworks. In this paper, we present ESGPred, a modular framework for predicting ESG scores by combining domain-specific language modeling with graph-based learning over a network that models the relationships among the companies being evaluated. In addition to presenting the rationale behind ESGPred and detailing its ESG score prediction process, we show that ESGPred can effectively predict both overall ESG ratings and their individual Environmental, Social, and Governance components. We evaluated ESGPred using a dataset of 5,266 sustainability reports published by 2,134 publicly listed companies between 2020 and 2023. In this evaluation, we used ESG ratings provided by LSEG Data & Analytics as ground truth. Our experimental results demonstrated that integrating graph-based relational learning improves prediction performance over purely text-based baselines.

In this work, we investigate permutation decoding for algebraic geometry (AG) codes arising from algebraic curves defined by separated polynomials. Using automorphisms of the underlying curves, we construct permutation automorphisms of the associated algebraic geometry codes and exploit the resulting orbit structure to determine information and check positions. We introduce a class of curves, called SAP curves (Separated Additive Polynomial curves), and investigate one-point AG codes defined on them. For these codes, we obtain permutation decoding sets that correct burst errors supported on coordinates associated with rational points sharing a common coordinate. We further identify a subclass of special SAP curves, including Hermitian curves, generalized Hermitian curves, and certain maximal curves, for which additional automorphisms yield more powerful decoding sets.

Hosted retrieval-augmented generation (RAG) and semantic search allow users to query valuable provider-held corpora, raising two competing demands: to hide each query and chosen result, yet reveal only the documents that the user is authorized to receive. Existing cryptographic approaches either make this costly by processing the entire corpus for every query, or sacrifice quality for efficiency by scanning a few clusters. We repurpose learned deep hashing as a private filter: a randomized binary code points the provider to a short candidate list, while encrypted reranking and oblivious key transfer protect the precise query and final selection. This shortlist short-circuits full-corpus cryptographic search without sacrificing retrieval quality: with 200-500 candidates, it closely matches full-corpus retrieval across five zero-shot corpora spanning 25K to 5.4M documents. On the full 2.68M-passage NQ corpus over a 10-Gbps link, our protocol only adds 0.38 seconds, or 6.5%, to a 128-token Qwen3-32B RAG pipeline. The released code satisfies directional metric differential privacy (DP) and substantially reduces embedding-inversion and property-inference leakage, demonstrating that a carefully learned shortlist can make private dense retrieval both accurate and practical.

This paper proposes a novel class of Bayesian Conjugate Distributions (BCDs) on extended special Euclidean groups SE_n(3) for global motion uncertainty representation and estimation. BCDs are constructed by restricting matrix Gaussian distributions on ambient Euclidean space to SE_n(3), avoiding singularity of local coordinates. The unique feature of BCDs is their conjugacy to a broad class of sampling distributions arising in motion estimation, enabling exact closed-form Bayesian inference. Optimal estimation with respect to BCDs is further investigated for implementing motion estimation. Two means are extracted as random state estimates, which maximize probability and minimize mean squared error, respectively. Moreover, within the BCD class, moment matching is shown to minimize the Kullback-Leibler divergence, and maximum likelihood estimation is also derived, which provide analytical methods to approximate underlying distributions with the optimality guarantee. As an application, a Bayesian estimator for pose and velocity estimation of an Unmanned Aerial Vehicle (UAV) is developed within an assumed density filtering architecture. The propagation and estimate extraction leverage the optimal estimation related to BCDs, and the landmark measurements are exactly fused with the BCD prior in closed form due to the conjugacy. Numerical simulations of a UAV flight indicate the advantages of BCD-based estimators over two classic Kalman filters on Lie groups with various initial uncertainties and measurement noise.

We consider a special block-structured integer program with $n$ blocks. Every block has a prescribed number of vectors that need to be selected while minimizing the objective $c^{\intercal}x + g(Ax)$, where $c$ is a rational cost vector, $A$ is an integral matrix with $d$ rows, and $g \colon \mathbb{Z}^d \to \mathbb{Q}\cup \{+\infty\}$ is an arbitrary function. We access $g$ via an oracle that minimizes $c^{\intercal}x + g(Ax)$ when all but $O(d)$ variables are fixed. Our algorithm runs in time $(ndΔ)^{O(d)}(Q + \text{poly}(|I|))$ where $Δ$ is the largest absolute value of an entry in $A$, $Q$ is the necessary time to evaluate the oracle for $g$ and $|I|$ is the size of the input. This model is a variation of the model studied by Dadush et al.~[IPCO 23]. Their model has only a single block but individual upper and lower bounds on the variables instead of prescribed block sums. For their model they present an algorithm with running time $t^d(dΔ)^{O(d^2)}Q$ where $t$ is the number of variables. We use our algorithm to solve mixed-integer linear programs with no upper bounds on the variables, a linear objective, and arbitrarily many constraints involving only fractional variables in time $(mΔ)^{O(m)}\text{poly}(|I|)$, where $m$ is the number of constraints involving integral variables and $Δ$ is the largest absolute value of coefficients of integral variables. This improves the dependence on $m$ in the exponent to asymptotically match the dependency in the integer setting. Further, we use our algorithm to generalize the parameterized algorithm for scheduling on uniformly related machines by Rohwedder~[ICALP 25] to obtain an algorithm for multiple knapsack with running time $w_{\max}^{O(d)}\text{poly}(|I|)$ where $w_{\max}$ is the maximum weight of an item and $d$ the number of different weights.

To keep long interaction contexts tractable, stepwise group-relative training for multi-turn LLM agents materializes each source trajectory into multiple response rows and valid tokens. This expansion creates an outer-accounting problem distinct from within-trajectory credit assignment: row-based outcome statistics can count one episode outcome multiple times, while token-level reduction can assign unequal aggregate objective mass across trajectories. We term these effects outcome-statistics multiplicity and objective-mass imbalance. We introduce Outcome-Mass Alignment (OMA), a two-component method that makes both accounting choices explicit. Outcome Alignment computes outcome statistics over unique source trajectories, while Mass Alignment controls aggregate objective mass after within-trajectory aggregation of optimizer-native token objectives. Neither redesigns the base optimizer's credit assignment, importance ratios, or clipping. Across matched ALFWorld GRPO comparisons at two model scales, where failed trajectories typically expand much more than successful ones, OMA improves matched-horizon endpoint success by 7.0-9.4 percentage points. Ordinary WebShop exhibits weak or reversed expansion-outcome coupling and correspondingly mixed effects, whereas in a success-oriented WebShop setting, WebShop-SOL, OMA improves success nAUC from 0.399 to 0.511. Component ablations and GiGPO and GSPO comparisons further support component separability and cross-optimizer compatibility. These results show that trajectory expansion always defines an implicit weighting rule, but its practical effect depends on how expansion couples to outcomes and optimizer-native updates.

Parameterised Boolean Equations Systems (PBESs) are used to verify properties of formal industrial models. We introduce a technique called selective chaining to simplify these PBESs by unfolding specific predicate variable instances when the unfolding results in a PBES that is likely to be simpler. For determining the latter we propose a set of heuristics. We show the formal correctness of the chaining operator and provide theoretical upper bounds using relevancy graphs on the size of the verification problem after chaining. Lastly, we study the effectiveness of our heuristics, showing that selective chaining can reduce memory usage and yield significant speed-ups solving PBESs that encode model checking problems for complex industrial systems.

Reward serves as the primary learning signal in reinforcement learning (RL). However, while reward magnitudes are typically held fixed throughout training, their temporal modulation remains underexplored. In this paper, we propose reward inflation, a gradual scaling of rewards over the course of training, and show that it can act as a healthy stimulus for RL. Theoretically, reward inflation induces an implicit recency weighting that upweights recent transitions during policy updates, enabling faster adaptation. We further show that reward inflation suppresses the emergence of dormant neurons by sustaining gradient signals as the policy saturates, suggesting a possible contribution to plasticity preservation. Empirical results on ALE games and MuJoCo tasks corroborate these findings, showing that an appropriate level of reward inflation benefits a broad range of tasks. Finally, we introduce Fed, an adaptive variant that adjusts the inflation level on the fly, and find that it often improves upon fixed inflation. Source code is available at https://github.com/northadventure/RewardInflation.

We present a deterministic algorithm for 3-SAT running in $1.32736^n$ time. The algorithm is built on a branching-and-chain framework augmented with a polynomial-time expansion procedure that joins short chains through additional clauses to form bounded-size blocks, which impose stronger constraints on the assignment space. After branching, the algorithm chooses between derandomized local search guided by the block structure and enumeration of block assignments together with selected port variables, reducing the remaining formula to 2-SAT. A weighted analysis of branching, partial-assignment, and local-search costs, together with inductive bounds for chains and block interactions, establishes the running-time guarantee. The framework is self-contained and also offers guidance for practical SAT solver design by identifying useful clause-overlap structures and suggesting when to switch between local-search and reduction-based strategies.

Generative artificial intelligence (GenAI) is now commonplace in higher education, yet unguided use can harm learning. Universities have responded with policies asking students to use GenAI for learning, not solutions, but how to translate these policies into classroom instruction remains untested. Here we report a randomized controlled trial with a six-week, end-of-semester follow-up (N = 431) in an introductory programming course comparing four instructional approaches, from a two-sentence reminder (Condition 1) to a multi-step constructive task (Condition 4). We found: (1) all four approaches produced learning-oriented prompting gains that persisted six weeks later and scaled with the designed engagement level, from 8 to 51 percentage points; (2) students in every condition self-reported sustained voluntary use, and confidence grew in step with skill; and (3) final examination performance did not differ between conditions, placing the possible boundary of the effects at the pathway from prompting skill to achievement. These findings provide instructors with an empirically calibrated set of options for translating GenAI-use policies into classroom practice.

Many failures of language models begin at a specific token, the onset. A monitor that reads the model during generation can classify the finished response, detect the failure once it is under way, or warn before it begins, and only a warning leaves time to stop or steer the generation. Monitors are usually evaluated with response-level metrics such as AUROC, which ignore when the monitor fires. We introduce PREWARN, an evaluation protocol for any failure with an annotated onset. It calibrates each monitor's threshold on validation data to an error rate chosen for the application, runs the monitor, and scores its alarms relative to the onset on all three tasks. As a case study we apply it to repetitive degeneration in three open language models, with early warning as the main task, using linear probes on activations and six untrained baselines. Response-level metrics saturate on this failure, and a rule that reads only the response length scores almost perfectly, while probes that AUROC ranks as equal warn very differently under PREWARN. Probes warn before the onset on most degenerate responses of two of the three models, low-rank adapters below the probe more than double the share of tokens just before the onset that it flags, largely by flagging degenerate responses long before the onset, and no untrained baseline warns much more often than random alarms. We release the probes with calibrated thresholds at https://huggingface.co/collections/marcodene/prewarn-degeneration-probes

The Third Generation Partnership Project (3GPP) Technical Specification 23.501 (Release 17) defines Maximum Bit Rate (MBR) enforcement at the User Plane Function (UPF) in fifth-generation (5G) networks. Supporting differentiated service-level agreements across the radio access and core networks requires the UPF to enforce MBR limits according to configured slice priorities. During overload, the challenge is to adapt these limits while preserving that differentiation and avoiding oscillations. Building on programmable enforcement and intelligent orchestration, we propose an agentic control loop combining UPF telemetry with locally hosted Large Language Model (LLM) selection from a finite set of admissible rate candidates. Deterministic logic validates each selection and derives Packet Forwarding Control Protocol (PFCP) updates. The evaluated runs show no oscillations or fallbacks and unchanged highest-priority rate limits. These observations establish configuration preservation; delivered per-class QoS protection remains unmeasured.

The transparency regime of recommender systems on Very Large Online Platforms (VLOPs) is undergoing significant change under the EU Digital Services Act (DSA). The DSA requires platforms to disclose information about their recommendation pipelines and grants vetted researchers access to non-public data, which may extend to recommender code. However, it is not known to what extent such code releases enable substantive and meaningful third-party auditing of a production recommender system. We address this question by proposing a code-level audit methodology for recommender systems and applying it to the X platform, which has voluntarily open-sourced its recommender pipeline. Our methodology is a backwards-walk traversal of the repository, tracing each decision-relevant component from the user-facing timeline back to its upstream producers, checking that it is reachable from the served pipeline, and classifying each under a four-tier inspectability scheme. Applying this procedure, we identify 13 transparency dead-ends on the live serving path: models, services, label lists, and configuration values that the public code depends on but the release does not contain. We find that the components that decide which posts are filtered and how they are ranked fall into the two least inspectable tiers, a pattern that intensifies with the 2025 integration of the proprietary Grok language model into moderation, annotation, and ranking. We further show that the release keeps 2023 code that the 2025 pipeline no longer reaches, with 36 parameter references that would pass for dead-ends without a reachability check. Finally, we discuss which additional artifacts would be required for meaningful auditing.

We study the impact of a Hertzian sphere on an infinite Rayleigh beam, which presents both interesting possibilities for numerical approximation as well as new insights in the multiple-contact-event regime. Here, Galerkin projection using an abstract continuous variable replaces the history-dependent convolution in the impact equations by a system of local-in-time ordinary differential equations. The corresponding approximate solutions are validated against direct numerical solution of the convolution-based formulation and show agreement. Unlike the corresponding Euler-Bernoulli beam impact problem, which is governed by a single nondimensional parameter, the Rayleigh beam impact dynamics is governed by two nondimensional parameters. A parametric study reveals three distinct impact regimes: one regime with a single contact followed by permanent separation, a second regime with a single contact that is sustained thereafter, and an intervening regime with initial separation followed by multiple contact events with diminishing amplitudes, leading eventually to sustained contact. Short- and long-time asymptotic analyses are used to characterize the boundaries between the permanent separation and the re-contact regimes in different parameter regimes. The three significant new aspects of this paper are the system itself (Rayleigh beam instead of Euler Bernoulli beam), the model order reduction achieved through Galerkin projection in an abstract domain, and the detailed parametric study leading to a simple qualitative picture of possible behaviours.

Vision-language-action (VLA) models are commonly adapted to downstream manipulation tasks via supervised fine-tuning (SFT) or online reinforcement learning(RL) post-training. SFT is prone to distribution mismatch, and existing RL approaches typically optimize all trainable components within a single RL loop at a uniform update rate, although these components play distinct functional roles. We propose TEMPO, a semantic-action decoupled, two-timescale RL post-training framework for VLA models. TEMPO freezes the pretrained vision-language backbone and adapts two components with dedicated RL loops: the semantic projection layer, which produces the latent action conditioning the action expert, and the low-level action expert itself. The semantic projection layer is updated infrequently and the action expert frequently, which aims to limit latent-action drift while allowing the action expert to learn from online feedback. Our analyses show that, compared with equal update frequencies, less frequent semantic updates reduce latent-action drift and improve policy performance. Experiments on the CALVIN benchmark and real-world manipulation tasks demonstrate that TEMPO outperforms both pretrained state-of-the-art VLA models and the RL post-training baseline, while reaching and maintaining higher evaluation rewards on three real-world tasks.

Recommender systems, which retrieve items of interest to users based on their interaction histories, have recently witnessed several key technological advancements, benefiting from the semantic representation space and model-data scaling. Nevertheless, practitioners and researchers have frequently encountered critical inconsistencies between simple offline evaluations and complex online results, and this gap still exists in the era of large models. In this work, we address several key factors that contribute to this inconsistency: 1) the \textit{large-scale continual learning challenge} that assumes a practical demand to maintain long-term model effectiveness under dynamic item pools and distributional shifts; 2) the \textit{combinatorial diversity challenge} that requires the generation strategy to solve the accuracy-diversity trade-off in both the finer token level and the holistic item level; and 3) the \textit{multi-task reward balancing challenge}, which is particularly critical for modern generative models that can accurately model the sequence distribution but are less controllable for the multi-task demands. To address these challenges and further pave the way for the development of more realistic recommendation solutions, we introduce KuaiRecV2, a deliberately constructed benchmark with industrial-level datasets from Kuaishou. Specifically, we provide a large-scale interaction dataset with million-scale multi-modal item information and long-term interaction records across 30 days, with a dynamic video candidate pool. Additionally, we provide a rigorous benchmark that evaluates retrieval accuracy, diversity, and multi-task performance in one rubric system, where an effective advancement occurs if and only if the new method improves all metrics. The code is available at https://github.com/cxli2333/KuaiRecV2