Papers

Virtual screening aims to prioritize active compounds from large chemical libraries within a limited experimental budget. When applying Boltz-2 to virtual screening, a key challenge is how to use limited experimental data from the target assay to improve the prioritization of active compounds. We investigated whether fine-tuning the Boltz-2 affinity heads with a small number of binary activity labels could improve early enrichment of active compounds in hit discovery. We compared fine-tuning with 40-300 labels in a retrospective evaluation on eight MF-PCBA targets. With 300 activity measurements, fine-tuning increased the number of actives in the top 1% by a geometric mean of 1.77-fold across the eight targets and improved average precision (AP) by 2.14-fold relative to the control without fine-tuning. We also investigated whether rescoring a subset of candidates could retain the improvement in hit recovery by reranking only the top-ranked Boltz-2 candidates with the fine-tuned head. Restricting rescoring to approximately 10% of the evaluation set retained hit recovery comparable to full rescoring. These findings show that affinity-head fine-tuning with limited activity labels improves early enrichment with Boltz-2 and that this benefit can be retained when rescoring a restricted set of candidates.

We prove a decomposition theorem for canonical processes generated by independent symmetric random variables with log-concave tails, without the $\Delta_2$ condition. The proof uses the classical chaining method, based on admissible partitions and chopping maps. The argument also recovers the majorizing-measure characterization.

Reduced-order models (ROMs) are widely used to accelerate the solution of parametrized partial differential equations, with projection-based methods relying on a linear representation of the governing equations in a low-dimensional space. Their efficiency, however, is fundamentally limited by the Kolmogorov N-width of the solution manifold: when the N-width decays slowly, accurate approximations require a large number of linear modes. This limitation has motivated nonlinear ROMs, which typically replace the linear representation of the solution field with nonlinear latent representations learned through autoencoder architectures. In this work, we propose a different perspective by introducing nonlinearity directly into the representation of the governing operators. Rather than compressing only the solution field, we learn nonlinear low-dimensional representations of the full-order operators while retaining an explicit reduced system of equations. Convolutional architectures are employed to exploit the locality and spatial structure of differential operators, and the resulting compressed operators are coupled with a nonlinear representation of the solution through a training strategy that accounts for both reconstruction accuracy and the error of the reduced-order solution. To address large sparse operators arising from fine and unstructured discretizations, continuous convolutions are used to operate directly on their non-zero entries. Finally, radial basis function interpolation enables the prediction of the compressed operators at unseen parameter configurations, providing an efficient online stage without assembling the corresponding full-order system. Numerical experiments on three parametrized problems demonstrate the potential of nonlinear operator compression as an alternative strategy for equation-based ROMs.

Traversing flexible slender structures involves strong bidirectional coupling between vehicle motion and structural vibration caused by a moving contact point. This paper presents a nonlinear modeling-and-control framework for an electric unicycle traversing a tensioned flexible cable. The cable is modeled using a tension-dominated finite-element formulation retaining vertical and lateral transverse dynamics, and is coupled to the unicycle roll and pitch dynamics through moving-contact kinematics. A modal truncation yields a control-oriented reduced-order model that preserves the nonlinear vehicle–cable coupling. Based on this model, a constrained nonlinear model predictive controller is formulated to regulate traversal, balance, actuator limits, and cable-span constraints. To reduce the computational burden associated with moving-contact interpolation, a frozen spatial contact-point interpolation strategy is introduced for the NMPC prediction model. Numerical studies under nominal, disturbed, constrained, and model-mismatch scenarios show accurate traversal, roll and pitch stabilization, attenuation of cable vibrations, and constraint-consistent closed-loop behavior. The frozen-interpolation formulation reduces and regularizes NMPC computation times relative to the non-frozen formulation, supporting its use as a computationally tractable approach for finite-element-based moving-contact vehicle–structure interaction systems.

What limits KV-cache compression at extreme bit-rates? We argue that it is not the choice of compression scheme, but how its budget is allocated across attention heads. Existing methods apply rank and bit-width uniformly, ignoring that each head has a different optimal mix of rank truncation and quantization. We show that co-optimizing rank and bit-width per head, using only standard low-rank projection and scalar quantization, dominates uniform allocation, with the largest gains at low bit-rates. Our method, KV-COBRA (Co-Optimized Bit-Rank Allocation), formalizes this as a resource-allocation problem: it balances rank-truncation loss against quantization loss within each head, then redistributes budget across heads to minimize total distortion. A fused Hadamard rotation equalizes per-channel variance, and reordering the SVD basis by attention-KL importance makes the solver query-aware. The same allocator extends to joint $K{+}V$ compression. On perplexity, zero-shot, and long-context benchmarks from $0.5$ to $4$ bits per dimension (bpd), KV-COBRA shows the smallest accuracy degradation among evaluated methods at low bpd, with no per-token overhead.

We consider a generalisation of the probabilistic reformulation of the supercooled Stefan problem from arXiv:1902.05174 allowing for non-integrable initial data. From an application of Ito's formula, we obtain a relationship between the Laplace transform of the initial data and a transform of the solution, together with its possibly discontinuous free boundary. Using this relationship we give a sharp existence result for global solutions, we extend previous results on the uniqueness of solutions, and we determine the asymptotic behaviour of solutions for a certain class of initial data.

The disaggregation and softwarization introduced by the Open Radio Access Network (O-RAN) architecture enable multi-vendor innovation but also expose the RAN Intelligent Controller (RIC) ecosystem to new runtime security risks. Existing O-RAN specifications define strong safeguards for onboarding, authentication, identity management, and secure communication; however, they do not provide a concrete mechanism for verifying whether deployed rApps and xApps remain in their intended, untampered state during operation. This paper introduces rApp/xApp attestation as a RIC-native O-RAN security use case for runtime integrity verification. Rather than proposing a new cryptographic protocol, the work defines how existing integrity verification techniques can be integrated into O-RAN through attestation modules, attestation agents, RIC application interfaces, and SMO-driven policy coordination. We map the use case to relevant O-RAN Alliance working groups, identify required standardization extensions, and demonstrate feasibility through a lightweight hash-based prototype implemented on the Near-RT RIC platform. Experimental results show attestation latencies below 40 ms across multiple cryptographic hash functions, indicating that runtime attestation can be performed without disrupting time-sensitive RIC operations when appropriately scheduled. Finally, we discuss remaining technical and standardization challenges, including trusted verification, known-good runtime states, scalability, mitigation policies, and future hybrid attestation mechanisms.

Annealing a semiconductor makes its vacancies mobile; they aggregate into multi-vacancy complexes that often carry spin. Such centres are a magnetic-noise source for any spin qubit among them, in silicon and silicon carbide as well as in diamond, and they are accordingly blamed for the decoherence of negatively charged nitrogen-vacancy (NV$^-$) centres in irradiated and implanted diamond, with two coherence records credited to removing them. Cluster-correlation-expansion simulations in which the spins outside each cluster keep exchanging spin projections in pairs reproduce the measured Hahn-echo coherence time of NV$^-$ centres in a substitutional-nitrogen bath to within 12%, inside the statistical uncertainty. Driven by published electron-paramagnetic-resonance parameters, the same calculation inverts the attribution. At a fixed paramagnetic spin density a multi-vacancy bath gives a Hahn-echo coherence time 3.0-6.1 times longer than a bath of isolated negative vacancies at 1 ppm, and 2.2-6.5 times longer from 0.01 to 10 ppm. A bath dephases the qubit because its spins exchange spin projections with one another, and two can exchange only if their transition frequencies match. A fine-structure splitting shifts those frequencies. Were every defect on the same crystallographic site, all would shift alike and still match, and the splitting would gain little, a factor 1.2. Real defects occupy symmetry-equivalent sites pointing in different directions, so neighbours land at different frequencies and stop exchanging: a factor 2.3. What governs the coherence time is therefore the fraction of bath pairs sharing a transition frequency, not the size of the zero-field splitting. Vacancy aggregation extends NV$^-$ spin coherence rather than shortening it.

Operational pseudospectral atmospheric models such as the ECMWF Integrated Forecasting System (IFS) run today almost exclusively on CPUs; GPU ports are under active development but not yet used in production. These models rely on the Spherical Harmonic Transform (SHT). Each time-step requires forward and inverse SHTs, and both passes depend on global pencil transposition that redistribute multi-dimensional arrays across compute nodes. At large node counts these global collectives dominate wall-clock time. We investigate GPU-resident lossy compression, using representative fields from the DYAMOND high-resolution operational dataset as input, and combining measured GPU compression throughput with SimGrid network simulation, we show that ZFP at 16 bits per value (rate-16), about the same storage budget as float16, matches the communication-time reduction of float16 truncation while delivering approximately $\mathbf{1600\times}$ lower mean relative error. ZFP at 8 bits per value (rate-8) achieves approximately $\mathbf{1.93\times}$ the speedup of float16 while retaining $\mathbf{4\times}$ lower mean relative error.

Machine learning interatomic potentials (MLIPs) can treat large-scale material systems with near first-principles accuracy and have been widely used to accelerate point-defect simulations. However, the training of MLIPs usually relies on large amounts of DFT data. This issue is particularly pronounced for charged defects, for which DFT calculations of large supercells are required to avoid long-range Coulomb interactions and finite-size effects, making the construction of datasets computationally expensive. In this work, we propose an efficient MLIP training scheme for neutral and lowly charged point defects based on small supercells (less than 100 atoms) and limited number of DFT calculations. The scheme requires only four DFT structural relaxations to construct the training dataset and the trained MLIPs dedicated for the defect can predict the defect formation energies in larger supercells (over 200 atoms) with small errors (mostly smaller than 0.3 eV). Using defects in GaN, SiO$_2$, and $\mathrm{Cu}_2\mathrm{ZnSnS}_4$ as representative examples, we evaluate the extrapolation capability of this scheme for predicting defect total energies and structural relaxations across different supercell sizes. The results show that an MLIP trained only on single small defect supercell produces severe errors when treating large supercells. Incorporating defect data from multiple supercell sizes improves the predictive performance of the MLIP, while adding pristine defect-free bulk supercells further enhances the accuracy. These results provide practical guidance for training MLIP models of defect systems with low computational costs.

We consider sequential Plurality elections in which each voter may abstain or vote for a single candidate. Each voter assigns utilities to all candidates; for each voter, this induces a (weak) order over the candidates. Ties are resolved uniformly at random, and voting has a small positive cost, so that a voter prefers to abstain when their vote cannot change the election outcome. We consider a variant of this model where, for each voter, we additionally specify a prefix of her ranking, so that she is only allowed to vote for a candidate from that prefix (or abstain). We prove that for this variant of the model, deciding whether a designated candidate is among the election winners in a subgame-perfect equilibrium of the associated extensive-form game is PSPACE-complete. This partially resolves an open problem from the work of Desmedt and Elkind [2010].

Understanding the entanglement cost of non-local quantum computation (NLQC) is relevant to complexity theory, cryptography, quantum gravity, and other areas. A central special case is $f$-routing, motivated in part by quantum position verification. Proving lower bounds on its entanglement cost in the fully robust setting has been a major open problem in NLQC. Motivated by this problem, we establish two related lower bounds. First, we study the shared-randomness cost of robust conditional disclosure of secrets (CDS). The connection between CDS and $f$-routing established by Allerstorfer et al. (Quantum 2024) makes understanding the randomness complexity of robust CDS a natural step toward lower bounds for the fully robust routing problem. We show that the shared-randomness cost of robust CDS is lower bounded by the logarithm of deterministic SMP communication complexity, even when communication and private randomness are unrestricted. Our lower bound is tight for the equality function. Second, we consider one-sided-perfect $f$-routing, in which the protocol is exact on one input class and has constant error on the other. By exploiting the positivity of the low-rank matrix in the method of Asadi, Culf, and May (ITCS 2025), we derive a lower bound on the entanglement cost in terms of sign rank. In particular, this yields a linear lower bound on the entanglement cost of routing for the inner-product function in both one-sided-perfect settings, matching the known upper bound.

Instruction-tuned LLMs faced with underspecified queries often commit to a single interpretation rather than ask for clarification, producing confidently wrong answers. In our experiments, prompting alone is insufficient: models either ask for clarification on every query or ask vague questions that fail to recover the missing information. Addressing this failure requires learning two coupled skills: when to ask rather than answer and how to ask a question that recovers the disambiguating information. Existing recipes either address only one of these skills or require a separately trained critic. We propose CIGAsk, an RL recipe that teaches both skills through two complementary reward signals within a multi-turn GRPO loop. Counterfactual Information Gain (CIG) compares the gold-answer log-likelihood under a frozen reference model with and without the user response, providing per-turn credit that guides how to ask. The Asymmetric Ambiguity Bonus assigns a signed reward at the terminal token based on the gold ambiguity label, guiding when to ask. Across three clarification benchmarks spanning table, passage, and open-domain QA, CIGAsk-7B outperforms the strongest external baseline despite using a smaller backbone. It also transfers across datasets without per-dataset tuning while preserving single-turn QA performance on out-of-distribution benchmarks.

As Large Language Model (LLM) agents are applied in continuously interactive environments, driving the evolution of their own capabilities becomes a core problem for achieving long-term autonomy. Currently, environmental knowledge is typically treated as an external fixed input rather than as part of the agent's ongoing evolution. Reinforcement learning methods usually optimize policies through environmental interaction but tend to adapt only to fixed task distributions or single environments. This paper proposes TTSE (Two-Track Self-Evolution), a dual-track online self-evolution framework that separates evolving knowledge into FACT (environmental facts, whose reliability is continuously verified through interaction evidence) and TIP (task-conditioned implementation procedures). From a decision-theoretic perspective, we decompose the agent's excess risk into environment-representation regret and conditional-execution regret, characterize the conditions under which environment-conditioned policies strictly outperform condition-agnostic policies, and bound the downstream risk in terms of FACT identification error and cross-condition mismatch cost. In practice, TTSE's ablation experiments on GDPevo validate the advantage of dual-track evolution. On the classic agent task benchmarks ALFWorld and ScienceWorld, TTSE further demonstrates superior task adaptation. Moreover, TTSE is broadly compatible with existing skill self-evolution methods; combined with the Bayesian-Agent algorithm, a single-track ablation validates the dual-track advantage, substantially improving the aggregate score across the five major domains of SOPBench over three independent repetitions. Finally, on the real end-to-end task benchmark PinchBench, TTSE is integrated into a general agent framework via retrieval-based injection and stably outperforms the baseline across three independent runs.

Analog-to-digital conversion dominates the energy and area of analog compute-in-memory (CIM) accelerators. This paper presents SimSkip, a successive-approximation-register (SAR) skipping scheme that combines a lightweight StrongARM-comparator similarity detector with SAR-skipping-aware training. When two held voltages are similar, the upper bits of the preceding code are reused to skip SAR comparisons. Detector-induced disturbance extracted from 55-nm CMOS simulations and bit-exact prefix saturation are embedded in the forward pass, and the network learns to tolerate them through quantization-aware training and knowledge distillation. On an ISAAC-based W4A4 WRN-28-10, SimSkip achieves 77.23% CIFAR-100 top-1 accuracy (mean of three seeds) versus 78.4% for the baseline, while reducing SAR comparisons by 48.2% across three hardware-modeled layers. A first-order model estimates a 27.2% net ADC energy reduction after detector overhead.

Classifier-free guidance (CFG) improves conditional generation in Flow Matching, but strong guidance can distort the generated distribution and reduce diversity. We provide a geometric account of this behavior by viewing Flow Matching as a time-varying gradient flow and characterizing how CFG reshapes its underlying potential. This view explains how stronger alignment can be accompanied by mean displacement and trajectory concentration, and motivates controlling guidance through the model-implied terminal posterior mean. We therefore propose Posterior-Mean-Capped CFG (PMC-CFG), a training-free, per-sample method that adaptively retains the strongest feasible guidance without additional network evaluations. Experiments on synthetic and large-scale image-generation benchmarks show that PMC-CFG limits guidance-induced distortion and concentration while improving the alignment–diversity trade-off, with particularly strong benefits when nominal guidance is large.

Rotating radio transients (RRATs) form an isolated neutron star (INS) population characterized by their sporadic and short radio bursts. In the fallback disc model, it was suggested that most INS populations pass through an RRAT phase during their evolution, and the mechanism producing these bursts could be the enhancement of the flux of open magnetic field lines through the poles of dead pulsars due to penetration of the inner disc into the light cylinder. To test this idea, we determine the RRAT phases that could be produced by this mechanism along the evolutionary curves of the INS families on the period-period derivative ($P$–$\dot{P}$ diagram). These RRAT phases form an RRAT region in the $P$–$\dot{P}$ diagram, which is in agreement with the known RRAT distribution. Using our results, we also investigate the descendants of central compact objects (CCOs). We estimate that the CCOs similar to the three CCOs with measured $P$ and $\dot{P}$ are not likely to become radio pulsars (RPs). The situation is rather different for Calvera, a recently identified CCO. In the model, Calvera-like sources become RPs with lifetimes $ > 10^{8}$ yr. Indeed, there is a cluster of RPs in the region of \PPdot~diagram where descendants of Calvera-like CCOs evolve in our model. We also estimate that they have a very low birth rate, and therefore their immediate descendants may not currently exist in our Galaxy.

Symmetric molecules with identical nuclei form distinct nuclear spin isomers (ortho and para) that cannot interconvert via radiative or inelastic transitions, preserving clues about formation conditions. In the interstellar medium, cyclopropenylidene (c-C3H2) and its doubly-deuterated isotopologue (c-C3D2) provide a prime laboratory for studying these ratios. While measured for the main species, spin ratios for c-C3D2 remain unavailable. We present new detections of ortho and para transitions of c-C3D2 toward three Taurus molecular cloud sources from the GEMS Large program, and revisit four literature cores to investigate the para-to-ortho ratio of c-C3D2 for the first time. Column densities were computed using RADEX with dedicated collision rate coefficients, yielding total c-C3D2 column densities of (2.5-5.3) x 10e11 cm-2 for the new sources. The derived para-to-ortho ratios are 1.5(+0.7-0.6) (B213-C16-1), 1.3(+1.3-0.7) (TMC1-C), and 1.0(+1.4-0.7) (TMC1-NH3, statistical). For revisited sources L1544, TMC-1C, TMC-2, and L1517B, ratios are 0.4(+0.1-0.1), 0.5(+0.1-0.1), 0.6(+0.2-0.1), and 1.4(+0.7-0.5), respectively. Overall, four sources show statistical ratios and three show higher non-statistical values. This is the first observational determination of the para-to-ortho ratio in a doubly deuterated molecule using a detailed radiative transfer tool. The elevated ratios in some sources likely stem from p-D2H+ or D3+ driven reactions, and are independent of H2 and c-C3H2 ortho-to-para ratios. These findings offer an essential benchmark for chemical models investigating c-C3D2 formation pathways.

How spatial structures manifest in multi-time correlations is a fundamental question in turbulence. We develop a non-Hermitian bosonic framework that unifies equal-time anomalous statistics and their temporal propagation within a common operator representation. A continuous Wegner flow reorganizes stochastic mode couplings in an extended wave-frequency space. Applied to the Taylor-Kraichnan model, the framework recovers the established equal-time hierarchy and constructs multi-time contributions through propagators acting on successively smaller sets of active fields. This construction separates uniform transport from intrinsic relative dynamics and shows how a higher-order equal-time state evolves under a generator acting only on the fields that remain dynamically active. Within an isotropic radial closure, we derive an explicit fourth-order two-time zero mode scaling function that describe how relative dispersion progressively weakens sensitivity to the initial separation. The framework thus connects spatial intermittency to temporal evolution by identifying how the sequence of observation times determines the propagation of equal-time anomalous structures.

The asteroid 1998 KY$_{26}$ has been the subject of thorough observation, both optical and radar, from shortly following its discovery on 28 May 1998, through 2 further close apparitions in 2020 and 2024. This has previously allowed an accurate characterization of the object, including an unusually rapid spin-rate with period 5.3516 $\pm$ 0.0001 minutes, a diameter of 11 $\pm$ 2 m, and a high albedo of $\sim{0.52}$. Furthermore, the presence of significant nongravitational accelerations (NGAs), with no detectable shedding of gas or dust, has stimulated the 'dark comet' categorization, with JAXA repurposing their Hayabusa2$\sharp$ spacecraft to rendezvous with the object in 2031. We follow-up the astrodynamical evidence pointing to the possibility this may actually be the lost Soviet Phobos 1 probe, and analyse various photometry and astrometry associated with 1998 KY$_{26}$ to further investigate the feasibility of this hypothesis. We find a first order approximation of the Phobos 1 spacecraft provides compelling agreement to 8 light curves of the object and further that modelling NGAs as solar radiation pressure (SRP) on a cylinder or solar panels provide significant reductions in residual of $\sim{11} \%$ with respect to the 260 astrometric and radar measurements. Although an investigation of this kind cannot be conclusive, nevertheless the results here add strong weight to the Phobos 1 hypothesis, and we find no clear contradictory evidence. The spin pole in Ecliptic J2000 coordinates, calculated from photometry based on the Phobos 1 assumption is $(\lambda,\beta)=({151^{+1}_{-2}}^{\circ},{+11^{+12}_{- 3}}^{\circ})$ , where the quoted ranges represent approximate 68.3% confidence levels.

We present a gate-based quantum computing solution of the homogeneous Bethe-Salpeter equation (hBSE) for the bound state of two massive relativistic scalar particles interacting via ladder-approximation scalar exchange. After Wick rotation to Euclidean space and O(4) S-wave partial-wave projection, the hBSE is reduced to a symmetric matrix eigenvalue problem of dimension N = 2^n. We decompose the resulting Hamiltonian into a sum of n-qubit Pauli operators and solve it with the Variational Quantum Eigensolver (VQE) using a Matrix Product State (MPS) tensor-network ansatz. For N=16 (n=4 qubits), the VQE recovers the maximum eigenvalue - which encodes the minimum coupling constant for binding - to better than 1 % mean relative error compared to classical diagonalization. Entanglement analysis of the BSE amplitude shows low, area-law-like entanglement over the tested sizes, which motivates the MPS ansatz. A critical analysis of three independent barriers - exponential Pauli overhead, approximately size-independent low entanglement, and decreasing VQE gradient scales for the tested ansatz, consistent with generic barren-plateau concerns at larger n - reveals that the specific problem studied here lies in a classically tractable regime: over the tested range it is well described by low-bond-dimension tensor networks and efficiently handled by MPS/Lanczos methods. This negative result provides, to our knowledge, the first entanglement quantification of the BSE amplitude in qubit encoding, establishes the Pauli Hamiltonian framework for future BSE variants, and identifies 2D Minkowski-space BSE, N-body bound states, and non-ladder kernels as physically motivated extensions where genuine quantum advantage may become plausible.

Electronic entropy is usually treated as a secondary correction to structural stability, but under strong electronic excitation it can become a primary thermodynamic driving force. Here we show that electronic entropy can drive both polymorphic and stoichiometric phase transformations in compressed iron oxides. Using finite-temperature density functional theory, we calculate the electronic-temperature-dependent Gibbs free energies of Fe$_2$O, FeO, Fe$_4$O$_5$, Fe$_3$O$_4$, and multiple Fe$_2$O$_3$ polymorphs, including $\alpha$-, $\iota$-, $\zeta$-, $\eta$-, and $\theta$-Fe$_2$O$_3$, over the pressure range 60–260 GPa. At 60-140 GPa, electronic excitation mainly reorganizes the relative stability of Fe$_2$O$_3$ polymorphs, driving transitions from $\iota$-Fe$_2$O$_3$ to $\eta$-Fe$_2$O$_3$. At 180 GPa, the free-energy landscape becomes strongly competitive as FeO is stabilized over an intermediate range of electronic temperature, while $\eta$-Fe$_2$O$_3$ becomes favourable at higher T. At 220-260 GPa, the lowest-free-energy phase at low T is the Fe-rich compound Fe$_2$O, but increasing electronic temperature stabilizes FeO. These results demonstrate that electronic entropy can control not only the relative stability of crystal structures at fixed composition, but also the competition between different iron-oxide stoichiometries. The predicted electronic-entropy-driven phase boundaries provide a route to nonthermal structural transformations in ultrafast and high-energy-density experiments.

Malware detection is a critical task in cybersecurity, and graph neural networks over control flow graphs have shown promising results for it. However, detectors are usually evaluated on a random split of a corpus collected over a single period, which cannot show how well a model generalizes to later samples. This study addresses that limitation with a strict temporal split: every model is trained on one period and scored once on a later one. Two corpora of control flow graphs, each node carrying 37 features, were extracted statically from 1,989 Windows portable executables: 459 graphs from 2024-2025 for training and 223 from 2026 for evaluation. Twelve variants and a flat-feature control were trained on the earlier corpus. The choice of message-passing operator changes robustness to the shift significantly, and every pairwise gap that survives correction separates an aggregating architecture from one built around a learned attentional readout. The ranking also reverses: the flat control, which sees node features but no topology, is the best in-distribution model and among the worst across the boundary, so a conventional benchmark would have rejected message passing. Neither recalibration nor ensembling substitutes for the operator choice. Attributions do not shift, but explanation validity is architecture-specific, and the most accurate operator on the later corpus is the hardest to explain. An architecture derived from the finding matches the best searched operator without search. The shift affects both malware and benign classes alike, so these are results about robustness to distribution shift, not malware evolution.

Massive three-loop form factors for vector, axial-vector, scalar, and pseudoscalar currents are vital for precision collider phenomenology. Extending the quarkonic framework to the more complex gluonic case, in Ref. (arXiv:2609.22034 [hep-ph]) we computed these contributions widely using advanced computer algebra, automated guessing, and large-scale PSLQ searches. Here, we outline the core strategy of the large-moment method and illustrate its main challenges, including solving record-sized recurrences and differential equations. Our high-precision evaluation provides an exact analytic representation around $s=0$ in terms of MZVs. We achieved a complete analytic series expansion around $s=\pm\infty$ for the first time, which depends on MZVs and three constants out of higher number spaces, along with reliable analytic continuations across $s \in (-\infty, \infty)$, serving as a landmark benchmark for modern symbolic computation.

Change point detection (CPD) identifies abrupt and significant changes in sequential data, with applications in human activity recognition, financial markets, cybersecurity, manufacturing, and autonomous systems. Traditional CPD methods often face computational challenges in high-dimensional settings and typically provide limited explanations for detected changes, which can restrict their practical usability. This paper introduces a CPD framework that improves scalability and interpretability by leveraging the Sliced Wasserstein (SW) distance. Our contributions are fourfold: (1) we transform multivariate sequential data into one-dimensional scores using the SW distance, making the resulting representation compatible with existing CPD methods; (2) we analyze the distributional behavior of random slices of the SW distance and show that, under suitable assumptions, they can be approximated by a Gamma distribution, providing a principled basis for threshold calibration; (3) we propose a self-adapting online CPD algorithm that combines this SW-based score with an adaptive quantile-based threshold; (4) we introduce a model-specific framework for generating contrastive explanations for annotated change points. Empirically, our method reduces false positives by at least $48\%$ on average compared with popular online and offline CPD baselines, while maintaining competitive or superior detection performance. Code is available at https://github.com/jsve96/SWCPD_Code. At the same time, it produces interpretable change-point annotations, making it practical for deployment in high-stakes applications.

GUI agents predict click coordinates as digit-token sequences, but standard text-LLM confidence estimation methods rank correct clicks from wrong ones only weakly. GUI-specific alternatives use K samples or new supervision, but still leave room for improvement. We trace part of this to place-value asymmetry: bounding-box correctness often makes higher-place digits more important than lower-place digits, so uniform aggregation weakens the signal that determines correctness. The fix is to weight each digit's Shannon entropy by its place value. We call this Place-Aware Coordinate Entropy (PACE). Across fixed-scale agents on ScreenSpot-Pro and ScreenSpot-v2, PACE wins both AUROC and selective accuracy on all primary comparisons in a single forward pass, matching or outperforming K-sample baselines at a fraction of the cost. PACE provides a per-click confidence estimate that turns coordinate-token internals into a practical confidence signal for GUI agent deployment.

Hyperbolic vision-language models (VLMs) represent image and text features in a geometry naturally suited to hierarchy, but their adaptation to downstream tasks has largely relied on fixed prompts. Existing prompt learning methods, meanwhile, treat class labels as a flat set and do not exploit available taxonomic structure. We address this gap with a hierarchical prompt learning plug-in for frozen hyperbolic VLMs. Given a fixed offline parent-class hierarchy, it augments a class prompt learner with a separate parent prompt learner, parent-level supervision, hyperbolic entailment regularization, and parent-feedback logit fusion. We instantiate the method with CoOp, CoCoOp and MaPLe, yielding HyPLO, CoHyPLO and MaHyPLO. Across the standard 11-dataset benchmark, all variants improve base-to-new generalization and cross-dataset transfer, and remain comparable to their prompt learning baselines under domain shift. Six hierarchical metrics and embedding analyses show that the method produces more taxonomically consistent predictions and induces a hierarchy-consistent organization of parent, class, and image embeddings in hyperbolic space. Its gains are largest when novel classes must be placed within a fixed taxonomy, and smallest for fine-grained confusions among sibling classes or shifts affecting only the image distribution.

Text-to-speech (TTS) corpora are costly to record, yet many utterances add little new phonetic information. Core-set selection reduces this cost by choosing a small training subset under a fixed audio-duration budget. We represent a corpus as a phonotactic graph that links each utterance to its most phonemically similar ones, and we first test whether this graph has structure. In Bangla and English corpora, its clustering is 199 and 56 times that of a size-matched random graph, and its modularity is more than twice that of a degree-preserving random graph. We then propose Community Representative, a selector that samples across graph communities and spreads its choices within each one, starting from utterances rich in rare phonemes. At every budget and in both languages, it covers more rare phoneme bigrams than random and entropy-based selection, and this lead holds on held-out utterances. TTS models trained on its 20% core-sets have a significantly lower character error rate (CER) than models trained on equal-duration random or entropy-based subsets in both languages. When all models train for the same number of epochs, the Bangla core-set model also outperforms full-corpus training (3.93% vs. 4.47% CER) with 4.5x less training time.

Deploying language-conditioned manipulation without a dedicated GPU requires efficient inference and action chunks that cover the delay between policy queries. We present vla.simd, a CPU inference engine that combines shared SIMD micro-kernels, reusable computation, and target-specific optimization. We relate query latency and execution horizon to action availability under lagged and time-aligned execution, distinguishing action supply from feedback frequency. Across six policies and four CPUs, vla.simd achieves approximately $1.4\times$ median speedup over compiled PyTorch references while preserving fp32 numerical fidelity. We also introduce IMPACT, an ACT-based policy with cached text representations and language-modulated visual features. IMPACT is the only language-conditioned policy in our evaluated set that supplies at least 30 actions/s on the Raspberry Pi 5: after a 90 s thermal soak, it supplies 33.5 actions/s in fp32 and 81.2 with int8. Separate GPU evaluations yield $76.4\%$ mean success across four LIBERO suites without robot pretraining; instruction-shuffling tests demonstrate selection among familiar goals. Trials with IMPACT on an SO-101 arm and SmolVLA on a UR10e with a Robotiq gripper demonstrate CPU deployment on two robot embodiments.

Semiclassical gravity provides a promising framework to address the black hole singularity problem through quantum backreaction characterized by a coupling parameter, making it essential to thoroughly explore the accretion properties around such objects. In this paper, within the framework of semiclassical gravity with type-A trace anomaly, we investigate the thermal radiative signatures of Novikov-Thorne accretion disks around rotating black holes across diverse parameter spaces, focusing on the energy flux density, temperature profile, and thermal luminosity. Our findings demonstrate that the introduction of a positive coupling parameter suppresses disk emission, leading to lower peak values of both energy flux density and temperature, as well as lower peak values and corresponding peak frequencies in thermal luminosity compared to their Kerr counterparts in general relativity. In particular, this discrepancy becomes remarkably pronounced when the black hole possesses a high spin. Mechanistically, increasing the coupling parameter reinforces the effective gravitational field strength of the central compact object, which pushes the equatorial stable circular orbits outward and increases the specific energy of the orbiting matter, thereby reducing the radiative efficiency and weakening the overall disk emission. This suppression mechanism offers a potential astrophysical window to distinguish semiclassical gravity from general relativity and to place observational constraints on the coupling parameter via thermal disk spectra.