Papers

Hamiltonian systems exhibit rich dynamical behaviors, among which escape is a typical phenomenon. Below the escape threshold, trajectories initiated inside the central potential well remain bounded. Once the energy exceeds this threshold, and the initial conditions lie within the escape basin, the system eventually evolves along unbounded orbits toward infinity. Predicting the escape dynamics is important for understanding the evolution of these systems. In this study, we employ the next-generation reservoir computing framework, a powerful machine learning architecture widely used in dynamical prediction, to address this challenging issue. To capture the intrinsic features of the system, we train the model using data from multiple energy values. Numerical experiments on two Hamiltonian systems show that, with only three time series as training data, each corresponding to an energy value below the escape threshold, the model is capable of not only predicting the escape threshold itself, but also reliably predicting the structure of the escape channels. As a primary contribution of this work, these findings demonstrate that escape dynamics can be inferred from subcritical data alone in the two systems studied, offering a promising paradigm for data-driven studies of Hamiltonian systems.

Modern agentic coding frameworks increasingly rely on community-shared rule files (e.g., AGENTS.md or .cursorrules) to guide autonomous code generation, yet the security risks of this pipeline remain underexplored. To bridge this gap, we introduce the package hallucination attack, where an attacker injects malicious prompts into benign rule files to induce coding agents to replace legitimate dependencies with attacker-controlled packages. To obtain effective malicious prompts injected into rule files, we propose PackHallu, an evolutionary optimization framework that iteratively rewrites these injected prompts using trajectory-level feedback and LLM-guided mutations. Evaluations across multiple benchmarks, LLMs, and agent frameworks show that PackHallu achieves high attack success rates and strong transferability across diverse models and agent combinations. Our findings demonstrate that coding agents are vulnerable to package hallucination attacks, highlighting the urgent need for stronger security safeguards in autonomous coding systems.

We study consumption and investment over a finite horizon when continuation utility must stay above a welfare floor that changes over time, and the floor at the horizon acts as a terminal wealth floor. A cumulative multiplier reduces the problem to a parabolic obstacle problem. When relative risk aversion exceeds one, the intervention boundary exists at every interior date. When it is below one, the boundary exists exactly at the dates where a discounted transformation of the floor is locally increasing and sets a strict running record, so it can disappear and later reenter. We prove local regularity of the boundary, construct the dual value from the obstacle solution, and verify the optimal cumulative multiplier, whose moments are controlled through its representation as a running supremum. Strong duality identifies the exact lower endpoint of feasible wealth. With stochastic income in a complete market, the floor becomes the value of an outside option, which links the boundary to limited commitment and endogenous borrowing capacity.

A reader's first highlights on a page are the cheapest personal signal a reading product has. The natural plan is to suggest what similar earlier readers marked, and to judge the result against popularity. We argue that the baseline to beat is reading position. In a time-ordered evaluation on one social highlighting platform (7,343 reader-page pairs on 1,511 pages after one highlight), ranking the sentences just below a reader's first highlight, with no other reader's data, puts the next highlight in the top five 47% of the time, against 26% for popularity and 29% for the better of two similarity methods. The baseline depends on the target: over all later highlights that ranking loses to popularity, while popularity discounted by distance from the latest highlight, at the scale with the best average precision of three tried, beats popularity and both similarity methods on both targets. In a comparison specified in advance, neither similarity method shows a gain over popularity in average precision over all later highlights, from one to five highlights, and a gain of +0.01 is excluded. Nor would a gain by itself show that a method has found a reader's preferences: synthetic readers who share one set of preferences produce one, and an evaluation out of time order shows a method where the reader went. The position results are exploratory and unconfirmed. Personalisation inside a document should be evaluated in time order and against reading position.

Coupled systems of ordinary differential equations (ODEs) are widely used to model complex physical and biological processes. In these applications, ODE model parameters must be calibrated to field data to enable prediction and parameter inference. In practice, the governing ODEs are an imperfect approximation to the true system, leading to non-negligible model discrepancy that must be incorporated to avoid biased parameter estimates. However, it is well known that highly flexible discrepancy models can confound discrepancy and simulator parameters, leading to poor identifiability. As a result, practitioners often must impose problem-specific prior constraints to adequately regularize the discrepancy, which can be challenging and cumbersome. In this work, we propose a novel calibration framework for coupled, multi-output ODE systems that enforces automatic, model-structural constraints on the discrepancy, improving identifiability without requiring application-specific discrepancy priors. Forward-model gradients are computed via sensitivity equations, solved jointly with the ODE system as a coupled initial value problem. Posterior inference is performed using an adaptive Metropolis-within-Gibbs sampler tailored to the resulting constrained posterior. We demonstrate the approach on three model systems: a mass-spring oscillator, a model of infectious disease spread, and a neuron firing model. We show that our method leads to improved parameter inference and predictive accuracy compared to standard calibration approaches.

Excited electronic states are difficult to accurately and tractably model, particularly those exhibiting multiexcitonic character. Scalable single-reference excited state methods typically fail to describe such states as they suffer from the defects of the mean-field approximation and/or neglect higher-rank excitation operators. To address this, we derive the [T] perturbative correction to the quantum self-consistent equation-of-motion unitary coupled cluster singles and doubles (q-sc-EOM-UCCSD) method. The method captures triple excitation effects through a classical step of second-order in many-body perturbation theory following a q-sc-EOM-UCCSD calculation on a quantum computer. The perturbative correction developed has a $\mathcal{O}(N^7)$ classical computational time scaling. We incorporate [T] into a hybrid compute strategy wherein eigenvectors of the effective Hamiltonian from the hybrid quantum-classical q-sc-EOM-UCCSD algorithm are relayed to a classical computer which performs the postprocessing required for the [T] correction. The benefits of [T] are quantified in a benchmark covering a variety of electronically excited states of the isoelectronic CH$^+$ and BH molecules. We find that the addition of the [T] correction to both trotterized and full operator variants of q-sc-EOM-CCSD can offer dramatic improvements over baseline q-sc-EOM-UCCSD. Our assessment clearly demonstrates the importance of considering higher-rank excitation operators in q-sc-EOM-UCCSD, quantifies the overall success of the [T] correction, and discusses some of its limitations.

Surveys often estimate the share of cases in each of several categories, such as the causes of an event or the reasons for a visit. Readers then compare the categories, but reliability rules usually check each estimate alone, not the differences between them. We ask instead which groups of categories can be told apart under the sampling design. The categories often fall into blocks, such as driver-related and vehicle-related causes. A grouping is called distinguishable when every pair of its groups in the same block differs by more than the design's minimum detectable difference, computed from the primary sampling units at the design's degrees of freedom. The maximal distinguishable groupings are the most detailed and can differ in size, while merging two groups can destroy distinguishability. So even if splitting any one group in two breaks distinguishability, a more detailed distinguishable grouping may still exist. We propose two methods. One merges categories until the grouping is distinguishable, then tests every single split. The other lists every distinguishable grouping of a block with few categories. When groups in different blocks must also differ, there might be no grouping that qualifies. An application to the National Motor Vehicle Crash Causation Survey, with 24 primary sampling units in 12 strata, gives the grouping maxima for two blocks of its categories. Of more than 27 million groupings of its 13 vehicle-related critical reasons, none with more than four groups is distinguishable, and its environment-related reasons resolve into at most four groups. Merging two groups of a distinguishable vehicle grouping destroys distinguishability in 5,239 of 31,773 cases. If groups from different blocks must also be told apart, no grouping qualifies. Analysts can apply these methods before publishing categorical comparisons from designs with few degrees of freedom.

A Markov chain with exit rates $q_i$ runs for $N$ steps with transition matrix $I+(T/N)Q$. On each irreducible face that carries start mass, we prove a local limit theorem for the occupation counts with an explicit constant and relative error $O(1/T)$, uniformly for $1\le T\le\varepsilon N$ with $\varepsilon\max_iq_i<1$ and on compact subsets of the open face. Up to $O(1)$, minus the log-probability of counts $n$ with support $F$ is $|F|-1$ times $\log(N/\sqrt T)$ plus $N$ times the Donsker and Varadhan functional of $I+(T/N)Q$ at $n/N$. Every admissible face has an explicit power of $T$ in its largest probability for large $T\le N^{1/3}$, and a constant given by a finite-dimensional variational formula. For $T=\tau\log N$ the face of the mode minimizes $\tau\lambda_F+|F|-1$ at first order, with $\lambda_F$ the exit rate. If an admissible face has $\Delta\ge1$ more states than another and an exit rate smaller by $\delta>0$, we locate every crossing of their largest probabilities for large $T\le N^{1/3}$ up to $O(1/\log N)$. For two irreducible faces the crossing is unique for large $N$ in $1\le T\le\varepsilon N$, with coefficient $1/2$ of $\log\log N$. Near a facet, the continuous-time occupation density is the facet density times an explicit kernel, with relative error $O(1/T)$ at times $x\le cT$ off the facet, $c$ small. The counts obey a lattice form with extra error $T^2/N$ for $T\le\varepsilon_1\sqrt N$. The kernel is Bessel for $x=o(T)$ at facet profiles where the removed state has equal entry and exit laws, and at a profile where the rate has a first correction, of order $m\ge3$, if and only if $x=o(T^{1-2/m})$. On the complete graph with the uniform start, each face of at least 2 states crosses a vertex exactly once.

Supercooled first-order phase transitions provide an attractive framework for generating observable gravitational waves (GWs), but the accompanying entropy injection dilutes any pre-existing baryon asymmetry, motivating baryogenesis during or after the phase transition. Cold baryogenesis offers a promising mechanism in which bubble collisions generate Higgs winding configurations and the Chern-Simons number through out-of-equilibrium dynamics. In the present study, we perform three-dimensional lattice simulations of a two-field model where the scalar field responsible for the phase transition is separated from the Higgs sector. Using realistic initial conditions obtained from bounce solutions, we follow the bubble expansion and collisions and study the generation of Chern-Simons number. It is found that efficient Chern-Simons production persists even for strongly non-degenerate potentials with large latent heat, which are favorable for enhancing GW signals. We also estimate the resulting baryon asymmetry using the simulated Chern-Simons dynamics and an effective CP-violating operator. Our results demonstrate that cold baryogenesis remains viable in the strongly supercooled regime, providing an important step toward establishing baryogenesis from cosmological first-order phase transitions.

We study explicit lightning-plus-polynomial (LP) approximation of $g(z)z^\alpha(\log z)^m$ on a closed unit-radius sector, where $0<\alpha<1$, $m$ is fixed, and $g$ is analytic nearby. Since the basic sector root-exponential rate is already known, we address three finer questions. First, a one-sided Abel–Poisson formula and a pole-free notched contour give a branchwise error bound that is uniform for grid shifts in compact subsets of $(0,1)$ and valid for every fixed logarithmic power. Second, for $z^\alpha$ in the dense and balanced regimes, an exact vertex-scale rescaling identifies a positive finite leading constant for a specified shifted sequence. Third, a rate-matched band of additional representation poles followed by polynomial compression gives the explicit representation dimension $N+2\sqrt{2\alpha\lambda_\sigma N}+O(1)$ while leaving the first $N$ tapered poles unchanged. A slightly overresolved hybrid sequence also preserves the exact leading constant. Matched-dimension experiments, including reentrant sectors, illustrate the allocation. These exact-asymptotic and dimension statements concern the stated constructions, not sector minimax approximation or total computational complexity.

Neural PDE surrogates are trained on numerical solver outputs that contain both physical evolution and solver-specific discretization errors. Because surrogates are also evaluated against held-out trajectories from the same solver, standard benchmarks cannot distinguish fidelity to the exact evolution from imitation of the numerical scheme. We introduce an empirical Fourier-symbol diagnostic that probes a trained surrogate's linearized one-step operator with individual Fourier modes and compares it with both exact-evolution and training-scheme references. To address architectural spectral bias, we train identical networks on schemes with orthogonal dissipative and dispersive signatures and compare their learned operators. In linear advection, the learned surrogates reproduce the training schemes' amplitude and phase errors, with the twin-scheme difference reaching more than 99.8% of the analytically predicted full-imitation ceiling. The same behavior occurs for a non-local Fourier neural operator and at the operator level for nonlinear Burgers dynamics. These results show that agreement with solver-generated test data does not by itself establish fidelity to the exact evolution. Fourier-symbol measurements provide a direct diagnostic of numerical provenance.

Simulation enables large-scale, low-cost robot data generation, but policies trained in simulation often fail to transfer to the real world due to the sim-to-real visual discrepancies. Existing approaches often rely on intermediate representations, which can discard rich semantic information or require additional perception modules at deployment. We address this visual sim-to-real gap with RoboRender, a framework that converts simulated trajectories into photorealistic RGB videos for policy learning. RoboRender trains a robot-oriented video generation model conditioned on simulated depth videos, language instructions, and robot RGB mask videos, preserving simulator geometry, robot motion, and action labels while synthesizing realistic textures, backgrounds, and distractors. The generated RGB videos are paired with simulator-provided states and actions to train policies for zero-shot real-world deployment. On robot video test sets, our video model outperforms depth-conditioned video generation baselines in generation quality. In real-world experiments across pick-and-place, articulated-object manipulation, and mobile manipulation tasks, policies trained on RoboRender-generated data achieve a 71% average success rate, outperforming raw simulation renderings and conventional visual domain randomization by approximately 7.1x and 3.6x, respectively. We further show that policy performance improves with more generated videos per simulation trajectory, increasing opening-task success by 65 percentage points. These results demonstrate that generative video rendering mitigates the visual sim-to-real gap for zero-shot policy transfer. Project website: https://robo-render.github.io/.

Reconstructing 3D Computed Tomography (CT) images from a few X-ray projections is a highly ill-posed inverse problem due to the loss of volumetric information. We propose PhyDiCT, a training-free framework that integrates a differentiable Physics-based forward model, grounded in the Beer-Lambert law, with a text-conditioned Diffusion as a strong prior to reconstruct 3D lung CT images. We refer to our approach as training-free since the prior model is used without fine-tuning, and our goal is to steer the denoising procedure to generate samples consistent with X-ray observations. We guide the diffusion generation using Split Gibbs sampling to jointly optimize for projection fidelity (reward) and consistency with prior knowledge. Also, we introduce a test-time refinement step that enhances image realism and anatomical coherence. We extensively evaluate our method on publicly available 3D CT datasets using both perceptual and semantic metrics, demonstrating that it surpasses existing plug-and-play diffusion and fully trained reconstruction approaches. Our findings highlight that combining a strong generative prior with the underlying physics of image formation substantially improves reconstruction quality, e.g., 7.5% improvement on SSIM compared to full training methods. Code will be released at https://github.com/batmanlab/PhyDiCT.

Large Vision-Language Models achieve strong VQA performance, but processing high-resolution, information-rich images requires substantial computation, motivating visual token reduction. However, existing methods often prune individual tokens or rely on fixed-size cropping, limiting their ability to preserve spatially structured information such as horizontally or vertically elongated text. To address this limitation, we propose ReFIT, an instruction-guided visual token reduction framework for efficient LVLM inference. ReFIT consists of Relevance-Guided Window Reshaping (RWR) and Instruction-Guided Token Refinement (ITR), where RWR captures instruction-relevant regions by adapting to their spatial characteristics, while ITR further removes unnecessary visual tokens. Experiments on four VQA benchmarks demonstrate that ReFIT improves answer accuracy while reducing computational cost, and qualitative results demonstrate its effectiveness in localizing relevant regions and removing unnecessary visual information.

For $h>0$, degree $m\ge0$, $0\le\beta<\pi$, and $q>0$, let $\Sigma_\beta=\{re^{i\theta}:r\ge0,\ |\theta|\le\beta/2\}$ and let $E_{m,\beta}(q)$ denote the best uniform error for approximating $e^{-hz}$ on $\Sigma_\beta$ by rational functions whose only possible finite pole is $s=-q/h$, with multiplicity at most $m$. For the linear scaling $q=\alpha m$ with a fixed $\alpha>0$, a Möbius map reduces the problem to polynomial approximation on a compact lens. Write $\gamma_\beta=2-\beta/\pi$ and let $\alpha_c(\gamma_\beta)$ denote the saddle-coalescence threshold. For $0<\alpha<\alpha_c(\gamma_\beta)$, let $H_\beta(\alpha)\in(0,1)$ be the exponential of the real saddle action. An exact boundary Faber integral, a global fractional-power contour deformation, and a two-saddle expansion yield the sharp two-sided order $E_{m,\beta}(\alpha m)\asymp m^{-1/2}H_\beta(\alpha)^m$, uniformly on compact subsets of $0<\alpha<\alpha_c(\gamma_\beta)$. On this noncoalescing interval, the exponential factor $H_\beta(\alpha)$ has the unique minimizer $\alpha_\beta^*=1/[\gamma_\beta\sin(\pi/(2\gamma_\beta))]$ and minimum $\tan(\pi/(4\gamma_\beta))$, recovering the classical half-line factor $\tan(\pi/8)=\sqrt{2}-1$ at $\beta=0$. Numerical-range estimates give matrix dimension-independent upper bounds for projected shift-and-invert Arnoldi, while finite-degree pole searches and nonnormal experiments illustrate the distinction between the sector-uniform law and individual matrix computations.

Best-of-N (BoN) is a common inference-time alignment method that selects the highest-scoring response among N samples from a reference model. Evaluating BoN policies from logged data is challenging under sample-only access because standard off-policy estimators require density ratios that depend on unavailable response likelihoods. In this paper, we propose a sample-only framework for evaluating and selecting BoN policies without access to these likelihoods. We show that the order-statistic structure of BoN allows the required density ratios to be expressed through score-rank probabilities that are estimable from samples alone. We then develop a doubly robust estimator of the BoN policy value (BoN-DR) that efficiently reuses a shared auxiliary sample pool across candidate budgets. We establish valid asymptotic inference even under reward estimator misspecification and prove the efficiency of our BoN-DR estimator. Since larger budgets can amplify errors in the score function and lead to reward overoptimization, we derive two selection rules: (i) maximizing the estimated policy value and (ii) maximizing a lower confidence bound on the improvement over the reference policy, which accounts for estimation uncertainty and provides a no-harm guarantee. Across synthetic experiments and GSM8K with multiple reference and reward models, our framework accurately estimates BoN policy values and selects effective sampling budgets.

Entropy-stable boundary conditions are critical for ensuring that the corresponding numerical scheme satisfies the discrete entropy inequality, thus mimicking the second law of thermodynamics for the compressible Navier-Stokes equations discretely. Provably entropy-stable adiabatic wall boundary conditions were introduced in Parsani 2014~\cite{parsani2014entropy}. These discrete wall boundary conditions were further generalized in Dalcin 2019~\cite{dalcin2019conservative} to a moving adiabatic solid wall or a wall with a prescribed heat flux for the compressible Navier–Stokes equations discretized by using summation-by-parts (SBP) and simultaneous-approximation-term (SAT) operators. Currently, no entropy-stable formulation using SBP-SAT operators for isothermal wall boundary conditions is available in the literature. This paper presents new isothermal no-slip wall boundary conditions that: 1) enforce the $T|_{y=0}=T^{wall}$ and $\vec{v}|_{y=0}=\vec{v}^{wall}$ conditions with the design order of accuracy while maintaining stability, 2) provide the correct sign of entropy production at the wall, and 3) mimic the entropy balance of the Navier-Stokes equations at the discrete level. The structure-preserving and design-order properties of the proposed methodology are demonstrated and verified on standard benchmark problems for compressible flows.

The spin density and Josephson current at equilibrium are studied numerically in a superconductor-normal metal-superconductor (SNS) junction with extrinsic spin-orbit interaction in the N segment. This system preserves structural inversion symmetry. We find that the Josephson current generates the spin Hall effect, resulting in an opposite spin density near the two edges of the N segment. We also find that a Zeeman magnetic field with gradient applied to the N segment generates an anomalous phase shift due to the inverse spin Hall effect, resulting in a diode effect when higher harmonics of the Josephson current are present. These results are consistent with those of a previous theoretical study based on diagrammatic perturbation theory. We find that the diode effect becomes more pronounced when stronger coupling between the N and S segments enhances the higher harmonics.

The difficulty of learning goal-reaching policies is often attributed to a "curse of horizon" that manifests as bias accumulation in temporal-difference backups and noisy advantage estimates. In this work, we identify an additional informational curse of horizon in goal-conditioned policy learning, where increasing the goal relabeling horizon can significantly reduce policy generalization and performance. Through a series of controlled experiments with oracle planners, we decouple the goal horizons sampled during training from those that the policy is asked to reach at test time. Even when evaluated only on a sequence of nearby subgoals, goal-conditioned behavioral cloning (BC) policies suffer from severe, training horizon-dependent performance degradation that is mitigated by reinforcement learning (RL) objectives. We explain this phenomenon as a horizon-dependent decrease in the conditional mutual information between actions and hindsight-relabeled goals, and find empirically that both BC and RL policies trained on longer-horizon goals exhibit a shift in sensitivity from goal to state information, as measured by the policy's input Jacobians. Motivated by this observation, we find that distilling the input Jacobians of short-horizon policies into long-horizon policies yields significant performance gains, especially in combinatorial manipulation tasks. Taken together, our results highlight goal relabeling horizon as an important consideration when learning generalist policies from offline data.

In this paper, we present robust portfolio selection models by incorporating a reward and penalty mechanism into portfolio management. We assume that the joint distribution of the losses of the underlying risky assets in a portfolio is uncertain but lies within a multivariate distribution set. Our goal is to identify optimal portfolio allocations by minimizing the worst-case conditional value-at-risk (CVaR) of portfolio loss under the reward and penalty mechanism and distribution uncertainty. Our models can also be used to investigate the problem of how to balance portfolio losses with their associated downside risk in portfolio management. We first derive an explicit closed-form expression for the worst-case CVaR under the reward-penalty mechanism, which generalizes several existing models and results regarding the worst-case CVaR, such as those studied in Jagannathan (1977), Chen et al. (2011), and Cai et al. (2024). We then apply this expression to obtain optimal portfolio allocations that minimize the worst-case CVaR under both a classical mean-covariance-based multivariate distribution set and a generalized mean-covariance-based multivariate distribution set introduced in Kang et al. (2019). Additionally, we utilize real market data to illustrate the application of the proposed models and the corresponding optimal portfolio allocations in portfolio management. Our empirical experiments show that portfolios based on the proposed models have the potential to outperform those based on several existing related models. Furthermore, the results demonstrate that incorporating downside risk into portfolio loss helps better manage risk and can achieve higher investment returns than considering either downside risk or portfolio loss alone. Moreover, our experiments reveal the trade-off between improving expected portfolio return and controlling the worst-case portfolio CVaR.

In two-agent dose-finding trials, the next cohort should help identify a combination for final selection. We studied a constrained knowledge-gradient (cKG) rule with one-cohort lookahead that updates independent Gaussian-process models of efficacy and continuous toxicity, reapplies a probability criterion for mean toxicity, and evaluates the resulting selection. We derived a deterministic calculation over a fixed set of dose combinations, holding fitted model parameters fixed during each hypothetical update. We compared cKG with constrained expected improvement (cEI) and two toxicity-only rules, targeted mean squared error (tMSE) and entropy, in four synthetic scenarios. In the primary obstructive sleep apnea (OSA)-derived scenario, averaged equally over strata and five probability cutoffs, cKG assigned fewer participants to combinations above the true mean-toxicity limit than tMSE (17.92% versus 27.08%), but selected such combinations more often at trial completion (18.80% versus 11.85%). Compared with cEI, cKG had higher mean simulated reduction in the 4%-desaturation apnea-hypopnea index (AHI4) at final selection (7.46 versus 6.72 events/hour), more above-limit final selections (18.80% versus 10.50%), and more above-limit assignments (17.92% versus 15.10%). Across scenarios, its efficacy advantage over cEI was smaller under stricter toxicity criteria. Continuous outcomes, uncalibrated toxicity limits, and a rule that still selects a combination when none meets the criterion limit clinical interpretation. Allocation and final-selection toxicity should be reported separately, alongside efficacy.

Recent advances in mean-field games and its multi-population variants enable large-scale heterogeneous multi-agent systems to be modeled through representative agents and their associated mean-field distributions. However, existing approaches do not explicitly account for uncertainty in the behavior of other populations. To this end, we introduce a new paradigm: risk-averse multi-population mean-field games, where each population optimizes a worst-case expected reward over dynamically feasible ambiguity sets of mean-field flows of a subset of the other populations. Employing an occupation-measure formulation along with tools from set-valued analysis, we establish, under mild assumptions, several theoretical properties of the multi-population game, including the geometric properties of the ambiguity sets and the existence of a novel risk-averse multi-population mean-field equilibrium. Further, we derive contractivity results of the fixed-point operator under entropy regularization and show that it can be utilized to learn the equilibrium. Finally, we propose a risk-averse fictitious-play scheme and show that exploitability decays to zero, despite the additional nonlinearity introduced by the worst-case objective. We report several numerical experiments to illustrate convergence and risk-averse behavior.

Agentic systems use large language models (LLMs) to carry out concrete tasks. Prior work often borrows abstractions such as scheduling, caching or isolation piecemeal from operating systems, so the mechanisms it builds share little common ground, and the shared view of the two forms of agentic system, Workflows and Agents, is limited. We construct an abstract machine that provides both. We treat the LLM as an Oracle and extend a two-stack pushdown automaton with one instruction, which hands the Oracle a whole stack as its query and appends the answer to that same stack. The machine thus performs two computations, the Oracle's and a Turing-complete one that we call the Priestess. A stack that the program only appends to grows autoregressively, as an agent's context does. Two symmetry breakings, S in storage and T in transitions, make a Priestess program the operating system of the programs the Oracle runs, and produce the Agent and the Workflow as the two placements of a task's program. For internally autoregressive Oracles, the two computations synchronize at the end of every answer under certain conditions, and through that synchronization we model caching and analyse scheduling. No guarantee that holds for every Oracle can fix which content crosses between the two computations, but such a guarantee does fix the boundary itself. The construction V fits the machine to a von Neumann computer. To show that it is realizable, we propose ArchNights, an extended RISC-V ISA and a Linux-style operating system implementing the machine by design. ArchNights-SE runs on gem5 as a computer system, becomes an agentic system when it runs an LLM as the Oracle, and will be open source. Agentic systems can then be designed as computer systems are. With a foundation built and a unified view, future work can share invariants and bounds, each with its conditions.

Representation Autoencoders (RAEs) generate images from pre-trained visual fea- tures, but their dense token grids make generative modeling expensive. Motivated by local feature correlations, we introduce PoolDINO, a learned affine pooling operator that merges neighboring tokens. Training the pooling operator jointly with the RGB decoder preserves the standard two-stage RAE procedure without a separate feature auto-encoder. On ImageNet-256, 4x token compression retains comparable generation quality under internal guidance, while 16x compression trades some quality for greater efficiency. At a fixed budget of 100 sampling steps, latent-sampling throughput increases by 3.7x and 9.0x, respectively, relative to the unpooled baseline. Classification and dense prediction evaluations show that comparable guided generation quality can coexist with weaker performance on other tasks.

We study continuous embeddings from two classes of generalized weighted Morrey spaces into generalized Morrey spaces. The Komori–Shirai-type spaces use the weighted measure of cubes in their normalization, whereas the Samko-type spaces use the Lebesgue measure. We first consider generalized Morrey spaces on bounded domains. In the non-radial setting, we establish embedding results under local radial comparability assumptions on the Morrey functions. As a consequence, we characterize the corresponding embeddings in the radial setting for the full range $0<p_1,p_2<\infty$. We then study embeddings from generalized Komori–Shirai-type weighted Morrey spaces associated with piecewise power weights and general Muckenhoupt weights. For the Samko-type scale, we treat both generalized and classical weighted Morrey spaces, again considering piecewise power weights and general Muckenhoupt weights. Finally, assuming that the weight is comparable to a positive constant on some cube or ball, we show that the continuous embeddings under consideration are not compact.

Modern web agents built on large vision-language models process webpages, select relevant UI elements, and translate model outputs into browser actions. Existing visual red-teaming approaches use adversarial visual content to manipulate this process. However, they primarily target model inference and do not explicitly account for structured input processing or action post-processing. Consequently, model-level success does not establish control over browser execution and cannot reliably characterize end-to-end agent robustness. To address this gap, we formulate red teaming for vision-grounded web agents as an end-to-end grounding-to-execution problem, and introduce WebMirage, a framework that crafts localized visual perturbations that cause agents to select attacker-controlled content and execute the corresponding browser action across varying webpage renderings. It uses a role-slot abstraction and webpage recomposition to capture competition among webpage elements, and dataflow analysis to align optimization with action post-processing. We evaluate WebMirage across four agent configurations and six VLM backbones on 2,250 tasks covering 13 public websites and a sandbox benchmark. WebMirage achieves an average attack success rate of 91.9%, compared with 17.4% for the strongest baseline, and remains effective against three agent-level defenses.

Self-improving LLM systems propose changes to themselves and keep those that score better on a small evaluation set. We treat this keep-if-better step as selection under measurement noise, model the correlated errors of the candidates in a single decision, and study empirically what happens when the evaluation set is reused. In runs where Qwen models rewrite their own instructions and every candidate is also scored on 600 held-out items, most proposals after the first are harmful, and the model gives the size of the winner's curse of a generation's best candidate. With a prior from a separate pilot, it matches the average overstatement of first-generation commits in native loops, though not setting by setting. In a pre-registered study, the final selection-set score of greedy loops exceeded held-out accuracy by 13 to 20 points with 16 selection items and by 1 to 5 points with 256. Held-out gains grew with the selection set on TREC but not on GSM8K, and the tested acceptance rules did not beat greedy acceptance over whole runs. Gains measured on the selection set also exceeded held-out gains when a current model refined a competent instruction, and in the validation scores of GEPA and MIPROv2. Scoring the starting and the current instruction on 64 items never used for selection removes the average bias of a loop's reported gain, but single estimates remain off by about 6 points. Self-improvement studies should report held-out gains with their uncertainty.

Emergent generative AI technologies, while remarkably beneficial across several tasks and domains, can suffer from unforeseen vulnerabilities and difficult-to-anticipate failure modes. Such failure modes can disparately affect vulnerable communities. This paper investigates two well-known vision language models (VLMs) and inspects how these models encode protected identities while generating images. We show that when VLMs are requested to transform human face images to reflect race identities (e.g., Hispanic, Black, White, or Asian) and sexual orientation (e.g., gay or straight), VLMs often comply with such requests. We demonstrate that compliance (and refusal) is non-deterministic, group-sensitive, and generated images are statistically distinguishable, often accentuating deep-seated stereotypes. We further show that representation biases introduced during image generation can re-emerge in textual reasoning and characterization tasks. Finally, we show that this worrisome finding is not limited to racial identity or sexual orientation, image generation can perpetuate bias around criminality, raising concerning possibilities of AI-assisted profiling.

Given the challenges of pharmaceutical development, there is increasing interest in computational approaches for identifying novel drug targets. Transcriptomics technologies enable the analysis of disease mechanisms at unprecedented resolution, while spatially resolved platforms further preserve the spatial organization of cells within tissues, allowing analysis to focus on specific regions of disease localization. To derive actionable therapeutic targets from these data, it is necessary to identify mechanisms that recur across a population rather than those observed in individual patients, which requires integrated analysis across disease conditions and donor cohorts. Here, we present the groundwork for an AI co-scientist framework for the sequential analysis of multi-batch spatial transcriptomics data that integrates interpretable mathematical models, including optimal transport, tensor factorizations, and topological data analysis, with LLM-powered agentic interpretation. Leveraging the complementary information provided by these mathematical models, the agents sequentially identify regions in which the disease is localized, molecular targets within these regions, and the direction in which each target should be modulated. Applied as a proof of concept to a public Stereo-seq study of the Alzheimer's disease (AD)-affected prefrontal cortex, we demonstrate that the agents converge on the white matter as the primary locus of the moderate-disease signal and identify a loss of myelin programme usage despite increasing oligodendrocyte abundance. The agents subsequently propose 29 candidate therapeutic targets, led by an oligodendrocyte differentiation module on which multiple mathematical perspectives converge, and flag 20 genes whose relevance to AD has not been previously reported. These results are hypothesis-generating and motivate further validation in larger, independent cohorts.

Optical chemical structure recognition (OCSR) turns a picture of a molecule into a machine-readable string, and it stands between decades of printed chemistry and any computational use. On real documents individual models remain unreliable. Over a 1000-molecule sample of MolParser, three readers score $46.0\%$, $43.4\%$ and $42.4\%$ exact match, yet at least one of them is right on $60.8\%$ of molecules. That gap is an ensembling problem, but not one voting can solve, because the readers seldom agree and seldom err together. We place the readers inside a multi-agent scaffold. Four language-model agents with cheminformatics tools argue about a candidate structure, write findings to an append-only knowledge base, and are held to render-and-compare gates that admit a correction only when it is demonstrably better. The debate reaches $61.9\%$, beating the best single reader by $15.9$ points and plurality voting by $9.2$ points ($p<10^{-9}$ for both, robust to correction for all $27$ tests we report), and is indistinguishable from the any-reader-right ceiling. That aggregate hides two opposing effects. On the $661$ molecules with no stereochemistry the debate is $3.9$ points above the ceiling ($p=0.004$), constructing structures no reader proposed, while on the $339$ stereochemical ones it falls below. We then swap the arbitrating model. A stronger arbitrator adds $2.2$ points (nominal $p=0.039$; neither this nor the $+3.9$ survives Bonferroni correction over all $27$ tests, so both are suggestive), and its mechanism is visible: it opens a correction on $423$ of $1000$ molecules against $120$, and an interaction test indicates it depends less on the readers (nominal $p=0.018$, not correction-robust). We conclude that the useful unit of progress is the arbitration policy rather than the reader, and that OCSR accuracy is better read against the union oracle than against the best single model.