Papers

Large language models (LLMs) are increasingly used in binary decompilation to refine the C-like pseudocode produced by traditional rule-based decompilers. While this pseudocode is useful, it is a heuristic and lossy abstraction rather than a faithful copy of the source code. It often contains decompiler errors, especially for aggressively optimized binaries where critical low-level details are obscured. We present ASAP, an assembly-source aligned pseudocode refinement framework for binary decompilation. ASAP learns source-aligned assembly representations from paired source and binary functions using joint function-level and snippet-level contrastive alignment. A Q-Former then compresses chunk-level assembly features into a fixed number of assembly tokens that condition the decompilation LLM alongside the decompiler-produced pseudocode. During refinement, we use stochastic pseudocode masking and a relative assembly-advantage loss to reduce the model's tendency to ignore assembly features and rely only on pseudocode refining. On two decompilation benchmarks across multiple compiler optimization levels, ASAP improves the average re-execution rate from 64.7% to 71.9% and the average recompilation rate from 91.6% to 96.6% compared with the strongest baseline, offering both a new perspective and a practical solution to binary decompilation.

Catastrophe bonds have become a central instrument for transferring catastrophic risk to capital markets, yet their payout structures are typically specified exogenously rather than optimized. In this paper, we treat the indemnity function as the decision variable and study the optimal design of CAT bonds from the sponsor's perspective. We establish the optimality of a piecewise linear indemnity function under both expected-disutility and spectral-risk-measure frameworks and thereby reduce an infinite-dimensional design problem to a one-dimensional optimization over the attachment point. The optimal design also eliminates the potential moral hazard associated with the discontinuity of threshold payout structures. We characterize the optimal attachment point under six risk criteria and obtain explicit or semi-explicit solutions. We further derive bounds for the optimal attachment point and develop comparative statics that provide practical design insights. Using U.S. severe storm losses from NOAA's Billion-Dollar Weather and Climate Disasters database for calibration, we show that the optimal CAT bond designs achieve substantial risk reductions across most risk criteria.

Reinforcement learning from verifiable rewards (RLVR) has produced large reasoning gains in language models, and verifiable video benchmarks make the recipe directly applicable to causal-temporal video question answering. We present a controlled study of what RLVR actually teaches video-language models about temporal structure. We fine-tune four open models with group relative policy optimization (GRPO) under three data recipes using annotated synthetic questions, self-supervised real-video pretext tasks, and a one-to-one mixture, plus a targeted real-video augmentation arm. We evaluate the model-recipe grid on 105 CLEVRER questions with one correct option and on two real-video transfer sets. The 8B ordered gain is 4.8 points (95% video-bootstrap CI [-5.6, 15.1]), while blind accuracy increases 10.5 points ([0.9, 20.6]). The 4B model gains an accuracy advantage for ordered over shuffled frames. On the real-video sets, the 8B synthetic-only checkpoint loses 26.7 and 25.2 points, with smaller regressions under mixed and augmented recipes. In a separate scene-disjoint, single-gold Qwen3-VL-8B experiment, supervised fine-tuning (SFT) and GRPO each complete 1,200 updates at three fixed seeds. GRPO minus SFT is -3.89 points in ordered accuracy (nominal 97.5% video-bootstrap CI [-7.36, -0.45]) and -2.78 points in ordered-minus-shuffled accuracy ([-6.60, 0.87]). Transfer differences are sensitive to parseability. Intervals quantify test-video uncertainty conditional on the fitted runs. The results show why benchmark accuracy, transfer, and presented-order sensitivity need to be evaluated separately.

This paper examines the post-2010 expansion of the Hungarian construction sector as a case study of large-scale government intervention in a single market. A political economy framework conceptualises the intervention as two mutually reinforcing policies - expansionary demand (public procurement and housing subsidies) and preferential allocation towards politically connected firms - operating under inelastic supply through two mechanisms: market-level subsidy capitalisation and firm-level rent extraction. Using a structured descriptive case study design, the paper draws on Hungarian Central Statistical Office and Eurostat macroeconomic statistics, firm-level data on Hungary's largest companies (CrefoPort), policy documents, and a regional comparison to assess four propositions derived from the framework. The evidence supports all four: real construction output grew faster than in any regional peer except Romania; construction producer prices persistently outpaced the GDP deflator; ownership of the leading firms shifted towards politically connected domestic owners, whose profitability rose sharply while their exports remained negligible; labour productivity per hour worked returned by 2024 to roughly its 2000 level; and connected firms captured a rising share of sector profits, as the subsidies' value was partly eroded by the price increases they induced. Competing explanations do not jointly account for this configuration. The findings illustrate how combined demand stimulus and favouritism produce extensive growth without efficiency gains.

We investigate the relationship between the dissipation potential $S_{\mathrm{NB}}$ and information entropy in stochastic Nambu nonequilibrium thermodynamics (NNET). Starting from pure diffusion, gradient-flow systems, and systems with Hamiltonian rotational currents, we derive the stationary distribution of the Fokker–Planck equation including a Nambu flow. We show that, when the Nambu flow preserves the dissipation potential and an Einstein-type relation holds between the diffusion matrix and the transport tensor, the stationary distribution is governed by an exponential weight determined by $S_{\mathrm{NB}}$. Consequently, the local self-information is directly related to the dissipation potential, demonstrating that $S_{\mathrm{NB}}$ should not be identified with the Shannon entropy itself but instead serves as the stationary quasipotential governing the nonequilibrium steady-state distribution. Using the distinction between the thermodynamic entropy and the dissipation potential established for an open gas–piston system coupled to a thermal bath and an external-pressure reservoir, we further clarify the Massieu interpretation of $S_{\mathrm{NB}}$, introduce a Nambu canonical distribution, and derive the covariance of fluctuations within the saddle-point approximation. These results provide a statistical interpretation of the NNET dissipation potential by identifying it as the stationary quasipotential underlying nonequilibrium steady states.

Paul Erdős conjectured that the only powers of two whose ternary expansions contain no digit 2 are 1, 4, and 256. Despite extensive computational verification, the conjecture remains open. We reformulate the problem using generalised base-4 repunits. Any admissible exponent is even, say \(n=2s\), and \[ 2^{2s}=4^s=3R_s(4)+1,\qquad R_s(4)=\frac{4^s-1}{3}. \] Thus \(2^{2s}\) omits the ternary digit 2 if and only if \(R_s(4)\) does. For each depth \(m\), the map \(\Phi_m:s\bmod 3^m\mapsto R_s(4)\bmod 3^m\) is a permutation of the residue classes modulo \(3^m\). Hence every prescribed \(m\)-digit ternary suffix has a unique canonical exponent seed modulo \(3^m\), and all exponents producing that suffix form an arithmetic progression. The lifting identity \[ R_{r+q3^m}(4)\equiv R_r(4)+q3^m\pmod{3^{m+1}} \] gives three lifts at each depth, exactly two of which preserve admissibility. The finite maps assemble into an isometric homeomorphism of the 3-adic integers conjugating \(x \mapsto x+1\) with \(y \mapsto 4y+1\). The unresolved global integer problem remains. No proof of the Erdős conjecture is claimed.

We revisit an established Quantum Control objective, characterize its exact local metric geometry, and reposition it as a rigorous benchmark for Search and Optimization. The trap-free topology of the landscape, established two decades ago, guarantees unhindered optimal pathways, but global topology alone does not govern convergence speed. Although the so-called Quantum Sphere is defined by the fourth power of the control field, we show that its yield gap admits an exact representation as a nonnegative linear combination of $\sin^2$ terms. From this single identity, we derive a two-sided quadratic sandwich with ratio $4/\pi^2$ on a certified, worst-case quarter-wave region, proving the optimal peak nondegenerate modulo its affine gauge and well-conditioned. For white-box optimization, on a halved (eighth-wave) region, the objective satisfies a Polyak-Łojasiewicz inequality with an explicit constant, granting Gradient Descent a linear rate of $1 - 16/(\pi^4\kappa_r)$, where $\kappa_r$ is the condition number of the relevant Hessian block. For black-box optimization, we prove that the elitist single-parent Evolution Strategy converges linearly, almost surely and in expected hitting time, from a certified sublevel set, despite the landscape's continuous gauge symmetry and its lattice of periodic optima; the certified level is explicit whenever the landscape is trap-free, and near the optimum the strategy provably sees only the quadratic tangent of the peak. Systematic numerical simulations corroborate the theory. Finally, we show that additive input (actuator) noise reweights each term of the representation by a Debye-Waller factor without disrupting the landscape's structure, and outline open challenges: output (signal) noise, far-field plateau escape, angular-synchronization certificates, and another physically scalable benchmark (SHG with Glass).

We study the one-dimensional Dirichlet spectrum and obtain new results concerning the origin $\mu$ of the ray contained in the spectrum and its isolated points. First, we prove the improved upper bound $\mu \leq (448875+1611\sqrt{67773})/46982 = 18.4809\ldots$. We also exhibit an explicit gap $(L,R)$ in the Dirichlet spectrum, which yields the improved lower bound $\mu \geq R = 5.9596\ldots$. The proof of the lower bound includes an exact computer-assisted finite search over admissible continued-fraction words. We further prove that all points $Z_k$, $k\geq 2$, introduced by Ivanov are isolated points of the Dirichlet spectrum. Finally, we determine the asymptotic behavior of the neighboring gaps around $Z_k$ and describe the equivalence classes of all irrational numbers whose Dirichlet value is $Z_k$.

We study tilings of closed surfaces by congruent most curvilinear polygons. By relating such tilings to branched coverings determined by planar diagrams and vertex types, we establish necessary and sufficient conditions for a given closed topological surface to admit a metric of constant curvature supporting a tiling with prescribed combinatorial data and number of tiles. We also characterize minimal tilings, which do not descend through nontrivial unbranched coverings to tilings with fewer tiles, in terms of common block decompositions of the associated permutations. Explicit permutation constructions yield infinite families of minimal tilings with prescribed vertex types, even when the planar diagram and vertex types are fixed. In the hyperbolic setting, these results also give finite-index subgroups of the universal tiling symmetry groups that are maximal among torsion-free subgroups, with the index equal to the number of tiles.

For a positive integer $k$, let $\beta_k(G)$ be the maximum cardinality of a vertex set inducing maximum degree less than $k$, and let $SF_k(G)$ be the maximum cardinality of a minimal $k$-star-forming set. The known inequality $\beta_k(G)\le SF_k(G)$ suggests asking when equality holds. We place this question in the framework of upper domination: at $k=1$, $\beta_1(G)=\alpha(G)$ and $SF_1(G)=\Gamma(G)$, so the classical equality $\Gamma=\alpha$ on bipartite graphs is exactly the first member of the proposed hierarchy. We prove the equality for complete bipartite graphs for every $k$, obtaining \[ \beta_k(K_{a,b})=SF_k(K_{a,b})=\max\{a,b,2k-2\}\qquad(a,b\ge k), \] and record the elementary low-degree case $\Delta(G)<k$. For $k=2$ we derive certificate restrictions for minimal $2$-star-forming sets in bipartite graphs. For chain graphs we go further: we prove $\beta_2(G)=SF_2(G)$ and obtain an exact formula for their common value. The proof uses the nested-neighborhood structure together with a classification of witnesses to the indispensability of a high internal-degree vertex. We retain the equality problem for chain graphs as a conjecture only for $k\ge3$, and formulate the broader bipartite conjecture. We also determine the upper domination number of every rectangular grid and combine it with the known exact dissociation number to compare $\Gamma$, $\beta_2$, and $SF_2$. In particular, $\Gamma=\beta_2$ on every even-by-even rectangular grid, while $\beta_2\le SF_2$ always; this motivates a grid equality conjecture whose even-by-even case would yield a three-parameter identity.

There have been persistent efforts to uncover low-dimensional macroscopic dynamics in large systems of interacting microscopic units, in which underlying physical principles govern their collapse onto low-dimensional manifolds. Here, focusing on interacting phase oscillator systems, we show that an unsupervised autoencoder framework can determine their intrinsic dimensionality, in agreement with the predictions of Watanabe-Strogatz theory. Systematic sweeps over the latent dimensions identify the analytically predicted reduced dimensions, while supervised regression confirms that the learned latent variables capture the full information contained in the theoretical Watanabe-Strogatz coordinates. We demonstrate this capability for two prototype systems: Kuramoto oscillators and resistively loaded Josephson junctions. We further establish the diagnostic specificity of the framework by showing that systems lacking such reducible dynamics do not exhibit a sharp dimensional reduction.

Three neural networks are compared as solvers of nonlinear ordinary differential equations: a multilayer perceptron (MLP), a Kolmogorov-Arnold network (KAN) whose edges are B-splines, and a KAN whose edges are Chebyshev series. Each network is trained only on the residual of the equation and on its initial or boundary conditions. The exact solution, or a numerical solution accurate to $10^{-13}$, is used only afterward to measure the error. Five test problems of differential orders one to four are investigated: the logistic equation, a singular Lane-Emden type initial value problem, a singular boundary value problem, the Falkner-Skan equation with suction, a fourth-order problem. The same training budget and the same choice of five activation functions are allocated to all networks. No numerical failures are encountered in any of the \nRuns{} training runs. For every problem, the best KAN is found to be at least as accurate as the best MLP while \paramsKANmin{} to \paramsKANmax{} parameters are used instead of \paramsMLP{}; the margin ranges from a tie to a factor of three. In the Falkner-Skan setting of an earlier study, the B-spline KAN is shown to be 34 to 220 times more accurate at every published test point. Much smaller MLPs are also found to be competitive, whereas at orders three and four, failures are experienced more frequently by KANs, and a coarse basis is required. The derivatives in the equation are distorted by the usual grid extension of a B-spline KAN; consequently, it is replaced with knot insertion, by which the logistic error is lowered more than a hundredfold, though no improvement is provided at order four.

Gas bubbles impose large penalties in water electrolysis at industrially relevant rates. Capillary-fed electrolyzers greatly mitigate them by wicking electrolyte to electrodes and venting gas from the backside, but only below a threshold current density. We show that the separator pore size sets this threshold through two opposing effects: smaller pores raise the supersaturation required for nucleation but reduce the permeability for electrolyte replenishment. A coupled electrochemical-transport model, anchored to a single measured onset, quantifies both mechanisms across 0.1-8 $\mu$m pores. Experimentally, reducing pore size from 8 to 1.2 ${\mu}$m raises the bubble onset limit from 0.18 to 1.73 A cm$^{-2}$ and lowers cell voltage by 0.11 V at 1 A cm$^{-2}$, primarily by suppressing bubble-induced overpotential. At 0.45 ${\mu}$m and below, the replenishment limit takes over and the cells dry out before bubbles form. Our findings establish the bubble-free operation limit as designable rather than purely empirical.

This paper studies flasque resolutions for linear reductive group schemes over arbitrary bases, extending Colliot-Thelene's theory. A flasque resolution exists for any linear reductive group scheme over a connected base scheme (resp. a quasi-compact, quasi-separated base scheme). The construction relies on Borovoi's fundamental group and descent theory involving automorphisms of Chevalley groups. Applications include R-equivalence for reductive group schemes over rings and over a field of cohomological dimension at most 2.

Time-dependent, three-dimensional (3D) oceanic multi-variables define coherent states of the evolving ocean to facilitate ocean diagnosis and advance ocean science to better inform environmental and hazard management. However, extracting quantitative evidence from these variables requires substantial and complex analytical effort. We introduce OceanMind, a multi-agent AI system that directly couples large language models (LLMs) with comprehensive time-dependent 3D ocean states for swift and effective diagnosis. OceanMind organizes the analytical process into four coordinated complexity stages: Query Routing, Skill-based Planning, Tool Execution, and Evidence-based Summary Generation. Specialized agents interpret user requests, construct and execute multi-step computational workflows, and synthesize quantitative evidence. To ensure reliable workflow construction, 63 reusable ocean-specific analysis skills serve as procedural manuals that guide the LLM agent in selecting data, conducting diagnostics, and applying analytical tools. With reflection and replanning mechanisms that use execution feedback to repair invalid plans, OceanMind ensures reliable analysis workflows across diverse needs. On a benchmark of 240 computational-workflow queries spanning the four stages, OceanMind outperformed general ReAct agents with the same registered tool pool, achieving relative improvements of 41.2% in effectiveness and 21.5% in efficiency. Beyond the benchmark, OceanMind reproduced published oceanic diagnostics, validated hypotheses, and supported environmental decision-making over global oceans. Overall, OceanMind advances LLMs by integrating them with time-dependent 3D ocean analysis, enabling scientific interpretation and enhancing formulation of environmental policies based on quantitative ocean evidence.

We study spontaneous scalarization of static, spherically symmetric black holes in scalar–Gauss–Bonnet gravity with a Ricci coupling $F=1 \beta\phi^{2}/4+\gamma\beta^{2}\phi^{4}+O(\phi^{6})$ that nonlinearly completes the quadratic one. For arbitrary couplings and potential, we derive the near-horizon regularity condition and horizon curvature invariants in closed form; scalarized branches terminate where this condition degenerates, at finite horizon curvature. A weakly nonlinear expansion fixes the bifurcation direction exactly: for a massless scalar, its governing coefficient is quadratic in $\beta$, changes sign at $\beta=1.1156$, and depends linearly on a quartic self-interaction. At equal horizon values of the scalar, completions differ in mass, to leading order, in proportion to $\gamma$. A radial-perturbation master equation for arbitrary couplings confirms the exchange of stability at the bifurcation and turning points. Consequently, the domain of radially stable black holes shrinks as $\gamma$ increases over the range studied; it is smaller for the exponential completion than for the quadratic truncation.

In general, measurement error constrains the performance of information engines. However, this may not hold universally. In this work, we propose a class of two-reservoirs information engines (TRIEs) with a two-level system and demonstrate that the monotonic relationships between the work output, the efficiency, as well as the work fluctuations of the TRIEs and measurement error are determined by the demon's control parameters. Utilizing this, we establish the conditions that allow the TRIEs to generate useful work under any level of measurement error, and identify the regimes in which these performance parameters can be optimized simultaneously by the measurement error. Furthermore, we demonstrate that when the demon's energy cost is supplied solely by the cold reservoir, the efficiency that accounts for this cost can surpass the standard Carnot bound without violating the second law. Finally, we design a dual Otto cycles TRIE and provide an experimental scheme for its implementation based on a trapped $^{40}\mathrm{Ca}^+$ ion. Significantly, our results can be applied to the multi-level system information engines.

Flow matching models excel at generative modeling, and many downstream applications require their samples to satisfy prescribed constraints, such as observed measurements and physical laws. However, existing constrained samplers often face a trade-off: enforcing constraints can substantially displace samples from the pretrained data distribution. To address this trade-off, we introduce MintFlow, a training-free constrained sampling framework that formulates constraint enforcement as a minimal intervention on the pretrained flow trajectory. MintFlow seeks the minimal perturbation of an intermediate flow state such that its subsequent evolution under the pretrained flow field satisfies the target constraint. By minimally perturbing the flow state while keeping the pretrained flow field unchanged, MintFlow enforces the constraint while minimizing unnecessary deviation from the pretrained distribution. An adjoint formulation yields a closed-form expression for this perturbation, eliminating expensive iterative optimization. Furthermore, MintFlow adaptively selects the intervention time to balance the required perturbation magnitude with its amplification by the remaining flow. Across a range of tasks in generative vision and physical system modeling, MintFlow achieves competitive constraint satisfaction while preserving the pretrained generative distribution substantially better than state-of-the-art constrained methods.

Online feature stores decouple feature computation from model serving, materializing features from upstream event streams into a low-latency store that inference reads at request time. This decoupling introduces a freshness gap - the interval between an event occurring at the source and its effect becoming visible in the served feature vector - that is widely documented as a cause of training-serving skew but has no formal treatment as a bounded, cost-quantified quantity. We model the online feature store as a materialized view over an event stream and define a per-feature freshness budget: the difference between the decision window a feature is consumed within and the staleness the serving pipeline imposes on it. We prove two propositions - a per-feature staleness bound, and a closed-form threshold below which a feature can never be served within budget - and validate them in a seeded, reproducible discrete-event simulation and on a real Apache Kafka, Redis, and PostgreSQL pipeline. The predicted collapse threshold reproduces on real infrastructure, with the decision-error transition landing where the closed form predicts. Reconciling simulated and real drop rates surfaces a methodological result of independent interest: the dominant simulation-to-infrastructure gap is not latency but a sampling-phase artifact - a simulation whose request clock and recomputation cadence are phase-locked systematically under-observes staleness. Randomizing that phase reconciles simulation and infrastructure to within 0.018 mean absolute error in per-feature drop rate. Results are scoped to the studied workloads and configurations and are not claims about any specific production feature-store implementation.

We study online convex optimization with one unbiased stochastic subgradient per round and an unknown finite conditional $p$th noise moment, $1<p\le2$. For every fixed interval $I$ of length $n$ and comparator path with $\Lambda_I=1+P_I/D$, one learner achieves \[ E[Regret_I(u)]\le\min(GDn, C[GD\sqrt{n(\Lambda_I+\log^2(2T))} +\sigma Dn^{1/p}(\Lambda_I+\log^2(2T))^{(p-1)/p}]). \] The learner uses none of $G,\sigma,p,I,P_I$, and the constant is universal. Interval adaptation adds to comparator complexity, preserving the distinct mean-gradient and noise exponents. The analysis controls calibration in expectation and limits the cost of observation-scale changes. Its general theorem compares to distributions over predictably available experts with relative-entropy dependence on a nonuniform prior. A common prior favors long windows and long restart lengths. With the statistics supplied, the interval cost becomes $1+\log(T/n)$, including the optimal full-horizon static rate. A change-of-measure lower bound identifies the noise power of this logarithm for learners retaining a full-horizon optimal guarantee, under explicit conditions. Static comparisons and deterministic partitions follow from the same decisions.

A dagger on a category assigns to every morphism $f \colon X \to Y$ a morphism $f^\dagger \colon Y \to X$, contravariantly and involutively, in the way the transpose reverses a real matrix. We show that on the categories of linear algebra and differential geometry, daggers encode metric and Riemannian structures. As is known, on a single vector space, an anti-involution of the endomorphism ring is the adjoint of a form that is determined up to a scalar. But more generally, on the category of all finite-dimensional vector spaces, a dagger is the adjoint for an involution of the scalars together with a unique unimodular Hermitian form on every space. The same holds for finitely generated projective modules over a commutative ring, and hence, through the Serre-Swan correspondence, for vector bundles, where a dagger is a choice of bundle metrics. The dagger structure expresses notions of orthogonality categorically: orthogonal complements and normal bundles acquire universal properties, and Hermitian modules form a dagger compact category. Moreover, daggers on Lie groups, Lie groupoids and their infinitesimal counterparts capture unitary representations and metric connections: a connection is metric precisely when its parallel transport is a dagger functor, or, infinitesimally, when its splitting of the Atiyah sequence commutes with the daggers. Finally, we classify the daggers on the locally graded category of differential operators of a Lie-Rinehart algebra. They correspond to metric-divergence structures, consisting of a real structure, a divergence, and a unimodular Hermitian form on every finitely generated projective module, and they act on operators as formal adjoints; Green's formula expresses the difference between an operator and its adjoint as a divergence.

Electricity price forecasting (EPF) supports scheduling, bidding, and risk management in electricity markets, yet existing benchmarks focus mainly on numerical inputs, leaving the forecasting value of forecast-time textual context insufficiently evaluated. We introduce TRACE, a reproducible benchmark of 7,300 zone–day instances pairing prices from five zones in a major U.S. market with official operational text available at the forecast cutoff. TRACE reconstructs official operational text at each cutoff, preventing post-cutoff information leakage. We evaluate TRACE for semantic alignment and forecasting value. Semantic assessments align with central movement and both tail risks in ground-truth prices, most consistently for upper-tail price risk. Forecasting value is reflected in a median 7.4% reduction in upper-tail pinball loss across time-series foundation models. A controlled cross-day text-mismatch ablation reverses the gains, falling below the no-text baseline.

Time series forecasting remains a critical challenge across numerous domains. Despite significant advancements, existing approaches struggle with complex phenomena such as regime shifts, cross-domain knowledge transfer, and multimodal data integration. This paper introduces Multi-Agent Collaborative Time Series Forecasting with Emergent Memory (MACTS-EM), a novel framework where specialised agents collaborate to achieve superior forecasting performance. The MACTS-EM architecture integrates: (1) domain-specialised forecasting agents for pattern recognition, anomaly detection, causal inference, and uncertainty quantification; (2) a meta-cognitive layer for dynamic agent allocation; (3) an emergent memory mechanism enabling cross-domain pattern transfer; (4) multimodal contextual integration; and (5) adversarial robustness components. Evaluation across financial markets, climate patterns, energy consumption, and pandemic propagation demonstrates that MACTS-EM outperforms existing approaches in most scenarios, with 8-12% improvement in forecasting accuracy, 22-27% better zero-shot transfer capability, 16-21% enhanced resilience during regime shifts, and 15-18% faster recovery after distribution shifts. Our findings suggest that collaborative, agentic approaches to time series forecasting represent a promising direction beyond traditional architectures, particularly for complex real-world scenarios requiring multi-resolution temporal understanding and contextual adaptation.

Interpretive Structural Modeling (ISM) is a well-known process for multi-criteria decision making. The success of ISM over other methodologies is its ability to model causal relationships, the binary scale of factors, and resulting hierarchical representation. Traditionally, the modeling process is performed by repeated interactions with subject matter experts until consensus is reached. This process is tedious, labor-intense, and most importantly limits the ability of ISM to scale to studies with hundreds of variables. Drawing on existing work of causal graph discovery with large language models (LLM) as imperfect experts, this work explores an integrated LLM-ISM approach for ISM. Pairwise, k-wise, rowwise, and full graph discovery methodologies are compared and evaluated. It is shown that causal graph discovery methods for ISM perform best using rowwise (SHD=160, F1-score=0.77) and full graph methods (SHD=135, F1-score=0.73).

The universal approximation property of dropout neural networks does not by itself describe the network size required for an accurate random realization. In this work, we study approximation of the unit ball of $W^{n,\infty}([0,1]^d)$ by ReLU networks whose edges are retained independently with probability $p$. The approximation error is measured uniformly over the input domain, and the guarantee holds with probability at least $1-\delta$ for a single sampled network. We construct networks of constant depth and size $\widetilde O_{n,d}(p^{-9}\varepsilon^{-\max\{d/n,2\}} \log(1/\delta))$. The construction combines bounded local subnetworks, localization on a successful approximation event, and a multiscale Taylor decomposition. Conversely, Sobolev capacity imposes a lower bound on the number of surviving edges, while approximation of a fixed affine function requires an output-layer cost of order $((1-p)/p)\varepsilon^{-2}\log(1/\delta)$ at sufficiently high confidence. For fixed $p\in(0,1)$ and $\delta<\min\{1/2,1-p\}$, the upper and lower bounds match in the accuracy exponent under a fixed or logarithmic depth budget. When $d\leq2n$, they also match in confidence up to logarithms of accuracy. We extend the lower bounds to $W^{n,r}$ targets with $L^s$ error, and distinguish this extension from the upper bound for $W^{n,\infty}$. The optimal retention dependence and logarithmic factors remain open.

Many scientific questions require reasoning about what was never observed: What if the conditions, interventions, or history had been different? Models can predict accurately on observed data yet fail on such what-if queries when correlated inputs are varied independently. A common remedy is to add controlled simulation data in which these factors are explicitly disentangled, but this requires access to a simulator, can be computationally expensive, and inherits the simulator's modeling assumptions. We introduce ReRoute, a framework for targeted scientific what-if prediction that combines factual data with partial mechanistic knowledge, without requiring controlled intervention data for adaptation. ReRoute fixes the queried input of a pretrained backbone to a reference value, reintroduces its variation through a known mechanistic pathway, and fine-tunes on the original factual data, while leaving downstream effects to the learned dynamics. We provide a causal identification result for this construction under explicit structural assumptions, with the core argument machine-checked in Lean. After showing that ReRoute achieves highly accurate counterfactual predictions in a controlled advection-diffusion system where exact responses are available, we turn to state-of-the-art climate emulation. On held-out coupled-climate interventions, ReRoute reduces aggregate climate error by 18.2-31.8% under severe CO$_2$ distribution shifts while preserving skill under standard conditions, at a small fraction of the cost of retraining on additional controlled simulations, without even accounting for the substantial expense of generating such data. Finally, on an emulator trained from historical ERA5 reanalysis, where no counterfactual reference exists, ReRoute preserves substantially more of the surface warming implied by the observed boundary conditions under a fixed-CO$_2$ counterfactual.

Speculative sampling accelerates diffusion generation by verifying inexpensive draft states in parallel while preserving the target law. Recent tree-based methods allocate the parallel compute budget more effectively than single-chain drafts, as demonstrated by Diffusion Greedy Rejection Sampling (D-GRS). D-GRS generates $K$ conditionally independent candidates per node, and sequentially tests them in their generation order. Yet the sampled candidates admit an informative ranking without additional target-model evaluations. To exploit this, we introduce Rank-Aware Speculative Sampling (RASS), a verification rule for speculative draft trees based on rank-aware list coupling. RASS orders draft candidates along the proposal-target mean displacement and samples a rank with weights optimized to minimize total variation between the selected-proposal and target laws. Finally, the selected candidate is maximally coupled with the target, with residual correction ensuring exact sampling for any choice of rank weights. We evaluate RASS on a Gaussian-mixture target, unconditional pixel-space generation on FFHQ, conditional generation on CIFAR-10, and latent diffusion with Stable Diffusion 3.5 using COCO2014 prompts. Measured by the ratio of standard to speculative sampling's target-model evaluation counts, RASS improves on D-GRS across the evaluated settings, with gains reaching approximately 20% on CIFAR-10 at matched compute budgets.

In two previous articles, we have explained how super-massive Kerr Black Holes in the nucleus of an active galaxy or quasar can generate a large amount of axion particles in their magnetosphere. In that third part, we explore the possibility of observable dynamics resulting from some interaction of an axion wind with the baryonic matter and the axionic Dark Matter in the surroundings of the closest AGNs. The axions that we have considered are the same as those Ultra Light Axions favored in cosmology as the main content of Dark Matter and which form Bose-Einstein Condensates, supporting a gigantic Schrödinger wave. In the paper we investigate what can be the nature of the interaction of the axion wind with the plasma and the hydrogen gas in the host galaxy and its halo. Despite the intriguing nature of a possible interaction, we show that only a kind of gravitational Tcherenkov-Vavilov effect can be efficiently operating with acoustic waves through a streaming instability. The instability is fast enough to operate efficiently in the central region of the host galaxy up to parsec scales and makes Lyman-alpha diagnostics of the acoustic waves of hydrogen gas and plasma possible, as well as the detection its hyperfine emission at 21 cm. However it turns out that scattering of the axion wind is very long and thus these winds could travel over very large distances and cross other galaxies. And then the wind travels without significant loss along its halo and even beyond up to quantum thermalization, which makes possible direct detection of these axions by conversion into gamma-rays on Earth, opening a possible axion astronomy in the local Universe. The parameters of Messier 87 indicate that it is a good candidate for developing these investigations both through the conversion of axions into gamma-rays and Lyman alpha diagnosis in the bulge.

It has recently been shown that nearest-neighbour retrieval provides a strong baseline for molecular fingerprint prediction from MS/MS spectra, with several variants matching or outperforming current deep learning models (Khoo and Barzilay, 2026; Liu et al., 2026; Gupta et al., 2026). Importantly, "nearest neighbour" encompasses a family of retrieval methods that differ in the information assumed to be available at inference. In this report, we systematically compare several nearest-neighbour variants and show how these differing assumptions affect performance. Our goal is to establish stricter baselines that enable more rigorous benchmarking and better measure progress in this area.

Operational land surface forecasting systems built on Mamba-family Structured State Space Models absorb non-stationary confounding events (unrecorded irrigation booms, dam-operation shifts, sensor recalibrations) into their state-transition matrices, silently biasing NDVI, LST, and crop phenology predictions long after the physical cause ends. This paper introduces SSU-LSF (State-Space Unlearning for Land Surface Forecasting), the first machine-unlearning framework purpose-built for geoscientific Mamba-based SSMs. We develop EKFac influence functions specialized to the Mamba state matrices via a closed-form matrix-exponential gradient, use spectral-radius-weighted elbow thresholding to localize a temporal confounding footprint $\Phi$, and apply Hessian-free projected gradient ascent within a KL-divergence trust region augmented by spatial total-variation (TV) regularization. Proposition 1 establishes that residual confounding is bounded by $\mathcal{O}\big((1-\rho(\bar{A})^{T_c})/((1-\rho(\bar{A}))\mu)\big)$, which grows with the window length $T_c$. Across three heterogeneous benchmarks and eleven baselines, SSU-LSF achieves confounding reduction rates of $0.773$ (CropHarvest), $0.821$ (NDVI-LST), and $0.859$ (ERA5), with worst-case clean-domain RMSE degradation of $4.2\%$ on ERA5, converging in 3–5 epochs at $8.4\times$ lower GPU-cost per unlearning request than full retraining. Code: https://github.com/Anidipta/SSU-LSF