Papers

Large language models (LLMs) are increasingly deployed across languages, yet their safety and reasoning capabilities remain predominantly evaluated in English. Translating existing benchmarks offers a scalable approach to multilingual evaluation, but preserving task-specific meaning across languages presents significant challenges, particularly for harmful requests, mathematical reasoning, and conversational interactions. We introduce a reasoning-guided translation framework for constructing multilingual evaluations designed to preserve the structure and intended meaning of diverse benchmark tasks. Our approach separates translation strategy selection from large-scale generation, combining reasoning-guided configuration selection with task-aware validation, targeted repair, and independent quality screening. Using this framework, we develop a source-aligned evaluation collection spanning seven safety, conversational, and mathematical reasoning resources across 11 languages. The resulting collection achieves 99.94% coverage of intended item–language pairs in the two subsets requiring repair. Evaluations of open-source LLMs reveal distinct language-dependent vulnerabilities, including inconsistent refusal behaviour and substantial reasoning degradation in lower-resource languages. These findings highlight the importance of task-sensitive benchmark construction and systematic quality assessment for understanding multilingual model behaviour beyond English.

Autoregressive generation underlies real-time, causal video synthesis and is the backbone of video world models. Today's dominant approach distills a bidirectional diffusion model into a causal few-step student, but existing pipelines make this conversion prohibitively expensive, requiring offline Probability Flow ODE} (PF-ODE) trajectories that cost tens of thousands of GPU-hours and terabytes of storage. We show it can be made dramatically more efficient. Pushing this efficiency to its extreme, we reduce trajectory supervision to a single sample, the strictest setting possible, and find that it causes severe quality collapse. We trace this to the teacher's trajectory geometry: the high-noise regime, governing semantic layout and motion, and the low-noise regime, governing fine texture exhibit fundamentally different curvature and optimization behavior, so a single monolithic student cannot jointly fit both, causing severe representation interference. Motivated by this, we propose Glance Forcing, a lightweight distillation framework that decouples the denoising trajectory into two complementary slow and fast experts, each guided by a tailored, stage-specific objective, while keeping the pretrained teacher entirely frozen. This decoupling further reveals a general mechanism underlying few-step distillation: PF-ODE distillation mainly straightens the few-step trajectory geometry, while the subsequent distribution-matching stage independently governs distributional quality. Using only one PF-ODE trajectory, Glance Forcing already surpasses the fully-trained state-of-the-art baseline, raising the VBench Total score from 84.04 to 84.22 while cutting trajectory-collection cost from thousands of GPU-hours to about one A100-GPU-hour, with markedly larger gains in dynamic degree, VisionReward, and instruction-following, and consistent preference from human raters over strong baselines.

Integrated sensing and communication (ISAC) equips cellular networks with the dual capability of data transmission and environmental sensing by sharing spectrum, waveforms, and radio hardware. While network-side sensing via transmission-reception points (TRPs) has been widely investigated, the use of user equipment (UEs) as sensing receivers introduces additional bistatic viewpoints and can improve target observability. UE-assisted operation is nevertheless constrained by small UE apertures and by waveform processing, especially when data-bearing communication signals are reused for sensing. The present paper investigates theoretical limits and operating conditions for UE-assisted passive target positioning, with emphasis on hybrid operation that fuses TRP–TRP and TRP–UE measurements. A unified OFDM-based signal-and-noise model is developed to cover both positioning reference signals (PRS) and communication waveform reuse with modulation-dependent noise enhancement. Based on this model, closed-form Fisher-information expressions and Cramer-Rao lower bounds are derived for delay and angular parameters and are mapped to positioning error bounds (PEBs) for two estimator classes: time-angle positioning based on TDOA + DoA and delay-only 3D ellipsoid positioning based on excess-delay/TDOA measurements. Monte-Carlo simulations across reference SNR, subcarrier spacing and occupied bandwidth, modulation order, radar cross section (RCS), sensing mode, and positioning algorithm identify regimes where UE participation and hybrid sensing provide significant gains and quantify the associated estimator-to-bound trade-offs.

Diffusion models are a powerful framework for learning complex probability distributions. While widely used diffusion models for continuous data rely on Gaussian noising processes, we instead develop diffusion models based on Poisson noise. More precisely, we introduce and study a new mathematical framework, called Mixed Poisson Diffusion Models (MPDM), with noising dynamics driven by multivariate counting processes. Our methodology is motivated by Poisson inverse problems. The success of Gaussian diffusion models has led to their use as prior models in a growing number of posterior sampling methods for inverse problems. However, most of these approaches focus on Gaussian inverse problems, in which observations are generated through a deterministic forward operator corrupted by additive Gaussian noise. By contrast, Poisson inverse problems, where observations are discrete and Poisson distributed, remain comparatively underexplored in this literature. Building on our Poisson generative dynamics, we develop a natural framework for posterior sampling in Poisson inverse problems and derive several posterior sampling algorithms based on the proposed model. We evaluate our approach on realistic Poisson inverse problems and show that the resulting framework is well suited to this class of problems.

Building foundation models for MRI has emerged as a promising direction for advancing medical image analysis and improving accessibility of task specific models. Existing approaches primarily focus on effective pretraining strategies and evaluating their performance across diverse downstream tasks. While these methods show performance gains, limited attention has been given to understanding task alignment between pretraining and downstream objectives. Unlike language models, where the two stages are naturally aligned, such alignment is less evident in vision based applications. Motivated by this gap, we explore continual pretraining. Specifically, continually pretraining a pretrained ViT encoder using Low-Rank Adaptation (LoRA) on a prostate segmentation task to enhance task alignment and downstream performance. We further perform layer-wise representational analysis using Centered Kernel Alignment (CKA) to assess representational shifts and their correlation with task outcomes. Our experiments show improved performance in 3 of 6 downstream tasks and comparable results in the others. Our Feature analysis reveals that, for segmentation tasks, continually pretrained encoder features exhibit greater similarity to those of the finetuned encoder than the original pretrained encoder, correlating with observed performance gains.

We introduce a compact Topological-Geometrical Descriptor (TGD) for planar two-dimensional carbon networks. Atomic coordinates are mapped to a periodic planar graph and represented through ring populations, topological-charge statistics, projected ring geometry, and dual-graph defect adjacency. The resulting modeling vector contains 19 physically interpretable features. We evaluate TGD on 120 published planar carbon allotropes using the DFT relative energy per atom with respect to graphene as the target. Repeated five-fold cross-validation with gradient boosting yields a mean absolute error (MAE) of 0.0826 +/- 0.0138 eV/atom and R2 = 0.803 +/- 0.074. Under the same validation protocol, TGD gives a mean MAE of 0.0828 eV/atom, compared with 0.0929 eV/atom for a 203-feature Matminer representation. We further analyze 360 vacancy-defected graphene configurations spanning 12 vacancy concentrations from 2 to 24 percent. When entire vacancy concentrations are held out, TGD gives a pooled MAE of 0.269 eV/vacancy and R2 = 0.861, outperforming simple baselines based on vacancy concentration alone or non-hexagonal-ring fraction alone. Sensitivity tests show that the inferred edge sets are stable to moderate changes in the adaptive bond factor. Finally, the mean absolute topological charge is examined as a one-variable empirical surrogate for the allotrope dataset and is treated explicitly as dataset-specific rather than universal. These results show that TGD combines compact prediction with direct structural interpretation while making clear the present restriction to single-layer planar carbon networks.

Smart healthcare IoT requires low-latency and reliable delivery of patient data, yet IP-based communication remains host-centric and offers limited support for mobility, caching, and request aggregation. This paper evaluates a smartphone-centric Named Data Networking (NDN) approach in which a patient's smartphone aggregates sensor streams and publishes a stable latest profile under a consistent name, enabling effective in-network caching. Using ndnSIM simulations on a realistic medical IoT topology, we compare NDN and IP under identical traffic and link conditions. Results show that per-sample naming provides little benefit, but smartphone aggregation with stable naming allows NDN caching to reduce delay after cache warm-up and achieve significantly higher delivery ratios under congestion compared to IP.

We define genus character $L$-functions of orders in quadratic function fields of odd characteristic. We express each as a product of two quadratic $L$-functions and explicit local factors at primes dividing the conductor, with a possible shift of the variable in one of the two quadratic $L$-functions. This product formula yields meromorphic continuation and functional equations. In the real case, we evaluate special values in terms of continued fractions and derive formulas for the product of class numbers of two imaginary quadratic function fields. These formulas are function field analogues of the Kaneko-Mizuno-Zagier formula.

We prove that every stable atomic (twisted) vector bundle on a projective hyper-Kähler manifold is 1-obstructed. In the appendix, we construct a simple atomic object that is not 1-obstructed, providing a counterexample to Beckmann's conjecture.

A well-known conjecture of Toda's asserts that the BPS invariants associated to a Calabi–Yau threefold are independent of the chosen Euler characteristic. We answer the question in the affirmative for a certain class of local surfaces, which we call warped local curves. Along the way we show that the moduli stack of sheaves on a warped local curve can be written as the global critical locus of a function. We also show that the moduli stack of sheaves on a curve fibration over $\mathbb{A}^1$ is the global critical locus of a function.

Classical film theory predicts a finite diffusion-limited current at porous electrodes. However, over-limiting currents (OLCs), which exceed this limit, are often observed and have been attributed to surface charge, confinement, or electrokinetic effects. Here, we use a frequency-domain framework to identify a different mechanism governed by the time scale of concentration variation. For perfectly reactive interfaces, the spectral Sherwood number, a frequency-dependent dimensionless measure of interfacial mass transfer relative to molecular diffusion, grows as $\Shnd\sim\omega^{1/2}$ at high frequency, while its phase approaches $\pi/4$. This response corresponds to the semi-infinite Warburg limit of the electrochemical impedance. A scaling analysis shows that the exponent and prefactor are determined by diffusion within a thin layer adjacent to the interface and are therefore independent of the global pore geometry. Numerical results for parallel plates, circular tubes, and twelve porous microstructures confirm this scaling. Geometry, flow strength, and flow direction affect the low-frequency response and crossover frequency, but not the asymptotic scaling in the high-frequency limit. This temporal route to OLCs requires neither surface charge nor confinement and could inform the design of systems for pulsed electrodeposition, mineral recovery and renewable-energy storage.

Counterfactual reasoning (asking what would have happened, had a decision been different) is a powerful tool for identifying causes of failure and explaining the behaviour of AI systems in sequential decision-making tasks. In Markov Decision Processes (MDPs), two sources of uncertainty complicate counterfactual inference: statistical uncertainty in the MDP transitions when the MDP is estimated from trajectory data, and uncertainty about the unknown causal model, as there are multiple structural causal models (SCMs) compatible with each MDP. Existing work addresses these two challenges in isolation. We address both simultaneously by introducing a framework for partial counterfactual inference in MDPs that is robust to both statistical and causal uncertainty. Our approach provides: 1) closed-form analytical bounds for the counterfactual probabilities that are valid across all SCMs entailed by a given interval MDP (IMDP); 2) a method to compute minimal robust interventions to achieve a desired counterfactual outcome via pessimistic value iteration on the induced counterfactual IMDP. Since the latter is constructed using our analytical bounds, our approach remains efficient and scalable to realistic MDP environments. Experiments across multiple case studies demonstrate our method's robustness and practical utility for identifying minimal sets of actions that, if changed, would have led to improved outcomes.

Routability prediction enables rapid evaluation of placement quality without routing, providing effective feedback for routability-driven placement optimization. Recent machine learning methods have achieved efficient routability prediction through data-driven approach. However, their prediction performance highly depends on the capability of input representations to capture essential circuit information. To address the limitations of existing circuit representations in simultaneously preserving logical topology, physical layout and routing resource information, this paper proposes HGRP, a heterogeneous graph neural network for routability prediction. HGRP models the cells, nets, and routing grids as distinct types of nodes and explicitly represents topological connectivity, spatial mapping, and grid adjacency, thereby constructing a heterogeneous graph representation that preserves multi-level circuit semantics. Based on the representation, a heterogeneous message passing mechanism is developed to enable information interaction and feature fusion among distinct types of nodes, effectively learning the complex correlations between circuit structures and routing behaviors. Experimental results on the CircuitNet dataset show that HGRP improves F1-score by 18.69% relative improvements compared with the state-of-the-art in the congestion prediction task. The source code is publicly available at https://github.com/ShenShan123/HGPR.

We construct a rank-9 uniform non-homogeneous vector bundle on ${\mathbb P}^5$. This yields the first counterexample to the longstanding open conjecture that every uniform vector bundle of rank $r<2n$ on ${\mathbb P}^n$ is homogeneous.

Let $\Omega=\Omega^+\subset \mathbb{R}^{n+1}$ be a bounded open set with $n$-Ahlfors regular boundary and let $\Omega^-=\mathbb{R}^{n+1}\setminus\overline{\Omega^+}$. For $x\in\partial\Omega$, $r>0$, define \begin{equation*} \varepsilon_{n,\Omega}(x,r):=\frac{1}{r^n} \inf_{H^+}\mathcal{H}^n\left(((\partial B(x,r)\cap H^+)\setminus \Omega^+) \cup ((\partial B(x,r)\cap H^-)\setminus \Omega^-)\right), \end{equation*} where $\mathcal{H}^n$ stands for $n$-dimensional Hausdorff measure and the infimum is taken over all open affine half-spaces $H^+$ such that $x\in \partial H^+$. Here $H^-=\mathbb{R}^{n+1}\setminus \overline{H^+}$. In this work we prove that $\Omega$ is a $C^{1,\alpha}$ open set, with $0<\alpha<1$, if and only if there exists some constant $C>0$ such that $\varepsilon_{n,\Omega}(x,r)\leq C\,r^\alpha$ for all $x\in\partial\Omega$ and all $0<r\leq \mathrm{diam}(\partial\Omega)$.

We develop novel methods for the fast construction of spectral approximations to smooth linear and weakly singular Volterra integral operators and for the fast evaluation of smooth nonlinear Volterra integrals. The proposed methods employ the Duffy mapping, which transforms the original indefinite integral into a definite integral and maps the original causal triangular domain to a square, thereby avoiding any artificial extension of the kernel function. For smooth linear and weakly singular Volterra operators, the resulting approximations are almost-banded and lower-banded, respectively. The $N\times N$ principal finite section of these approximations can be constructed in $O(N)$ and $O(N^2)$ flops, respectively, via recurrence relations. For Volterra integrals with a smooth kernel that depends nonlinearly on a given function, recurrence relations lead to fast evaluation with a complexity of $O(N^2\log N)$ when the function is approximated by a degree-$N$ Chebyshev approximant. These spectral approximations and evaluation procedures give rise to fast and stable spectral methods for solving Volterra integral equations, as well as a broad class of problems involving Volterra integral operators. Extensive numerical experiments show that these spectral methods significantly outperform existing methods in terms of stability, speed, accuracy, and ease of implementation.

Cooperative success rests in part on implicit communication. Partners converge on conventions: regularities that give an action meaning beyond its direct effect. Yet conventions are usually observed only indirectly: a score drop when partners are swapped shows that a convention existed, but not what it was or where it was used. In the card game Hanabi, the exact probability that a play fails can be computed from the hints received and the cards in view. Players who use conventions succeed more often than this prediction allows; the resulting margin is the convention gap. Off-belief learning provides a hierarchy of agents (levels 1-5) whose first level uses no conventions by construction and whose later levels use progressively more. For humans and these agents on one scale, we ask what the conventions are and where they are used. On what: in two independently collected corpora of human games, humans show a larger gap than every agent level. On the conventions both populations use, however, humans do no better than level 2. The human advantage is concentrated in a situation that no agent level learns to use: hints that point to a card other than the ones they touch. On where: conventions depend on what the player already knows about the card, not on how recently it was hinted, and they are used most in the mid-game; humans use them earlier and later in the game than the agents. In interventions on the agents, blocking a single high-gap situation costs little, because meaning reroutes to other situations, but blocking several together costs far more than the sum of the single costs. Thus, the convention gap turns conventions from an inferred cause of score loss into a measurable, comparable, and causally testable object.

We present a novel method which combines the invertibility of the differential filter of Germano, with convolutional networks to accelerate plasma turbulence simulations. Rather than learning the mapping from coarse-grid LES fields to fine-grid DNS, the proposed method learns the mapping from coarse LES fields to the differential-filtered DNS fields. We then use the inverse of Germano's filter to reconstruct the fine-scale details and hence calculate exactly the nonlinear terms in the governing equations and subgrid-scale terms in the LES equations. The method is tested on the 2D Hasegawa-Wakatani (HW) and modified Hasegawa-Wakatani (mHW) plasma turbulence models at $256^2$ and $512^2$ DNS resolutions using 2, 4 and 8 times smaller LES grids in each direction. We compare the model with DNS results on energy, enstrophy, spectra, probability distributions, poloidal and radial fluxes and computational cost. Validation extends to 1000 simulation time units for HW and 100 for mHW, whose faster dynamics lead toward laminarization. For the $256^2$ cases, the finer LES meshes preserve several integral and spectral features, whereas the coarsest mesh yields unphysical behavior. In contrast, for the $512^2$ cases, even the coarsest LES mesh preserves the turbulent character of the flow and maintains acceptable statistics for both the HW and mHW models. Moreover, we show how our methodology is superior to traditional algebraic numerical LES models, which traditionally fail to capture important design features. Finally, benchmarks show an MPI task-time reductions of 4.82–49.45, with a maximum wall-clock speedup of $1.55\times$ when compared to full CPU socket.

This paper investigates the use of Behavior Cloning (BC) for autonomous driving from a bird's-eye view (BEV) perspective in simulated urban environments. BC uses supervised learning to mimic expert driving behaviors. Previous works have applied BC in the CARLA simulator but did not fully address the challenges of traffic light compliance. Our approach enhances BC by integrating a kernel density estimator to adjust training sample weights based on action density, thereby improving the learning of rare but critical actions such as stopping at red lights and accelerating at green lights, specially in scenarios of scarce number of expert demonstrations. Using BEV inputs, which provide an abstract top-down view of the driving environment, our method simplifies the policy learning process. The trained convolutional neural network (CNN) outputs steering and acceleration actions based on these BEV inputs and additional state variables. Experimental results in the CARLA simulator demonstrate that our weighted BC method significantly improves driving performance, achieving higher route completion compared to standard BC. This weighted approach proved to be crucial in learning correct driving behaviors, particularly in test environments not encountered during training, highlighting its potential for enhancing autonomous vehicle navigation.

Hong-Ou-Mandel (HOM) interference offers robustness in high-noise environments due to its phase-insensitive nature, while N00N-state interference provides enhanced phase sensitivity for precision measurements in low-noise settings. However, their realizations have typically required separate experimental platforms. Here, we report a single hybrid quantum interferometer that can dynamically transition between HOM and N00N-state interference by adjusting two time delays, $\tau_1$ and $\tau_2$. This tunability enables operation in phase-insensitive, phase-sensitive, or hybrid modes within a unified setup, as confirmed by theoretical analysis and experimental demonstration. Our system offers a versatile platform with potential for quantum sensing applications, paving the way for integrated quantum technologies capable of adapting to diverse noise conditions.

Gravitational-wave (GW) ringdown signals provide a powerful probe of the strong-field regime of general relativity (GR) through black hole spectroscopy and tests of the Kerr hypothesis. However, unmodeled propagation effects can bias the inference of remnant properties and the interpretation of ringdown features. We investigate the impact of gravitational lensing on ringdown inference and tests of GR using GW150914-like simulated signals spanning the long-wavelength, wave-optics, and geometric-optics regimes. We perform complementary time-domain post-merger and frequency-domain full inspiral–merger–ringdown analyses using unlensed waveform models. We find that lensing in the wave-optics regime can produce substantial biases in the inferred remnant mass and spin and can strongly affect the inferred modal content of the ringdown. In particular, it can spuriously favor the presence of the first overtone, with $\log_{10}\mathcal{B}^{221}_{220}\sim40$, even though the injected signal has no support for it. Parameterized ringdown tests can likewise yield strong apparent support for deviations from the Kerr quasinormal mode (QNM) spectrum, with $\log_{10}\mathcal{B}^{\rm modGR}_{\rm GR}\sim8$. In contrast, the evidence for higher-order modes remains largely unchanged for the simulated signals considered here. The parameterized analysis using the entire signal further shows that microlensing can bias the inferred frequency of the dominant $(2,2,0)$ QNM. These results demonstrate that unmodeled microlensing can produce systematic biases and spurious signatures in ringdown analyses, potentially affecting precision tests of the Kerr hypothesis.

Large Language Models (LLMs) are increasingly deployed in multi-agent environments that produce unanimous agreement, both in intended interactions that need coordination and in unintended emergent interactions. Consequently, understanding how LLMs with competing incentives reach consensus is crucial to better interpret and design LLM collectives. We introduce an evaluation, based on the Network Consensus game, to assess the mechanisms driving this process. Here, a network of agents must agree on one of two colours to receive a payoff, but agents are offered different payoffs for the agreed colour. We vary components such as communication access and network structure and assess six open-source models and one closed model (ChatGPT 5.6-sol). We find that open-source LLMs reach consensus less often than humans in comparable experiments and only if messages are communicated across the global network. ChatGPT 5.6-sol, however, is able to reach consensus with communication limited to its neighbours and with no communication ability. This is achieved through lower stubbornness and higher majority-following tendencies, and by providing more information about its neighbours' states in messages. However, situating agents in a highly clustered graph reduces consensus, regardless of these traits, demonstrating that both system design and individual model traits drive collective outcomes.

We develop a harmonic framework for spherically symmetric Friedmann-like geometries with constant scalar curvature but non-zero trace-free Ricci curvature. This preserves Friedmann-type isotropic expansion while introducing a radially varying tidal field and supporting anisotropic stress. We formulate the scalar Laplace-Beltrami operator as a Sturm-Liouville problem and establish its self-adjoint domains and regularity at minimal spheres. Robertson-Walker harmonics are recovered in the maximally symmetric limit. Trace-free curvature terms are expressed as spectral projection kernels. Rotational symmetry preserves angular labels and parity, while the tidal profile couples different radial eigenmodes. A finite-shell calculation displays this off-diagonal structure, verifies consistency between radial representations, and confirms the expected weak-field scaling. A minimally coupled scalar field remains diagonal in the exact basis, indicating operator-dependent mixing rather than mixing solely due to inhomogeneity. We formulate transverse-traceless kernels, whose dynamical application requires a specified model for the supporting anisotropic matter. This converts curvature-induced mode mixing into a spectral problem for perturbations beyond maximally symmetric cosmologies.

Solving linear systems of equations is a fundamental task in machine learning and optimization. In this work, we extend the original WZP algorithm from real symmetric linear systems to complex Hermitian linear systems. We then study the success probability of the modified WZP algorithm from a resource-theoretic perspective. By expressing the postselection success probability as the expectation value of a positive operator, we derive upper and lower bounds of the maximal success probability over separable, incoherent, and real pure states. These bounds further establish connections between the maximal success probability and geometric entanglement, coherence, and imaginarity. We further show that, under permutation averaging, the averaged success probability is an affine function of the coherence fraction, with its dependence determined by the alignment between the success operator and the equal-superposition state. We also investigate the preprocessed WZP algorithm under subgroup-constrained preprocessing. For the two-level family, the optimal success probability is determined by the maximal projection weight onto the orbit of the target projector. Our results provide a unified perspective on the WZP success probability under different operational constraints.

We establish new restrictions on the virtual cohomological dimension of word hyperbolic groups admitting Anosov representations into $\operatorname{PO}(p, q)$ ($p > q \geq 2$) or the exceptional Lie group $\operatorname{G}_2'$. Our main result states that, if $P_k$ is the parabolic subgroup of $\operatorname{PO}(p, q)$ stabilizing an isotropic $k$-plane in $\mathbb{R}^{p, q}$, a group admitting a $\{P_1,P_q\}$-Anosov representation into $\operatorname{PO}(p, q)$ has virtual cohomological dimension strictly less than $p$. As a result, a $\mathbf{H}^{p, q-1}$-convex cocompact subgroup of $\operatorname{PO}(p, q)$ cannot act properly discontinuously and cocompactly on $\mathbf{H}^{p, q-1}$. The main theorem also implies that a group admitting a $P_{\alpha}$-Anosov representation into the exceptional Lie group $\operatorname{G}_2'$ has virtual cohomological dimension strictly less than $4$, where $\alpha$ is a short restricted root of $\operatorname{G}_2'$. As an application, we prove that the homogeneous space $\operatorname{G}_2'/ \operatorname{SU}(2, 1)$ does not admit any compact quotient.

Sensor-based beam prediction is usually evaluated against exhaustive beam training, leaving open when sensing adds useful information beyond a deployed tracker's link feedback and when it should be withheld. We separate system gain from the marginal value of sensing in action-restricted causal replay of recorded 64-beam sweeps from four vehicular DeepSense 6G scenarios. The replay exposes only serving-beam and explicitly probed candidate powers and counts probes and model invocations separately. The proposed controller combines a scheduled post-setup re-sweep with probe-verified sensing admission and blacklist backoff. In three scenarios where feedback-only tracking is within 0.25 dB of the recorded-sweep oracle, the controller stays close to feedback-only in mean loss while rarely invoking the model, whereas always-on sensing performs worse. In the sole high-innovation scenario, the controller's system gain is 0.82 dB (95% confidence interval [0.30, 1.40]) over the plain feedback-only tracker and 0.15 dB over a probe-limited re-sweeping reference, with fewer probes than either. Sensing and the re-sweep both correct setup lock loss. At the native decision interval, enabling sensing in the same controller yields no statistically resolved loss reduction but cuts probing by 21%. In this scenario, accurate sensing reduces loss relative to the sensor-disabled controller under emulated link noise, staleness or longer decision intervals. Weak or delayed sensing can instead increase loss despite probe verification. These exploratory results characterize sensing as a conditional substitute for RF probing or a source of loss reduction, depending on the limitations of link feedback and on the accuracy and timeliness of sensing.

Effective social interaction requires more than fluent responses: agents must reach plans that fit the participants' needs and circumstances. Existing evaluations of social agents mainly rely on LLM judges that rate the conversation, without checking the resulting plan. We introduce HICO-Bench, a benchmark of 180 everyday coordination scenarios, such as resource allocation, scheduling, and authorization, in which two agents with private information converse and then independently record the plan they agreed on. A program checks Feasibility (the two records match and satisfy both participants' private conditions), and an LLM judge checks Integrity (the conversation supports the recorded plan); passing both is a Verified Success. Each scenario has two versions that share the same public context but differ in one private condition, so the correct plan differs between them. Across thirteen LLMs, 36% of conversations rated highly for goal completion fail Verified Success. To learn from outcomes that arrive only at the end of a conversation, we propose Paired-Credit Policy Optimization (PaCPO). After each turn, PaCPO asks the agent for a tentative plan, checks it in both versions of the scenario, and rewards the turns after which the tentative plans fit their respective versions. PaCPO raises Qwen3-8B's Verified Success from 20.4% to 28.3%.

We reanalyze the total solar eclipse images taken with a consumer camera during the Artemis II lunar flyby, using the RAW data as ground truth for our previous analysis of the released JPEG image. Calibrated with field stars, the RAW data show that the lens vignetting and the variation of the pixel solid angle reduce the signal of diffuse light by up to $\sim$70% at the edge of the field, and that the signal on the dark side of the Moon is most likely foreground light rather than an instrumental offset. Between 10 degrees and 15 degrees elongation, the radial profiles of the F-corona follow power laws with indices of $-2.0$ to $-2.3$ (east), $-2.5$ (west), and $-2.4$ to $-2.7$ (north and south), consistent with space-based observations, and the surface brightness along the ecliptic is $\sim$26 MJy sr$^{-1}$ ($\pm 7\%$) at 10 degrees. The isophotes have flattening indices of $\sim$0.6-0.7 and superellipse indices of $\sim$1.5-1.6. Comparison with the ZodiSURF zodiacal light model supports a radial dust density index of $\alpha \simeq 1.3$ down to $\sim$0.2 au and suggests that the vertical distribution of the dust broadens toward the Sun. The JPEG image retains nearly the full vignetting, despite its metadata, and its tone response is not a single power law; together with the zero-point treatment, these effects account for the steep and asymmetric profiles reported previously, while the flattened morphology and $\alpha \simeq 1.3$ are confirmed. We identify quantities robust in JPEG images and show that the tone response can be reconstructed from field stars.

Software engineering agents have advanced rapidly, but scaling their training remains constrained by the limited availability of high-quality executable tasks. Development histories contain valuable signals for constructing such tasks, yet they often require substantial reconstruction due to incomplete environments, missing tests, and inconsistent artifacts. We present SWE-Stream, a system for scaling executable SWE training data through progressive construction and elastic streaming execution. SWE-Stream progressively validates and refines task candidates, reusing existing artifacts when possible and applying targeted completion only when necessary. An elastic scheduler asynchronously coordinates construction stages under shared resource constraints, adapting execution budgets based on observed workload demands. Using development histories from 10,728 repositories across six programming languages, SWE-Stream constructs 80,874 validated tasks and collects 118,688 successful trajectories. Fine-tuning Qwen3-30B-A3B-Thinking-2507 on these trajectories achieves resolve rates of 69.0%, 49.3%, and 45.42% on SWE-bench Verified, Multilingual, and Pro, respectively. In a controlled comparison with approximately matched training budgets, SWE-Stream achieves a 45.6% resolve rate on SWE-bench Verified, outperforming three existing data sources. Moreover, its streaming construction strategy improves efficiency, delivering $1.45\times$ the accepted-task throughput of batch construction and accepting 13.2% more tasks than fixed-limit streaming.

We identify a normalization defect in a determinantal construction of fourth-order Hahn difference equations and compare its scalar equation with Foupouagnigni's 1998 elimination and its published 2001 counterpart. On an explicit positive orthogonal family, all five prescribed coefficients vanish: the formula gives \(0=0\) and imposes no condition. Along an explicit regular perturbation, division by \(\varepsilon^2\) before specialization recovers a nonzero equation on that family, the same scalar equation as the correctly executed old shift-and-eliminate procedure. On the common nondegenerate domain specified below, both constructions give the same scalar operator up to a nonzero rational factor. The first-order degree bounds follow directly from valid identities of 1998; preserving cancellations gives a bound that can be arbitrarily sharper on explicit admissible inputs, together with exact top coefficients and cancellation criteria. Explicit coefficient formulas do not by themselves yield a new scalar equation on this domain. Our comparison does not exclude corrective or expository contributions; additional computational gain or exceptional coverage must be demonstrated.