Papers

The recent observation of the high-energy nuclear recoil event LZ230616 with a recoil energy around 248 keV provides an interesting possibility to probe inelastic doublet dark matter (iDM). However, the electroweak doublet DM scenarios face strong constraints from solar DM capture for DM masses around the TeV scale which give rise to correct thermal relic density. In this work, we revive the inelastic doublet DM (iDM) scenario via type-II seesaw. In particular, we consider an inert lepton doublet (ILD) DM. While the minimal ILD scenario gives the correct relic abundance around the TeV scale, it is strongly constrained by direct detection and solar DM capture. In the type-II seesaw, the presence of the scalar triplet generates the required mass splitting and provides additional annihilation channels, allowing the correct relic abundance for much larger DM masses. The larger DM mass also helps evade the solar capture and indirect detection constraints. Additionally, the scalar triplet gives sub-eV neutrino masses required by the oscillation data.

Automatic dense packing is widely desired in warehouse operations but remains a fundamental challenge in robotic manipulation. Existing work on irregular-object packing largely targets simulation with idealized contact, treating the object as an isolated rigid body. The gripper often enters as a discrete, post-hoc feasibility check, if considered at all, and the perception and contact drift accumulated during execution are not addressed. We present a closed-loop pipeline that integrates perception, gripper-aware placement optimization, and force-guided execution on a real manipulator. The optimizer represents the object together with the gripper as a single composite body of hierarchical sphere trees. It searches over five degrees of freedom on a GPU within a CMA-ES framework, with the vertical coordinate grounded analytically against the current heightmap. During execution, a force-monitored vertical descent stops on first contact. A post-release consolidation push then closes the residual lateral clearance that gripper-aware planning leaves behind. The container is re-perceived between placements so that drift does not accumulate. We validate the system on a Franka Emika Panda robot packing a 3D-printed set of flat, curved, and concave objects, and a YCB object subset. An ablation study isolates the contribution of gripper-aware optimization, the consolidation push, and mesh-derived geometry to end-to-end success, achieved density, and computational cost. We further benchmark against the heightmap-minimization method as a baseline representative of prior irregular-object packing work.

Cellular embeddings of graphs in surfaces have well-defined duality and minor (edge contraction and deletion) operations that interact in a natural way. A pseudosurface is obtained from a surface (compact 2-manifold) by a finite number of identifications of finite sets of points. Points that are created by the identifications do not have a neighborhood homeomorphic to an open disk and are known as pinchpoints. Embeddings of graphs in pseudosurfaces have been considered, both implicitly and explicitly, since the 1960s. Usually the condition that all pinchpoints correspond to vertices of the graph is imposed. However, this makes it difficult to define duality and minors for pseudosurface embeddings and have these operations interact in the expected way. We define the class of pseudocellular embeddings of graphs in pseudosurfaces, which allow pinchpoints at places other than vertices, in particular in the middle of faces or edges. A subclass known as quasicellular embeddings corresponds to previous embedding models due to Deneen, Shute, and Thomborson and to Huggett and Moffatt. Pseudocellular embeddings also generalize other structures, including the edge-point ribbon graphs of Ellis-Monaghan, Kauffman, and Moffatt, and cyclically ordered graphs or cogs (also known as rigid-vertex graphs). Duality and minor operations for pseudocellular embeddings have very simple and straightforward definitions using topological quotient operations. Pseudocellular embeddings of edgeless graphs have nontrivial structure, and we define some minor operations for those that are related to `t-minor' operations on bipartite graphs. We develop a family of polynomial invariants for pseudocellular embeddings and discuss connections to other polynomial invariants.

Traffic sign recognition (TSR) is an important perception task for autonomous driving and advanced driver-assistance systems, where a system must both localize traffic signs and determine their semantic classes efficiently. This work presents a TSR system based on YOLOv2 for simultaneous detection and classification. Two complementary modifications are studied. First, YOLOv2 is extended with intermediate prediction layers, forming a branched architecture that can terminate inference early for easy cases and reduce computation time. Both whole-image and cell-wise branching strategies are investigated. Second, geometric information is introduced to reduce classification errors between visually similar signs. An unsupervised Bayesian image-segmentation method produces binary representations that are compared with class-specific geometric templates inside YOLOv2 bounding boxes. This information is used either during inference or as an additional signal during training. A dedicated dataset is constructed by combining GTSDB and GTSRB samples using seamless cloning and controlled image transformations. Experiments cover ten traffic-sign classes, with 3,000 training and 300 test samples. The selected branched architecture reports 0.647 s runtime and 0.680 mAP, compared with 0.6607 s and 0.680 mAP for baseline YOLOv2. Geometric verification during inference increases mAP to 0.713, while the geometric-feature training variant achieves 0.697 mAP with a reported runtime of 0.6608 s.

For every integer $d\ge4$, we prove strong stability for a subfamily of classical rotational hypersurfaces in $\mathbb H^d$ with normalized mean curvature one. The examples are complete, two-sided, properly embedded, and nowhere umbilic, with topology $\mathbb {R}\times\mathbb{S}^{d-2}$. An explicit positive supersolution yields a quantitative stability inequality for all compactly supported test functions. Consequently, endpoint horospherical rigidity fails in every ambient dimension at least four.

We show that there are integers $N$ and odd $D$ such that no algorithm can decide, from the rational coefficients of a homogeneous polynomial vector field $F$ in dimension $N$ of degree $D$, whether the origin is Lyapunov stable for $\dot Y=F(Y)$. This proves a conjecture of V. I. Arnold. We provide a Lean formalization of the proof. The smallest dimension $N$ for which we were able to prove undecidability is $N = 5$. An analogous undecidability result holds for global asymptotic stability and global exponential stability for non-homogenous vector fields. For dimension $N=2$ and homogenous vector fields, we establish decidability for all commonly used stability notions, building heavily on existing results.

Halo models provide the standard framework for interpreting cosmic infrared background (CIB) anisotropies and inferring the connection between dusty star-forming galaxies and their host dark matter halos. Recent studies have shown that inaccuracies in the modelling of halo clustering may bias the inferred parameters governing star formation in dark matter halos. In this work, we investigate whether a more detailed description of halo clustering can alleviate these shortcomings. We incorporate a scale-dependent, non-linear correction to the halo bias within a halo-model framework based on a simple SFR-halo mass parametrisation. We consider both a full mass-dependent implementation and a computationally efficient approximation. We also investigate the impact of updated (sub)halo mass functions. We validate the revised framework using mock observations derived from a dedicated simplified version of the SIDES-Uchuu simulation, combining the SIDES empirical galaxy model with the Uchuu N-body simulation, specifically designed to match the SFR prescriptions adopted in our halo-model. The corrections modify the predicted CIB clustering by up to 30% on intermediate scales, while the effective implementation reproduces the full treatment at the sub-percent level. The updated model removes the scale-dependent discrepancies previously identified and accurately reproduces the measured power spectra over the full multipole range. In MCMC analyses, the input SFR parameters are recovered with substantially reduced biases, with all input parameters recovered within their $1\sigma$ confidence intervals. Our results demonstrate that accurately modelling non-linear halo clustering is essential not only to reproduce CIB anisotropies, but also to reliably recover the underlying galaxy-halo connection. The framework presented here provides a validated foundation for future analyses of observational data.

Multi-hop retrieval failures are not uniformly distributed across queries: they cluster in structurally predictable subpopulations. We prove two results formalizing this structure. First (CWAR Reducibility): confident-failure reduction is achievable if and only if retrieval features carry mutual information about success, a condition satisfied by LLM-judge pipelines but substantially weaker in dense-only settings, explaining the AUC-AC gap between regimes. Second (Feature Regime Complementarity): no single ANN score feature achieves best predictive performance across all failure regimes; the dominant feature differs between datasets (query length on MuSiQue, hop-1 concentration on HoVer), and a constructive witness pair shows each is necessary in one regime and non-contributory in the other. We instantiate these principles in RegimeAbstain, which computes a Retrieval Confidence Score (RCS), a logistic function of up to nine query-ANN structural features, all available without any additional LLM call, and uses it to implement a calibrated abstention policy. We define the Confident-Wrong-Answer Rate (CWAR) metric and evaluate across three multi-hop benchmarks (MuSiQue, 2WikiMultiHopQA, HoVer) and two retrieval architectures (LLM-judge and dense-only), covering five failure regimes with CWAR from 14.5% to 62.1%. RCS achieves best or co-best AUC-AC in all five conditions against eight confidence baselines. On MuSiQue (LLM-judge), RCS reduces CWAR from 39.5% to 20.6% at 50% coverage (47.8% relative reduction), with ECE=0.035. A model trained on MuSiQue transfers to 2WikiMultiHopQA with only -0.5pp AUC loss, confirming the domain-agnostic structure of regime features.

A desirable property of a world model is the ability to learn continually across tasks, adapting to new environments without forgetting what the agent has already learnt. In particular, the ability to retain and reuse knowledge obtained from prior experiences underpins an agent's ability to efficiently adapt to novel environments, as the dynamics of the physical world can often be described in recurring mechanisms. However, the world model's measure of adaptation entangles two abilities: the speed and capacity to learn unseen tasks, and the reuse of knowledge already acquired, since incoming tasks carry novel content alongside what recurs. In order to isolate knowledge reuse from prior experiences, we propose a compositional continual learning benchmark for world models in robot manipulation. Specifically, we design each task curriculum with compositional tasks that combine aspects of the tasks seen in the sequence. We further factorise this composition along the axes of action and perception to better understand how different input modalities bottleneck knowledge reuse. We evaluate state-of-the-art world models under canonical continual learning methods, alongside a modular world model whose dynamics backbone contains explicitly reusable components. Results show that modularity balances reuse against forgetting better than conventional methods, but none solve the problem fully, leaving clear room for continual world models built to reuse without forgetting. More details are available on our project website: https://object814.github.io/Compositional-Continual-Learning/.

We prove that, over any field, the bilinear complexity of multiplying a $3\times 2$ matrix by a $2\times m$ matrix is strictly greater than $24m/5$. In particular, every exact bilinear algorithm for multiplying a $3\times 2$ matrix by a $2\times 5$ matrix requires at least $25$ multiplications. Together with the Hopcroft-Kerr upper bound, this proves that the $\langle 3,2,5\rangle$ matrix multiplication tensor has rank exactly $25$. The proof has been formally verified in Lean 4, with the formalization available at https://github.com/fallnlove/mm325_proof.

Graph Convolutional Networks (GCNs) are highly sensitive to label noise, since corrupted supervision can propagate through the graph and degrade learned node representations. This work proposes PCC+GCN, a hybrid framework that uses Particle Competition and Cooperation (PCC) as a graph-based label-refinement stage before GCN training. PCC identifies suspicious labeled nodes through particle domination dynamics and determines whether their labels should be preserved, removed, or reassigned before GCN training. The framework also allows the graph used by PCC to be augmented with feature-based $k$-nearest-neighbor edges, while the GCN itself is trained on the original graph structure and node features. The proposed method was evaluated on ten graph datasets from the NoisyGL benchmark under conventional Uniform, Pair, and Random label noise, as well as under instance-dependent label noise. A detailed hyperparameter analysis was also conducted on Cora, CiteSeer, and PubMed. Under conventional noise, PCC+GCN achieved the highest overall average accuracy and the best average rank among the evaluated methods, with an average gain of $1.67$ percentage points over the baseline GCN across the clean setting and all noisy scenarios. Under instance-dependent noise, PCC+GCN remained competitive with the best-performing robust methods while requiring substantially lower execution time, being the fastest robust method on eight of the ten datasets. The results indicate that PCC-based label refinement provides an effective and computationally efficient preprocessing strategy for improving GCN robustness under noisy supervision.

Catastrophe (cat) bonds overcome some lack of reinsurance by sourcing capacity from the wider capital markets. We present a new type of cat bond addressing the known trade-off between moral hazard and basis risk. As our main contributions we propose a trigger mechanism which is entirely transparent and simpler to evaluate compared to indemnity modeling techniques, as well as a methodology to price this cat bond. This is relevant for insurers and public authorities in a world where natural disasters are occurring with increasing frequency and severity due to climate change, but also for players willing to enter the cat bond market for whom the lack of transparency of this asset class has been a significant obstacle. Our trigger is derived from a cost random field which separates the physical hazard, a vulnerability function and the exposure. This allows the trigger to take a flexible form between parametric and modeled loss, in case exposure is taken into account. We present a case study based on historical windstorm events impacting Germany. Using wind speed data from historical storms, we fit a max-stable random field on a resolution which is standard in the reinsurance industry. The availability of industry loss and exposure data allows us to calibrate the vulnerability component to historical observations. Besides measuring the basis risk associated with our trigger, we perform a full model assessment and discuss numerical results.

Motivated by experimental studies on cuprate-manganite heterostructures, we study a correlated bilayer Hubbard model subject to a local interfacial exchange field using cluster dynamical mean-field theory in the underdoped cuprate regime. We find that superconductivity coexists with a manganite-induced Cu spin polarization. Remarkably, throughout this coexistence regime, the superconducting solution carries a larger magnetic moment than the corresponding constrained normal state. Analysis of the magnetic response shows that this reversal is driven by correlations, which enhance magnetism already in the normal state and even more strongly in the superconducting solution. At stronger exchange fields, superconductivity collapses abruptly, displaying hysteresis consistent with a first-order transition, and a strongly spin-polarized pseudogap emerges. These results show that, under strong electronic correlations, superconductivity can enhance rather than suppress interfacial magnetism.

We develop a representation-theoretic framework for computing projection constants of finite-dimensional subspaces of $C(K)$, where $K$ is a compact homogeneous $G$-space. Within this framework, we characterize the finite-dimensional $G$-invariant subspaces for which the $G$-equivariant projection is unique. These are precisely the finite orthogonal sums of full isotypic components in the Peter–Weyl decomposition of $L_2(K)$. For such subspaces, averaging shows that this unique $G$-equivariant projection is minimal; it is the restriction to $C(K)$ of the $L_2(K)$-orthogonal projection. Consequently, their projection constants are given by the $L_1$-norm of an explicit reproducing-kernel slice. We apply this method to spaces of $d$-homogeneous polynomials in high dimension. The examples range from Fourier analysis on the torus, through spherical harmonic analysis on real and complex Euclidean spheres, to genuinely noncommutative harmonic analysis on the unitary group, where representation theory becomes indispensable. For all these families, we obtain precise high-dimensional asymptotics at the square-root-of-dimension scale.

Generative AI has entered research programming, yet there is little evidence about which tasks researchers hand to it or how they decide whether its code is correct. We draw on 527 free-text responses to a 2025 survey of researchers who write code, most of them at U.S. universities. In each response, a researcher recounts a single task from their own work, the way they used an AI tool for it, and what they did to assess the result. We coded the task and the evaluation strategies reported, and related both to programming experience, research area, and confidence ratings. Use was concentrated in five tasks: data handling, visualization, debugging, mathematical/scientific computing, and statistical analysis. Evaluation was informal and individual. Over half of accounts described running the generated code, while automated tests and review by another person were rare. Use cases and evaluation strategies varied little with programming experience, but confidence did: Less experienced programmers trusted the AI more than themselves, and experienced programmers the reverse. Evaluation confidence was not associated with the strategies reported. Its strongest correlates were confidence in the tool and in oneself. Validating AI contributions to scientific code rested largely on individual judgment, outside shared infrastructure for testing or review. Interfaces could support task-appropriate evaluation rather than leave it to the user.

A selective predictor acts as a safety gate: it returns an output only when the prediction appears sufficiently trustworthy. Deployments increasingly require this reliability to be certified at a target precision for every reporting unit of interest, such as a tool, policy label, or patient subgroup. The main difficulty is often not whether a granted certificate is valid, but whether finite calibration data can produce one at all. As the gate becomes safer or more fine-grained, some units may receive too little evidence to certify. We make this notion of availability computable through classical exact-binomial inversion and formulate reporting-partition selection, under a fixed group order, as a dynamic program that exposes the trade-off among safety, granularity, and served traffic. The resulting frontier reveals a large population opportunity that finite-sample estimation nearly erases: a truth-informed planner gains $0.157$ mean coverage over support balancing, whereas a naive estimator recovers only $0.005$, making recovery from finite data the central challenge. Constructing candidate partitions on one planning split and selecting among them on another recovers part of this gap, improving mean coverage over support balancing by $0.060$, with the direction reproduced in $59$ of $60$ model effects across three intent-routing datasets and two architectures. A complementary validity-preserving lever, reallocating the familywise error budget across reporting units, recovers additional coverage both with population quantities and noisy estimates. The same frontier recurs, with predictor-specific ceilings, across LLM tool-calling, content moderation, lesion classification, and recommendation. Certified availability is therefore a plannable deployment resource that determines when a safety gate can be certified, at what granularity, and over how much traffic.

In this work, we study asymmetric $(1\to2)$ quantum cloning for pure states belonging to orbits generated by the orthogonal group. For every dimension $(d\geq3)$, we determine the full region of achievable pairs of average single-copy fidelities. Using orthogonal covariance and the Brauer algebra, we reduce the optimization over all CPTP maps to a finite-dimensional spectral problem and construct explicit optimal cloning channels. We then apply the general results to four-mode fermionic states in the even-parity sector. Using triality, we identify the fixed-concurrence fermionic families with the orthogonal orbits considered in the first part of the paper. In particular, for pure Gaussian states we show that the symmetric locally optimal cloner is unique. We also compare local and global cloning and find that the channel optimal for the joint two-copy fidelity is different from the one optimal for the single-copy fidelities. Finally, we show that the locally optimal cloner cannot be implemented using only operations which preserve fermionic Gaussian states.

Quantum sensing uses nonclassical states to improve measurement sensitivity, but the same states that provide metrological gain also decohere more rapidly. This limits the usable interrogation time and, in practice, often precludes improvement in ultimate sensitivity - realized useful quantum advantage has consequently remained rare. Here, we restore persistent quantum advantage by embedding Fock-state enhancement within a quantum heterodyne (Qdyne) protocol, decoupling sensitivity from the decoherence-limited interrogation time {\tau}. Measurements are acquired at short {\tau} where the quantum-enhanced gain is optimal, while precision accumulates with the total measurement time. Demonstrated on the motional mode of a trapped 40Ca+ ion, we observe quantum-enhanced precision that persists to measurement times seven orders of magnitude beyond the dephasing limit, scaling as T^{-3/2} with no indication of saturation. Using the n=3 Fock state, we reach a frequency precision of 0.5uHz relative to an 86MHz carrier, achieving a fractional precision ~6x10^{-15}. This represents a quantum-enhanced gain of 7.1(10) dB over the n=0 state, in agreement with Fisher information predictions. This is the first demonstration of Qdyne beyond solid-state spin-defects. Further, by using the quantum harmonic oscillator to perform frequency mixing, we extend operation beyond 1GHz, two orders of magnitude above the ceiling of pulsed dynamical-decoupling implementations. These results recover quantum advantage at the long timescales required to improve ultimate sensitivity, with direct implications for nanoscale NMR and quantum logic spectroscopy.

Impact rates predicted using pre-impact source-population models and those inferred from observations disagree by a factor of a few to $\sim200$ across the Solar System planets and the Moon. An exciting opportunity for an additional observational constraint is the discovery of asteroids before impact and their tracking until impact. We evaluate this prospect based on both known objects and forecasts of the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST). First, we identify impacts from known objects, confirming multiple past cases and predicting two coming impacts: 2015 FK488, a $\sim1.8$ km Centaur whose nominal orbit predicts an impact with Jupiter in $\sim200$ yr, and 2022 KG1, a $<10$ m NEO recently removed from the impact-risk lists but whose nominal trajectory leads to an Earth impact in $\sim260$ yr. Both orbits are poorly constrained. Second, assuming the model-based impact rate, we predict that LSST will discover 7 of the 59 synthetic impactors quantified in our previous analysis. All of the discovered impactors will impact Jupiter, and they have a size range of $\sim200-2000$ m and expected discovery-to-collision lead times of $1.94-257.4$ yr. Two of them are discovered inside Jupiter's Hill sphere, offering rare opportunities to observe the transition from a temporary bound orbit to impact before it happens. If instead we assume the observation-inferred impact rates, LSST will at most discover $\sim20$ future Earth impactors ($>10$ m) and $\sim300$ future Jupiter impactors ($>10$ m), with a wider range on both ends for discovery-to-collision lead times. These results demonstrate that LSST can identify a subset of planetary impactors years to centuries before collision, providing a new observational pathway for testing source-population models and the impact-rate mismatch.

Planetary impact rates are difficult to infer from observations alone because impacts are rare, observational records are target-dependent, and converting observed flashes, bolides, and craters to an intrinsic impact rate depends on uncertain selection effects and conversion parameters. Building on the simulations in Paper I, we examine whether observational systematics can explain the order-of-magnitude mismatch between observed and modeled impact rates on Earth, and whether comparable discrepancies extend to other Solar System bodies. We convert 12 observed impact records for Earth, the Moon, Mars, and Jupiter to a common pre-atmospheric impact rate for $D>10$ m, propagate uncertainties through the conversion chain, and compare the resulting rates with the intrinsic rates derived in Paper I. We find discrepancies for Earth (observation-to-model median ratio $2.2-21.0$), the Moon ($3.4-132.1$), Mars ($14.2-118.6$), and Jupiter ($2.6-47.7$), with the dominant source of the discrepancy differing by body. The simulated Jupiter-family comet (JFC) contribution substantially reduces the Earth discrepancy, suggesting that an additional dynamical source may help explain the gap. The lunar records disagree with each other by two orders of magnitude, with the discrepancy dominated by size-extrapolation uncertainties. The Mars mismatch is dominated by crater-to-impactor conversion. The Jupiter comparison depends mainly on uncertainties in the inferred number of bound impactors and in the observational completeness thresholds. Further investigation of this mismatch requires incorporating cometary modeling and tighter constraints on size-extrapolation scaling and improved observational completeness.

Memory systems for large language models have focused predominantly on efficient retrieval, whereas the decision of whether retrieved memories should be trusted has received comparatively little attention. When the memory store contains conflicting positions, standard retrieval-augmented generation (RAG) blindly injects memories and amplifies hallucinations: in models susceptible to memory injection, the RAG hallucination rate under conflicting memories is markedly higher than that of a memory-free baseline. Inspired by memory signaling mechanisms in the prefrontal cortex, we propose the Memory Decision Layer (MDL), a zero-parameter memory decision controller situated between the retrieval and generation stages. Its core is a three-signal complementary encoder that fuses relevance, reliability, and task risk through QR-based orthogonal subspace projection and a meta-working-memory signal into an interpretable decision representation that quantifies the trustworthiness of retrieved memories. Building on this encoder, MDL explicitly decouples confidence from consistency and introduces risk inversion and explicit abstention. Evaluations on mainstream large language models and multiple open-source datasets show that MDL reduces the hallucination rate under conflicting memories by about 56.04% in general scenarios and approaches zero hallucination in high-risk scenarios. The controller is fully white-box: it relies purely on geometric operations, requires no trained parameters, and adds only about 0.14 ms per decision – roughly 50x faster than the embedding-retrieval step that precedes it and four to five orders of magnitude faster than an LLM self-evaluation call.

Previous studies have shown a mismatch between the simulated rate of asteroids impacting the Earth and the observed rate, posing an important problem for planetary defense. Extending this analysis to other planets offers a new opportunity to examine minor body populations and impactor models. We present a unified simulation framework that estimates intrinsic $D\gtrsim10$~m impactor rates on the eight planets and the Moon. We construct size-calibrated source populations of Near Earth Objects (NEOs), Main Belt Asteroids (MBAs), Jupiter-family comets (JFCs), Centaurs, and scattering trans-Neptunian objects (TNOs), and propagate them with a 300-year $N$-body simulation. To avoid relying on black-box-like collision flags, we identify close encounters and estimate impact rates in two complementary ways: a direct count of realized impactors from object-level minimum-distance calculation and a collision expectation based on population-level impact parameter statistics. We predict over 300 years there will be $6.0^{+3.8}_{-2.7}$ impactors on Venus (all NEOs), $10.0^{+11.3}_{-4.9}$ on Earth (NEOs and JFCs), $\sim$$2981^{+481}_{-469}$ on Jupiter (mainly MBAs and JFCs), $1228^{+2759}_{-1017}$ on Saturn (mainly Centaurs), and none on the remaining planets with the $1\sigma$ upper limits spanning $<1.8$ (inner planets) to $<1\times10^{8}$ (outer planets; weak constraint). Jupiter uniquely receives both fast unbound impactors ($>50\, \mathrm{km\,s^{-1}}$) and slower bound ones reaching down to $\sim$$20\,\mathrm{km\,s^{-1}}$. In a companion paper, we compare our predictions with observations and test whether the gap extends across the Solar System and whether current telescopes can observe these collisions.

Reinforcement learning is increasingly used to align image generators with reward signals, and Flow-GRPO recently extended this paradigm to flow-matching models by treating the denoising sampler as a stochastic policy that can be optimized from reward feedback. Training in this setting is unstable in a way specific to multi-step denoising: the policy update changes systematically across denoising steps, with importance ratios drifting below one, becoming increasingly dispersed, clipping at different rates, and leaving fewer usable samples late in training. Prior work treats these effects as separate failure modes and addresses each with a hand-tuned stabilizer. We show instead that they arise from a single per-step quantity, which we call path variance. This quantity is determined exactly by the sampler's Gaussian transition kernel and can be estimated cheaply during training. This reframes instability as a resource that can be measured and budgeted rather than a collection of symptoms to repair. Our method, $\lambda$-Controlled GRPO, calibrates importance-ratio behavior from this predicted law rather than from noisy empirical statistics, and allocates gradient effort across denoising steps according to their predicted cost. The two scales governing the update are fixed by standard policy choices rather than introduced as free tuning parameters. On a text-to-image model under two reward settings, rendering difficult target text scored by optical character recognition and matching human preferences scored by a preference model, $\lambda$-Controlled GRPO improves both text accuracy and preference reward over the strongest empirical stabilizer. It also keeps late-step path variance within its intended budget, precisely where the baseline systematically overshoots. The result is a Flow-GRPO update calibrated by its own transition law rather than stabilized after instability appears.

Current Vision-Language-Action (VLA) models for autonomous driving operate primarily through feedforward inference across the perception–reasoning–planning hierarchy. While modern architectures maintain temporal recurrence within the perceptual module, early perception remains blind to downstream reasoning and navigation goals, processing visual inputs agnostically without prioritizing cues informed by prior decisions. To bridge this gap, this paper introduces PRIME, a learned feedback mechanism that conditions the VLA perceptual queries on a novel Situational Memory. By aggregating latent representations of past perception, reasoning, navigation goals, and predicted behaviors across an L-step window via cross-attention, PRIME enables intent-driven perceptual attention at minimal computational cost, adding only a maximum of 29.7M parameters (0.41% of the 7.3B-parameter base model). Evaluated on the Bench2Drive closed-loop benchmark, PRIME achieves a state-of-the-art Driving Score of 82.47 (+4.73 over ORION) and a Success Rate of 60.00% (+5.38 percentage points), the highest reported Driving Score among published VLAs trained on Think2Drive demonstrations.

We need studies on conversational AI (CAI) at scale to understand human behavior and shape CAI design. However, fragmented reporting of systems and study configurations hinders replication, extension, and knowledge accumulation. We present Gricea, an open-science platform representing studies as configurable, deployable research artifacts that researchers can run, inspect, share, and reuse. Informed by a formative analysis of prior CAI research, Gricea couples study procedures, participant-facing systems, and conversational task behavior in. In a replication study using Gricea, we replicated configurations 93% of eligible CUI 2026 papers; while also flagging missing information in 96% of papers that hinder faithful replication — further motivating Gricea's need. In a user study, researchers and practitioners from diverse backgrounds successfully constructed runnable studies addressing various open-ended research questions. Together, these findings demonstrate Gricea's support for constructing, reproducing, and extending CAI studies through shared research artifacts, enabling cumulative knowledge building through open science.

We introduce QuranicMMLU, a benchmark for evaluating generative AI on Quranic Arabic across multiple dimensions of linguistic complexity. Existing Quranic benchmarks center on general question answering and semantic retrieval, without probing specific linguistic competencies or stratifying by cognitive demand and verse difficulty. We construct a five-pillar Quranic taxonomy spanning Phonology, Morphology, Syntax, Semantics, and Pragmatics, with 31 leaves covering phenomena from tajwīd and root-and-pattern morphology to occasions of revelation and inter-surah coherence. For each leaf we generate questions stratified by Bloom's cognitive level and verse perplexity, then have LLM as a judge to independently answer and score every item and route the annotations to manual review. The resulting dataset comprises 980 human-reviewed questions, each issued in both open-ended and multiple-choice form. We benchmark 12 systems on these items and find that the Islamic-specialized model leads, yet every system scores higher on multiple-choice accuracy (average 84%) than open-ended answer quality (average 60%): the two rankings agree closely (Kendall's {\tau}=0.73), but multiple-choice scoring hides failures that surface only once answer choices are removed. QuranicMMLU thus offers a rigorous, linguistically grounded framework for evaluating Arabic NLP in the Quranic domain.

The large-scale integration of renewable energy into electrical power systems places a growing reliance on grid-forming converters (GFMs) to overcome critical stability and control hurdles. By emulating a slowly varying voltage source behind a tunable resistive?inductive impedance, GFMs enable precise setpoint tracking and vital grid support. In this regard, virtual admittance (VA)-based inner-loop control provides exceptional flexibility in shaping the converter's small-signal immittance at connection terminals, allowing for advanced GFM functionalities. This article addresses a novel internal-model-control-based VA approach, which inherently accounts for the effects of the inner current-control loop and the voltage feedforward. Unlike the state-of-the-art VA control, the proposed VA control ensures that, at frequencies outside the influence of outer loops, the converter's terminal admittance closely matches the target admittance, which is shown to be relevant for stability. It is revealed how, with the state-of-the art VA control, a small-signal instability in the harmonic range may arise if a low-pass filter with a high cut-off frequency is used for the voltage feedforward. However, the proposed VA control ensures stable operation regardless of this cut-off frequency. The methodology is experimentally validated using a laboratory prototype that features three-phase voltage-source converters.

Slow positron beams enable diverse applications, from surface-sensitive materials studies to positronium physics, but progress is limited by source intensity and brightness. In linac-based sources, mono-energetic, slow positrons are produced by moderating the broad, divergent distribution of fast positrons produced by high-energy electrons incident on a high-Z target. This process is intrinsically inefficient. Conventional linac-based designs place the moderator close to the target. Tungsten moderators become less efficient when heated by high-power, fast positrons, leading to defect-related losses. Cryogenic moderators, like solid neon, melt when exposed to high-power, fast positrons. Following the method of O'Rourke et al., we use Monte Carlo simulations combined with a diffusion model to investigate how moderator geometry, material, and incident fast positron energy affect slow positron production. We compare single tungsten foil and multi-foil configurations with solid neon moderators in reflection and transmission geometries. We find that for fast positron energies below 300 keV, solid neon offers order-of-magnitude higher efficiency than tungsten due to its larger diffusion length, whereas at MeV-scale energies multi-foil tungsten has higher efficiency due to the maximization of scattering. We also find that including a target proxy in simulation increases low-energy efficiency up to fourfold by allowing a fraction of initially reflected positrons to return to the moderator. We conclude that substantial gains in start-to-end efficiency at linac sources will likely require combining a decelerating cavity with a decoupled cryogenic moderator.

We study incidence problems for sticky sets of tubes. In the first part, we prove a sharp Furstenberg set estimate for sticky sets of tubes in $\mathbb R^3$ satisfying convex Wolff axioms. In the second part, we prove the Kakeya set conjecture in $\mathbb R^4$ for sticky sets of tubes satisfying convex Wolff axioms.

We compute the gluonic contributions to the three-loop heavy-quark form factors for the vector, axial-vector, scalar, and pseudoscalar currents. In the low-energy limit, $q^2/m^2 \rightarrow 0$, we used guessing algorithms to derive closed-form difference and differential equations from the rational sequences associated with the multiple zeta values and other constants appearing in the expansion coefficients, which required up to 26000 coefficients in the most demanding cases. Part of the results are obtained analytically in terms of harmonic polylogarithms and square-root valued iterated integrals by solving the obtained differential equations. For the remaining contributions, arbitrarily deep series expansions around the singularities of the form factors can be obtained by matching local expansions at intermediate points and by exploiting the differential equations obeyed by the functions associated with each transcendental constant in the expansions around $s = q^2/m^2 = 0$. For all these calculations advanced computer algebra methods have been employed. By using analytic continuation methods for differential equations matching the expansions at different values of $s$ and using PSLQ, we derive analytic results in the high-energy limit, $q^2/m^2 \rightarrow \infty$. The expansion coefficients are expressed in terms of multiple zeta values up to weight $w = 6$ together with three additional constants, two related to sixth-root-of-unity letters and one associated with quadratic-form implied iterated integrals. We also derive deep expansions about the threshold and pseudo-threshold. Numerical results are presented in the whole kinematic range and compared to results in the literature.