Analysis of PDEs
Generalized minimizers for the capillarity energy outside convex obstacles
Gabriele Nani and Simone Vincini
In general, the capillarity energy outside an unbounded obstacle may fail to admit volume-constrained minimizers, as infimizing sequences may escape to infinity. We study this minimization problem outside a closed convex obstacle $C$. Via a $\Gamma$-convergence result for the capillarity energies outside sequences of closed convex obstacles converging in the Hausdorff sense, we prove that the infimum of the functional is attained by a generalized minimizer, namely, a finite union of regions, each of which minimizes the capillarity energy outside a limit of translations of $C$ in its volume class, with the total volume satisfying the prescribed constraint. Each generalized minimizer can be suitably approximated by an infimizing sequence, thus providing a description of its asymptotic behaviour. As an application, we study the isoperimetric profile associated with the capillarity energy and investigate how the asymptotic geometry of the obstacle affects both this profile and generalized minimizers and their connectedness.
Read the full abstract ▾
Computation and Language
SMI: Ontology-Grounded Semantic IDs for Hierarchy-Aware Medical Concept Representation
Ziyang Song, Lia Shen, Yixuan Li et al. (+6 more)
Biomedical
language models (LMs) provide strong representations for clinical tasks, but they often fail to capture the hierarchical structure among medical concepts. This limitation is especially pronounced for rare and fine-grained clinical codes in long-tailed vocabularies, where hierarchy-agnostic LM representations become less informative as concept frequency decreases. We introduce Semantic Medical Identifier (SMI), a training-free framework for constructing ontology-grounded Semantic IDs and hierarchy-aware medical concept representations. For each diagnosis, procedure, or medication concept, SMI defines its Semantic ID as the coarse-to-fine path through an expert-defined medical ontology. SMI then aggregates frozen biomedical LM
embeddings across the corresponding ontology levels to form hierarchy-aware medical concept representations that preserve coarse-to-fine clinical semantics. The base SMI uses uniform aggregation across ontology levels, while a level-weighted variant selects global level weights on validation data. Our theoretical analysis shows that SMI yields more informative medical concept representations than hierarchy-agnostic LM
embeddings, particularly for rare and fine-grained clinical concepts. Experiments show that SMI captures meaningful hierarchical structure and improves clinical relation prediction. On MIMIC-IV, SMI improves mortality prediction, readmission prediction, drug recommendation, and rare drug prediction. On eICU, SMI yields more robust patient representations under cross-hospital distribution shift.
Read the full abstract ▾
All authors Ziyang Song, Lia Shen, Yixuan Li, Qincheng Lu, Weihao Li, Haoran Xu, He Zhu, Xinyu Wang, Yue Li
High Energy Astrophysical Phenomena
Multimessenger Constraints on Nuclear Acceleration in Neutrino-Emitting Seyfert Coronae
Chengchao Yuan
High-energy
neutrinos from Seyfert
galaxies probe particle acceleration near supermassive black holes, but the accelerated nuclear composition remains uncertain. We connect
stochastic acceleration to nuclear and radiative cascades to assess nuclear contributions. Acceleration and losses limit the energy per nucleon available to nuclear cascades; without secondary reacceleration, fragmentation cannot raise this scale. For NGC~1068, joint
neutrino and gamma-ray constraints depend on the nuclear mixture: helium-dominated injection permits larger nuclear power fractions than the heavier mixtures considered. For the harder candidate {
neutrino} spectrum of NGC~7469, the adopted model yields more energetic nuclear cascades, while the source-specific flux selections allow larger nuclear fractions for the heavier mixtures than in NGC~1068. These results show that composition constraints depend on nuclear yields as well as acceleration conditions and available power.
Read the full abstract ▾
Robotics
Perception Beyond the Visible: Incompleteness-Aware Scene Understanding for Robot Autonomy
Rohit Mohan
Reliable
robot autonomy requires
reasoning about the world from partial and noisy observations. However, many perception systems model only visible scene regions, rely on fixed semantic categories, and assume reliable sensor measurements. These assumptions break under occlusion, novel object encounters, and sensor degradation. In this paper, I present a research
vision for incompleteness-aware scene understanding. I organize this
vision around three fundamental challenges: incomplete evidence, incomplete knowledge, and incomplete sensor reliability. I discuss my research contributions toward addressing these challenges through amodal perception, semantic occupancy forecasting, open-world scene understanding, label-efficient learning, and uncertainty-guided multi-modal fusion. Building on these contributions, I outline future research directions toward belief-centric 4D scene representations. These representations aim to jointly model geometry, semantics, motion, and uncertainty. I also discuss active sensing and safe adaptation strategies that allow
robots to reduce uncertainty through interaction. The goal is to develop calibrated and actionable scene representations that support reliable autonomy in dynamic, open-world environments.
Read the full abstract ▾
Computation and Language
Known-Voice Cloning and the Limits of Personality Transfer for Character Design
Johanna Magdalena Kuch, Maud van Lier, Daksitha Senel Withanage Don et al. (+3 more)
Recent advances in zero-shot voice cloning systems make it possible to reproduce a recognizable voice from a short recording and attach it to arbitrary virtual characters. Prior work suggests that familiar vocal identity can carry over associated personality impressions to new targets such as
robots or virtual characters. However, these works have mainly examined familiarity and personalization effects rather than directly testing whether a cloned familiar voice systematically transfers personality impressions to an embodied virtual character. Such transference raises the concern that cloned voices could be used to manipulate people into engaging with virtual characters as capable social actors. Across two studies, we investigated whether personality impressions transfer when a virtual character speaks with the cloned voice of a familiar person. Study 1 examined this question in a real-world university setting using the cloned voice of a lecturer familiar to the participating students. Study 2 examined the same question online using voices of social media content creators and members of their communities. Drawing on the concepts of identity activation, trait transfer, and person reconstruction, we show that evidence of personality transfer was limited in both studies. Observed effects were selective and did not generalize across traits or study contexts. In the online study in particular, recognizable voice identity cues activated expectations about the speaker but did not reconstruct the broader person impression. Qualitative responses further indicated that judgments were frequently shaped by perceived artificiality and by whether voice and appearance fit together. Our findings moderate common concerns that cloned voices straightforwardly manipulate people into perceiving artificial characters as capable social actors.
Read the full abstract ▾
All authors Johanna Magdalena Kuch, Maud van Lier, Daksitha Senel Withanage Don, Patrick Gebhard, Tanja Schneeberger, Elisabeth André
Social and Information Networks
When Platforms Pay Users to Correct Misinformation: Revenue Incentives and Governance Capacity in Weibo Annotations
Chichen Lin, Huaijin Xie, Wenhan Xie et al. (+5 more)
Crowdsourced misinformation governance systems such as Community Notes and Weibo Annotations depend on sustained corrective labor. How monetary incentives shape participation and subsequent governance outcomes remains poorly understood. We conduct a mixed-methods study of Weibo Annotations, analyzing 102,994 posts and 431,407 comments from 2023–2026, alongside interviews with 15 contributors. We find concentrated corrective labor, limited observed display, and weak quality–engagement
alignment. Policy introduction was associated with sharp increases in contributor participation and annotation production. Model-assisted quality followed its expected trajectory, evidence for additional display growth was inconclusive, and engagement per annotation fell below expectation. Public discussions and interviews reveal substantial verification demands, difficult-to-interpret review processes, uneven platform knowledge, and gaps in contributor protection. We identify four recurring bottlenecks in governance orchestration: participation, review, reach, and protection. Effective incentive design therefore depends on whether platforms can support these downstream processes as corrective labor expands.
Read the full abstract ▾
All authors Chichen Lin, Huaijin Xie, Wenhan Xie, Weijian Fan, Xiao Han, Yijie Jin, Xiaojie Hao, Jiaying Wu
Dynamical Systems
Analytic centralisers of quadratic germs and the Brjuno dichotomy
Wen Si
For $Q_\alpha(z)=e^{2\pi i\alpha}z+z^2$ with $\alpha$ irrational, we prove that the analytic centraliser of $Q_\alpha$ is trivial whenever $Q_\alpha$ is not analytically linearizable. Together with the quadratic Brjuno–Yoccoz
theorem this gives a complete dichotomy, for the centraliser is infinite cyclic at the non-Brjuno parameters and conjugate to the group of nonzero complex dilations at the Brjuno parameters. The result extends the dense-parameter conclusion of Avila–Cheraghi–Eliad to every nonlinearizable member of the quadratic family. The proof starts from Inou's graph-globalization construction, where we retain the original neutral fixed point on a Riemann surface and verify the branch
topology and finite-valence estimate, including multiplicities. Nonlinearizability, through the equality case of Schwarz's lemma, forces the surface to be the plane, and a direct quadratic growth estimate then makes the two entire projections
polynomial, so that the established
polynomial commutant classification yields the integral-iterate conclusion. We also determine the convergence set of the formal commutants and classify the analytic iterative roots of every nonzero iterate.
Read the full abstract ▾
Combinatorics
Proofs of four Takigiku–Tsuchioka conjectures on level $2$ principal characters of $A_{2\ell-1}^{(2)}$
Liuquan Wang
We prove four
conjectures of Takigiku and Tsuchioka consisting of eight Rogers–Ramanujan type identities. Two
conjectures involve triple sums and a third involves quadruple sums, and their product sides are certain level $2$ principal characters of $A_{13}^{(2)}$. The fourth
conjecture gives multisum representations of certain level $2$ principal characters of $A_{2\ell-1}^{(2)}$. The proofs rely on the Bailey pair method, finite sum identities involving Gaussian
polynomials, and the quintuple product identity.
Read the full abstract ▾
Materials Science
Composition-controlled optical fingerprints and anisotropy of self-trapped excitons in double perovskites
Simone Grillo, Offir Zachs Maishlos, Liya Koren et al. (+7 more)
Double perovskites (DPs) offer a compositionally versatile platform for photocatalysis and optoelectronics, but their soft
lattices favor strong
exciton-
phonon coupling and self-trapped
excitons (STEs), which produce broad, red-shifted photoluminescence (PL). Here, we combine optical spectroscopy and ab initio calculations to study lead-free, Sb-doped Cs2BInCl6 nanocrystals with B-site alloying from Na to K. Alloying tunes the PL energy from about 2.48 to 2.70 eV, while absorption features near 3.8 eV remain nearly unchanged. Calculations reveal that Sb doping enables bright radiative recombination channels and delocalizes the excitonic wavefunction. We identify STEs responsible for the large PL red shift, associated with an upward shift of dopant-hybridized in-gap states. Na-to-K substitution selectively changes the STE
lattice distortion, explaining why PL varies strongly with composition while absorption does not. Theory further predicts pronounced PL polarization anisotropy that reflects the STE wavefunction's spatial structure and depends on B-site composition. Time-resolved anisotropy measurements closely agree with these predictions, providing experimental fingerprints of the excitonic wavefunction and its
evolution with alloying. Together, these results disentangle the roles of doping and alloying in lead-free DPs and establish polarization-resolved spectroscopy as a probe of STE wavefunctions.
Read the full abstract ▾
All authors Simone Grillo, Offir Zachs Maishlos, Liya Koren, Rotem Strassberg, Rachel Lifer, Edi Maizenberg, Umberto De Giovannini, Angel Rubio, Yehonadav Bekenstein, Ofer Neufeld
Computer Vision
PlanPath: Exploration-Guided Orchestration for Patient-Level Pathology Reasoning
Jiusong Ge, Mingzhao He, Qidong Liu et al. (+8 more)
A patient's pathology case spans tens of slides from distinct specimens and anatomical sites, each carrying non-redundant, site-specific evidence. Forming a patient-level conclusion therefore requires locating the slides relevant to a given question and integrating complementary findings across them. However, existing pathology
agents analyze one slide at a time and describe morphology rather than measure it, leaving evidence acquisition across a patient's specimens and patient-level quantification underexplored. To address this gap, we propose PlanPath, an agentic framework for patient-level pathology visual question answering (VQA). PlanPath turns the question into an exploration plan that specifies the evidence required, routes to the slides likely to carry it, and drives coarse-to-fine acquisition from tissue architecture down to individual
cells. Tissue- and
cell-level tools supply the counts, proportions, and extents, which PlanPath aggregates across slides into a traceable answer. To enable evaluation, we build two patient-level VQA
datasets for renal and gastric cancer, averaging 15 WSIs per patient, with pathologist-verified question-answer pairs derived from diagnostic reports. Experimental results show that PlanPath outperforms the strongest baseline by 9.4% and 12.1% in overall accuracy on the two
datasets and further improves existing slide-level
agents when plugged into our framework. The
dataset and code are available at: https://github.com/JSGe-AI/PlanPath_1.
Read the full abstract ▾
All authors Jiusong Ge, Mingzhao He, Qidong Liu, Chunbao Wang, Chunze Yang, Di Zhang, Jiashuai Liu, Mireia Crispin-Ortuzar, Huazhu Fu, Chen Li, Zeyu Gao
Fluid Dynamics
Some generalizations of the tornado model
S. N. Artekha
A possible mechanism of tornado generation is associated with the Schwarzschild criterion for convective instability. The supercooled cloud-air mixture near the axis starts to descend and, due to the super adiabatic temperature gradient, the overheated near-surface air mass rises. The complete system of equations for full model includes equations for the stream function, for the azimuthal velocity component, and for the change in density. The refined model considers rotating convection within the framework of ideal hydrodynamics. The simplest model uses a separate equation for the stream function and can be applied to the initial stage of
vortex generation. Within the framework of the simplest model, an analytical solution is combined from 8 linear dependencies by height. Rotation increases due to the conservation of angular momentum. Within the framework of the refined model, some unidirectional differential rotation arises, but it depends on boundary conditions (can be local). The full model studies the tornado development as rotating convection from the very beginning to the quasi-stationary stage. Velocity components increase in a coordinated manner. Two cases are considered: a single whirlwind and a two-eyed tornado. Radial flows converge towards the central part of the tornado near the earth surface, and there is an outflow from the "wall" to the axis and to the periphery near the top of the tornado. The maximum vertical velocity component is achieved at a certain radial distance at a certain height. The maximum azimuthal velocity component is achieved in the tornado "wall". Tornado generation criteria are discussed.
Read the full abstract ▾
Differential Geometry
Odd-dimensional lens-space Yamabe counterexamples
Jakob Loren
Let $n\ge5$ be odd and let \(L_{3q}^n=S^n/C_{3q}\) be the standard lens space, where $q>3$ is
prime. We prove that for all sufficiently large $q$, \[ \sigma(L_{3q}^n) > (3q)^{-2/n}Y(S^n)=Y(L_{3q}^n,[g_{\rm rd}]). \] Here \([g_{\rm rd}]\) denotes the quotient-round conformal class. Combining
spin bordism with the surgery
theorem of Ammann–Dahl–Humbert, we establish \[ \sigma(L_{3q}^n)\ge \min \{q^{-2/n}Y(S^n),\Lambda_n\}, \] where $\Lambda_n>0$ depends only on $n$. For sufficiently large $q$, this exceeds the quotient-round value, showing that the quotient-round conformal class need not maximize the smooth Yamabe invariant in any odd dimension $n\ge5$.
Read the full abstract ▾
Mesoscale and Nanoscale Physics
Dark state preparation in a zero-field-split molecular triplet
Maria Pfahler and Andrea Donarini
We propose a protocol to prepare and probe dark states in a single-molecule electron
spin resonance based on atomic force microscopy (ESR-AFM), thus extending existing techniques developed to address Rabi oscillations. To this end, we simulate in analogy to atomic physics the coherent manipulation of a three-level system in the V-configuration using a multifrequency
magnetic field. More specifically, we consider a neutral molecule with a zero-field-split (ZFS) first excited triplet and analyze within a pump-probe scheme the dynamics of the total triplet population. In the triplet subspace, the dark state is a coherent superposition formed by two ZFS states which does not evolve in time despite a continuous resonant driving towards the third one. We characterize the properties of this state comparing it to its fully incoherent counterpart. Within a density-
matrix formalism for open systems, we then compute the expected dark state signatures both for ideal scenarios and for the realistic parameterization of ESR-AFM on pentacene.
Read the full abstract ▾
Cryptography and Security
ARACDE: Retrofitting Databases with ABAC using External Authorization Engines
Karthikeya S M Yelisetty, Shamik Sural and Santosh Nagarakatte
This paper proposes a methodology for decoupling the authorization logic of a relational database from its core data management engine, enabling specification of advanced access control policies with minimal changes to the code base. Our decoupled authorization architecture, which we call ARCADE, externalizes access request resolution to a formally verifiable policy engine (e.g., Cedar), while the database itself enforces those decisions. We showcase the capabilities of ARCADE on two widely popular open-source databases, namely MySQL and PostgreSQL. Purpose built lightweight plugins add an Attribute-based Access Control (ABAC) layer that augments the native discretionary and role-based access control capabilities of these databases. ARCADE supports highly expressive policies using attributes of subjects, objects, and the environment, and the use of Cedar enables easy auditing and verification of the specified policies. It follows a design principle that mitigates potential risks in segregating authorization from the database core through low footprint, network isolation, and a policy composition methodology that does not negate any native authorization. Experiments on TPC-C, TPC-B, and a custom workload show that ARCADE incurs negligible performance overhead, for instance, less than 3.9% throughput penalty for MySQL (1.3% for PostgreSQL) on industry standard TPC-C
benchmark.
Read the full abstract ▾
Astrophysics of Galaxies
The Connection Between the Growth of Massive Black Holes and Their Dynamical Histories
Lee A. Whitaker, Michael Tremmel, Angelo Ricarte et al. (+1 more)
Massive black holes (MBHs) grow through both gas
accretion and mergers and it remains unclear how variations to MBH growth and dynamical histories contribute to the spread in empirical scaling relations between MBHs and their host
galaxies. We examine the ex- and in-situ growth of MBHs within large-scale cosmological simulations and compare the
evolution of an outlier population of under-grown MBHs to the general population. We use the Romulus25 uniform volume cosmological simulation and the RomulusC cluster zoom-in simulation to examine the growth and merger histories of MBHs. We identify central MBHs that are undergrown relative to other MBHs in similar mass
galaxies and investigate their
accretion, merger, and orbital histories within the simulations. The contribution of ex-situ growth increases with MBH mass, with mergers becoming increasingly important in the most massive
galaxies. We find no difference between cluster and field environments in terms of the importance of ex-situ growth, which is generally very low. Under-grown MBHs have experienced fewer mergers and have had complex dynamical histories, spending significant time 'drifting' within the
galaxy or 'competing' with another MBH. Other under-grown MBHs have experienced a 'late entrance' into a
galaxy which until then had no central MBH. Our results demonstrate that MBH assembly is dominated by in-situ growth, which in turn is impacted by the dynamical histories of MBHs within assembling
galaxies. Prolonged off-centre
evolution can suppress
accretion and contribute to the formation of under-grown central MBHs at $z=0$.
Read the full abstract ▾
All authors Lee A. Whitaker, Michael Tremmel, Angelo Ricarte, Priyamvada Natarajan
Machine Learning
BridgeTwist: Twisting Schrödinger Bridges for Training-Free Conditional Sampling
Mengyu Li, Qianqian Qu and Jun S. Liu
Generative models are often required to serve many downstream conditional tasks. In the training-free setting, a pretrained unconditional model is reused as a prior and combined with a task-specific likelihood to sample from the corresponding posterior. While such training-free samplers have been developed for
diffusion and flow models, their counterpart for Schrödinger bridge (SB) models remains unexplored, despite the growing appeal of SBs as flexible
stochastic transport models between general source and target distributions. We propose BridgeTwist, a training-free conditional sampler for pretrained SB generative models, formulated as a tempered twisted sequential Monte Carlo scheme. Its twist is a closed-form plug-in surrogate, obtained by combining the velocity and score fields exposed by SB pretraining into an endpoint predictor and evaluating the likelihood there. Since this surrogate is less reliable near the source, we temper the twist with a time-varying exponent, downweighting its influence at early steps. We further introduce a bridge-shaped inference-time
diffusion schedule that concentrates the proposal noise in the interior of the path. Telescoping importance weights then cancel all intermediate twist factors, yielding an asymptotically exact sampler for the posterior. Empirically, BridgeTwist outperforms training-free baselines on class-conditional and text-prompted image generation, and is competitive with specialized inverse problem solvers on inpainting, deblurring, and super-resolution.
Read the full abstract ▾
Dynamical Systems
Multidimensional Szemerédi Theorem for Pairwise Independent Jointly Intersective Polynomials
Nikos Frantzikinakis and Davide Sclosa
We develop a new approach to the Joint Intersectivity
Conjecture of Bergelson, Leibman, and Lesigne for commuting transformations. The
conjecture asserts that the necessary divisibility condition is the only obstruction to
polynomial multiple recurrence and, consequently, to the multidimensional
polynomial Szemerédi
theorem. We establish the
conjecture for pairwise independent
polynomials, obtaining the first general result for sequences of arbitrary length beyond the classical one-transformation setting. Our argument is inspired by the Furstenberg-Katznelson approach and its
polynomial enhancement due to Bergelson-Leibman. Its main novelty, however, lies in the development of new techniques that allow us to bypass the use of the
polynomial van der Waerden
theorem, on which these methods traditionally rely, and for which no suitable analogue is currently available in our setting. In particular, we prove a Nil-Bohr returns
theorem for compact extensions in systems with nilpotent structure, as well as a nil-plus-null decomposition for
polynomial multicorrelation sequences that preserves joint intersectivity and positivity.
Read the full abstract ▾
Machine Learning
Bayesian Test-Time Inference of Task-Aligned Similarity from Weak Interactive Feedback
Weida Liang and Kenji Kawaguchi
Language model systems repeatedly make test-time decisions by comparing similarity: they retrieve exemplars, rank candidates, and revise responses. Yet similarity is not unique: lexical, semantic, structural, and stylistic resemblance can point to different candidates, and the task-aligned notion can change across instances. We introduce BASIS (
Bayesian Axis Selection via Interactive Signals), a training-free
Bayesian method that uses weak test-time feedback to infer which criterion in a supplied, task-specific library governs the current decision. It maintains a joint posterior over candidate targets and criteria, updates it from scalar scores, pairwise preferences, or verifier outputs, and selects probes by expected information gain. In a finite noisy model, we characterize when the target-criterion pair is recoverable and show why fixed-criterion and fixed-mixture rules can fail under criterion shift. Across concept recovery,
reasoning exemplar selection, and response refinement, BASIS improves six-probe concept-recovery Success@1 from 56.0% to 68.1%,
reasoning accuracy from 72.3% to 73.1%, and refinement win rate from 61.4% to 63.8% over tuned static mixtures. It also exceeds compute-matched best-of-N selection on
reasoning (72.1%) and refinement (61.2%). Gains are largest under criterion mismatch and with larger candidate pools, and matched controls indicate that adaptive criterion weighting explains much of the benefit, with joint target-criterion inference adding further value. BASIS selects only among supplied criteria and does not discover new ones.
Read the full abstract ▾
Computer Vision
Texture Interpolation as Texton Matching
Ido Madar, Gal Oren and Shai Avidan
Texture interpolation asks what should lie between two textures. Standard methods typically blend pixels or features at corresponding spatial locations, implicitly assuming that co-located elements should be mixed. For unrelated textures, however, this correspondence is arbitrary: a crack at one location has no particular relationship to a fiber at the same location in another image. We study whether this hidden pairing choice is itself a major determinant of interpolation quality. We represent textures as collections of local appearance vectors, or textons, and replace positional pairing with appearance-aware matching while keeping the encoder, representation, and decoder fixed. This isolates the effect of correspondence from changes in model capacity or training. Across texture pairs, appearance-aware matching consistently improves the distributional quality of interpolated midpoints over positional feature blending. Soft matching produces the strongest midpoint statistics, whereas one-to-one assignments yield more balanced transitions and closer agreement with the target endpoint. The same intervention also improves a second pretrained texture interpolator without retraining. These results identify correspondence as a first-class design choice in texture interpolation: meaningful transitions depend not only on how strongly two textures are mixed, but also on which local appearances are mixed with one another.
Read the full abstract ▾
Machine Learning
Kwai Summary Attention V2 Technical Report
OneRec Team, Chenglong Chu, Guohong Mu et al. (+30 more)
Summary-
token architectures such as Kwai Summary
Attention (KSA) compress distant context into learnable Summary
token per chunk, and hybrid variants interleave these layers with full
attention to retain exact access to the original text. In such hybrid KSA-Full architecture, the interleaving inserted summary
tokens are served as a streaming flow role across different layers. Although effective, however, it also introduce some bottlenecks: (1) the different layers summary compression chunk must be same, therefore hard to reach a flexible KV Cache compression, e.g., different layer set different summary compression chunk window. (2) in the full-
attention layers, the summary
token introduced additional computation cost, but such layer ignore them and still pay quadratic prefilling, therefore unfriendly to the computation side. Motivated by them, we developed the KSA V2, which not only achieves a more flexible KV cache
memory compression but also reduces the computation cost in Full
attention layer. Specifically, we found that the summary
tokens can serve two roles: the KV
memory and sparse indexer.
Read the full abstract ▾
All authors OneRec Team, Chenglong Chu, Guohong Mu, Guowang Zhang, Hao Peng, Hui Wang, Jian Liang, Jiangxia Cao, Jiao Ou, Jiaxin Deng, Lu Ren, Muhao Wei, Ruiming Tang, Tianle Xu, Wenwu Ou, Xingmei Wang, Zhixin Ling, Zhiyuan Liang, Ziqi Wang, Chaoyi Ma, Haixiao Gao, Kuo Cai, Renpu Liu, Ruochen Yang, Wenlong Xu, Xinqi Jin, Yan Sun, Yifan Wang, Yingzhi He, Yiyu Wang, Yusen Huo, Zhaojie Liu, Zihan Zhou
Statistical Mechanics
Extending the mean first-passage-time formalism to the reconstruction of high free-energy barriers: A top-start approach
Ying Jiang and Ivan Saika-Voivod
Rare events, such as
crystal nucleation in mildly supercooled
liquids, are characterized by waiting times that can be much longer than the timescale of the underlying microscopic dynamics. The mean first-passage time (MFPT) formalism [J. Wedekind and D. Reguera, J. Phys. Chem. B 112, 11060 (2008)] provides a route to reconstruct one-dimensional free-energy profiles from the MFPT and steady-state distribution measured along a reaction coordinate, but conventional implementations become inefficient for sufficiently high barriers because spontaneous crossings are rarely observed. We introduce a "top-start" extension that initiates trajectories near the barrier top and reconstructs the free-energy profile from their subsequent
evolution away from the barrier toward lower-free-energy regions. We establish the equivalence between the top-start and two-region constructions using a Markov-chain transition-
matrix formulation. We validate the method using one-dimensional random-walk simulations across low, medium, and high free-energy barriers, with transition-
matrix calculations providing an exact
benchmark for the corresponding discretized models. The top-start reconstruction remains accurate in the high-barrier regime where conventional MFPT sampling becomes prohibitively inefficient.
Read the full abstract ▾
Sound
AMemOps: Evaluating Memory Updates in Long Audio Streams
Peiwei Chang, Li Zhang, Mark D. Plumbley et al. (+2 more)
Audio
memory must recognise recurring sources across acoustic changes while keeping their attributes up to date. Overall accuracy does not reveal how these demands interact. We introduce Audio
Memory Operations (AMemOps), a
benchmark for latest-attribute
retrieval, preservation of unchanged attributes and absent-source rejection in long audio streams. Paired controls vary recordings, training identities and earlier events to examine their effects on updating. Training on other recordings of the test instruments increases a reference delta-rule
memory's update advantage over matched
attention by 13.4 percentage points. At
retrieval delays of 120\,s, matched histories separate benefits of consistent repetition from costs of conflicting earlier values, while additive-
memory controls identify benefits of erasure.
Read the full abstract ▾
All authors Peiwei Chang, Li Zhang, Mark D. Plumbley, Philip J.B. Jackson, Wenwu Wang
Machine Learning
Does Pretraining Teach Chemical Language Models Chemistry?
Leif Sieben and Yoel Zimmermann
Chemical
language models (CLMs) are pretrained on large corpora of molecules with the aim of learning representations that transfer to diverse downstream tasks. Here, we systematically study the effects of pretraining using a series of such CLMs across 64 molecular property prediction
datasets. We find that the downstream gains of pretraining in CLMs do not scale and performance saturates below gradient-boosting baselines. Across the regimes evaluated, only supervised pretraining on traditional physico-chemical descriptors consistently improves downstream performance for the CLMs tested whereas some forms of pretraining result in models worse than a randomly initialized encoder. While models can resolve molecular structure well, they do so no better than a randomly initialized model, suggesting that, at best, CLMs learn descriptor-like properties in addition to low-level
token statistics during pretraining, but surprisingly little chemistry beyond what classical molecular features already capture.
Read the full abstract ▾
Robotics
ForkCast: Predictive Recovery for Vision-Language-Action Policies via Counterfactual Futures
Chenru Wen, Huajian Zeng, Haokai Ding et al. (+5 more)
Vision-
language-action (VLA) policies perform manipulation tasks using camera images and
language instructions, but object slip or visual occlusion can disrupt execution. Recovery is typically learned during policy training or delegated to a monitor that triggers a scripted repair after detecting a failure. We introduce ForkCast, a predictive recovery framework that compares counterfactual futures to determine whether and how to intervene in a frozen VLA policy. At each decision point, ForkCast evaluates three candidates: executing the policy's nominal action chunk (Commit), adjusting the camera viewpoint (Reobserve), and replanning the arm trajectory (Replan). A candidate-conditioned world model forecasts the latent outcome of each candidate at a common future horizon, and a shared task-progress head scores the predicted outcomes. By evaluating the policy's own continuation alongside the repair candidates, a single comparison based on task progress determines both whether and how to intervene, without a failure detector, an intervention threshold, or updates to the VLA backbone. On five simulated RoboTwin 2.0 tasks with injected occlusion and object slip, ForkCast increases pooled success from 13.3% for the frozen $\pi_{0.5}$ policy to 30.0% and transfers to two additional VLA backbones without updating ForkCast. After adaptation on real-
robot candidate rollouts, ForkCast also improves success over the base policy under both clean and visibility-limited conditions. Code will be released to benefit the research community.
Read the full abstract ▾
All authors Chenru Wen, Huajian Zeng, Haokai Ding, Jian Kang, Anqing Duan, Dawei Zhou, Dezhen Song, Xingxing Zuo
Machine Learning
Tiny Recursive Message Passing
Tommaso Salvatori, Matteo Forasassi, Alexis Toumi et al. (+1 more)
Message passing neural networks have shown remarkable performance in algorithmic and constraint
reasoning, due to their ability to encode the structure of the task directly in the architecture. Recent work has shown that better results can be achieved using recursive
transformers trained with deep supervision and truncated recursions, which make extremely deep loops stable to train, and equipped with test-time procedures such as discrete deduction
lattices. In this work, we apply the same procedures to message passing networks, and show that the two ingredients are complementary, and lead to parameter efficient models: we obtain about $99\%$ test accuracy on Sudoku-Extreme with a model of $16$k parameters that trains in about $10$ minutes on an H100, or in $9$ hours on a CPU. We further show that the cost of truncating gradients grows with the number of rounds information needs to cross the graph, making the choice of wiring a trade-off between robustness to truncation and transfer across problem sizes. We conclude by showing that our models transfer zero-shot to larger instances than those seen during training, outperforming popular neural baselines on N-queens and Boolean satisfiability with a fraction of their parameters.
Read the full abstract ▾
All authors Tommaso Salvatori, Matteo Forasassi, Alexis Toumi, Giovanni de Felice
Astrophysics of Galaxies
Cosmic Evolution Survey: LOFAR 144 MHz imaging of the COSMOS field
E. Vardoulaki, T. Shimwell, M. J. Hardcastle et al. (+18 more)
We present the first deep, targeted LOFAR 144 MHz source catalogue of the COSMOS field, together with the first 144 MHz source counts in this field. The data comprise 48 h of observations. Over the 2.6 sq. deg. region adopted for the catalogue analysis, the final Dutch-array image reaches a median rms sensitivity of ~100 uJy/beam at an angular resolution of 15" x 6". The resulting catalogue contains 1984
radio sources; a peak S/N >= 5 selection is adopted for the resolved/unresolved classification and source-count analysis. We characterised the catalogue using Monte Carlo source-injection and recovery simulations, including dedicated simulations of extended sources, to quantify completeness, flux-density redistribution, and morphology-dependent source-recovery effects. The catalogue reaches 90% completeness at 1.48 mJy. The resulting Euclidean-normalised differential source counts are measured down to ~0.4 mJy and show good agreement with LoTSS DR3 144 MHz source counts and with the VLA-COSMOS 3 GHz counts scaled to 144 MHz, with ~5% of catalogue sources having multiple 3 GHz counterparts within a beam-matched search region. Comparison with the VLA-COSMOS 3 GHz catalogue shows good astrometric agreement, with median offsets of Delta_RA = 0.35" (+0.03"/-0.05") and Delta_Dec = -0.83" (+0.02"/-0.03"), approximately 6% of the corresponding synthesised-beam dimensions, and no significant radial dependence across the field. The resulting catalogue provides a low-frequency view of COSMOS and a reference
dataset for future deeper LOFAR observations and multiwavelength studies of this extensively characterised extragalactic field.
Read the full abstract ▾
All authors E. Vardoulaki, T. Shimwell, M. J. Hardcastle, L. Doukas, L. K. Morabito, C. L. Hale, M. T. Sargent, V. Smolčić, G. Zamorani, G. Gozaliasl, F. An, M. Aravena, A. L. Faisst, M. Vaccari, E. Treister, A. Finoguenov, P. Best, J. S. Kartaltepe, C. Casey, S. Jin, V. Mainieri
Dynamical Systems
Existence and Uniqueness Proof for the Transport Equation and the Orthogonal Dynamics Equation for Zwanzig's Projection on Gaussian Spaces: A Vanishing Viscosity Method
Christoph Widder
We prove the existence of unique solutions ${u\in H^1(0,T;L^2_\gamma(\mathbb{R}^n))\cap L^2(0,T;H^1_\gamma(\mathbb{R}^n))}$ of the transport equation $u_t=F\cdot\nabla u+f$ and the orthogonal dynamics equation ${u_t= Q_\text{Z} (F\cdot\nabla) Q_\text{Z} u+f}$, where $F$ is a time-independent vector field, $f$ is a forcing term, $P_\text{Z}=1-Q_\text{Z}$ is a Zwanzig projection (conditional expectation) on $L^2_\gamma(\mathbb{R}^n)$ defined by $P_\text{Z}h \propto \int h(\cdot ,y) \rho(\cdot,y)\rm{d}y$, and $\gamma$ is the standard Gaussian measure with density $\rho$. To this end, we develop a vanishing viscosity method based on the maximal regularity theory for the Ornstein-Uhlenbeck operator on $L^2_\gamma(\mathbb{R}^n)$. The following assumptions guarantee the existence and uniqueness of strong solutions: ${f\in H^1(0,T;L^2_\gamma(\mathbb{R}^n))\cap L^2(0,T;H^1_\gamma(\mathbb{R}^n))}$, ${g\in H^1_\gamma(\mathbb{R}^n)}$, $F\in W^{1,\infty}(\mathbb{R}^n,\mathbb{R}^n)$ and $\rm{ess\,sup}_{\mathbb{R}^n}\left(x\cdot F-\rm{div}F\right) < \infty$, where $g$ denotes the initial value function.
Read the full abstract ▾
Statistics Theory
Inference for Fréchet Mean Sets
Ryunosuke Miyazaki
We construct confidence sets for the entire Fréchet mean set by inverting studentized pairwise risk inequalities. Gaussian multiplier calibration localized around empirical near minimizers gives asymptotic coverage at least the nominal level, uniformly over distribution classes satisfying empirical-process conditions. The construction requires neither unique ness nor a
manifold structure. We characterize the limiting binding directions and give conditions for exact asymptotic coverage. On the intrinsic circle, exactness holds for every fixed non-point-mass law and uniformly over a family in which two means merge at a quar tic minimum. With ideal multiplier
quantiles, the construction reproduces the Euclidean Wald set under both nonsingular and singular covariance. Under regular local conditions, it is first-order equivalent to intrinsic Wald inference.
Read the full abstract ▾
Computer Vision
ST-GRU PredRNN-V2: A GRU-Based Reformulation of PredRNN-V2 for Video Prediction
Augusta van Haren, Hanna Hoogen, Luca Pattavina et al. (+1 more)
Video prediction requires models to capture complex spatiotemporal dependencies and to retain relevant information over long sequences. PredRNN and its successor PredRNN-V2 address this challenge with spatiotemporal Long Short-Term
Memory (ST-LSTM) units that maintain two separate
memory states, at the cost of a relatively complex recurrent building block. In this work, we investigate whether the simpler gating of Gated Recurrent Units (GRUs) can be transferred to spatiotemporal predictive learning. To this end, we develop a spatiotemporal GRU (ST-GRU) unit as a GRU-based alternative to the ST-LSTM. The ST-GRU replaces the ST-LSTM's three gates per
memory state with two GRU-style gates (update and reset), and merges the temporal
memory state into the hidden state, which reduces the number of parameters by about 5%. We incorporate the ST-GRU into PredRNN-V2 and evaluate the resulting model on the Moving MNIST
video prediction
benchmark, following the original experimental protocol. Averaged over ten predicted frames, the ST-GRU-based PredRNN-V2 achieves an MSE of 55.1, an SSIM of 0.883, an LPIPS of 0.065 and a PSNR of 19.53dB. It does not surpass the original ST-LSTM-based PredRNN-V2 on MSE, SSIM or PSNR, but it achieves a lower LPIPS and performs on par with the earlier ST-LSTM-based PredRNN, while using a simpler recurrent
cell. These results suggest that ST-GRUs are a promising building block for spatiotemporal predictive learning and motivate further work on their design and integration.
Read the full abstract ▾
All authors Augusta van Haren, Hanna Hoogen, Luca Pattavina, Quincy Rommen
Solar and Stellar Astrophysics
Anchoring the Chromospheric Activity$-$Rotation Relation: A Study of the Pleiades and Group X
A. Núñez, M. A. Agüeros, J. L. Curtis et al. (+1 more)
We present a study of the rotation and the chromospheric H$\alpha$ activity of low-mass
stars in the $\approx$120 Myr-old Pleiades and the $\approx$400 Myr-old Group X open clusters. We combine published and new rotation periods ($P_\mathrm{rot}$) from TESS and ZTF, and use LAMOST, SDSS, and MDM spectra to measure H$\alpha$ equivalent widths (EWs). While overall the Group X distributions of $P_\mathrm{rot}$ and EWs are as expected, detailed comparisons to the Pleiades and to the 700-Myr-old
benchmark Praesepe produce some surprises. For F through mid K
stars, the Group X color$-$period sequence lies between those of the Pleiades and of Praesepe, as expected. But the $P_\mathrm{rot}$ for its late K and early M dwarfs and the Praesepe slow-rotator sequence overlap, indicating that the 700 Myr-old
stars have not spun down significantly despite being nearly twice as old as their Group X counterparts. In the Rossby number$-L_\mathrm{H\alpha}/L_\mathrm{bol}$ plane, rapid rotators in Group X are saturated, as is seen in older clusters, whereas the rapid rotators in the Pleiades show a weak but measurable dependence of $L_\mathrm{H\alpha}/L_\mathrm{bol}$ on rotation. This implies that the chromospheric activity$-$rotation relation attains its mature form between 120 and 400 Myr. Dividing the sample by color reveals that K, M0, and M1 dwarfs are no longer saturated by $\approx$120 Myr, M2 and M3 dwarfs become unsaturated between $\approx$400 and $\approx$700 Myr, and later M dwarfs remain largely saturated through field ages even as their saturated $L_\mathrm{H\alpha}/L_\mathrm{bol}$ levels decline, indicating that chromospheric
evolution is not determined by rotation period alone.
Read the full abstract ▾
All authors A. Núñez, M. A. Agüeros, J. L. Curtis, C. A. Ospinal Achicanoy