Papers

This paper studies a finite-horizon mixed $H_2/H_\infty$ feedback Nash game for stochastic evolution equations on a separable Hilbert space. The drift generator is unbounded, the remaining coefficients are bounded, and the one-dimensional Brownian diffusion depends on the state, control, and disturbance. The $H_2$ channel is an LQ state–control energy. For the disturbance channel, the stochastic LQ uniform-convexity characterization yields equivalence between strict induced $L^2$ attenuation and unique strongly regular mild Riccati solvability. Simultaneous bounded-generator approximation of the Lyapunov equation and the state justifies quadratic identities for strongly continuous mild operator solutions. These identities verify both Nash inequalities and full-output strict attenuation from a strongly regular coupled Riccati pair. An invertible feedback block and a contraction argument establish locally unique coupled solutions on a sufficiently short terminal interval for every positive attenuation level. For a specified stochastic heat-equation model at $\gamma=0.09$, a uniform invariant rectangle further proves full-horizon existence, bounded infinite-dimensional feedbacks, and operator-norm spectral convergence. Numerical computations reproduce the projected gains and compare selected best responses. Global coupled solvability for general coefficients remains an explicit hypothesis.

There is considerable interest in using time-varying electricity prices to shape consumer demand response, and better align energy demand with renewable production. However, optimal prices generally vary over time in response to complex signals such as weather forecasts, sunrise/sunset times, and day-of-week patterns; and existing methods are not able to make efficient use of such rich contextual information. Here, we propose a neural-network-based algorithm for contextual energy pricing, modeling pricing as a Stackelberg game and leveraging a mean-field solution representation from Mehrabi et al.(2024). The approach learns constrained mappings from contextual features to feasible price signals. We validate our approach by simulating the energy grid in several US cities, and show that incorporating contextual information can considerably increase the value of the demand response programs.

We present the Motion-Aware Audio-Visual Complexity Metric (MAV-C), a reference-free framework for the joint objective estimation of audio-visual complexity. The metric combines entropy-based audio features (temporal, spectral, and spatial) with visual features (Sobel gradient magnitude, chromatic uniqueness, and optical flow) via a parametric fusion stage, producing a continuous joint complexity score CAV (t) [0,1]. We validate MAV-C on two datasets: a controlled synthetic corpus (SYN) of stimuli with known signal characteristics and a naturalistic gameplay corpus (GAM) of 60 clips drawn from the SAFEPLAY-X dataset. On SYN, the metric exhibits strong validity: the audio score CA and visual score CV are each insensitive to changes in the opposite modality (CoV < 0.003), the joint score CAV spans [0.00,0.90] across all parameter combinations, and single-axis feature sweeps produce monotone trajectories (Spearman up to 0.995). On GAM, CV differs significantly across content categories (Kruskal-Wallis p = 0.021) while CA does not, and the two sub-scores are uncorrelated (r = 0.03), confirming they operate on independent signal dimensions. OFAT sensitivity analysis identifies a two-tier parameter hierarchy, with modality balance (wa) and visual regularization (v) as most significant tunable parameters. Full subjective calibration is planned as future work.

End-to-end backpropagation has been the dominant mode of training in deep learning, allowing for the coordination of parameter updates across layers of a neural network. Prior studies have explored alternative – and, in some cases, simpler – training mechanisms, showing that they can sometimes achieve performance similar to backpropagation. However, the architectural conditions under which locally optimized networks, which avoid end-to-end backpropagation of error, can learn representations comparable to those learned through end-to-end training remain unclear. We aim to answer this question in the context of self-supervised learning, an important framework for large-scale pretraining in artificial intelligence. Here, we investigate how network width and depth affect the efficacy of greedy layer-wise and end-to-end self-supervised training in convolutional networks. We find that in wider networks, the benefits of end-to-end backpropagation over greedy layer-wise training shrink: in relatively shallow and very wide networks, we even observed higher performance in models trained with greedy layer-wise training. Subsequent analysis of the representations formed by these networks shows that very wide greedy-trained networks exhibit more favorable representational geometry than do networks trained end-to-end with backpropagation. This work shows that width can compensate for restricted credit assignment and identifies differences in representational geometry as a potential mechanism for their improved performance.

Aerial base stations (ABSs) can rapidly establish connectivity in ad hoc, infrastructure-deprived environments, but their broadcast, line-of-sight transmissions leak far beyond the intended service area, exposing communications to passive eavesdropping and interference. Prior physical layer defenses based on cooperative jamming typically assume known eavesdropper locations, simplified statistical channels, or continuously repositioned jammers. We instead pose the problem as a radio frequency (RF) containment: confining usable signal to a user-defined, arbitrarily-shaped target zone while denying it elsewhere independent of eavesdropper location. We present ARCTAN, a gradient-based optimization framework that jointly optimizes the position, orientation, and transmit power of stationary ABSs and cooperative jammers (CJs) by backpropagating through site-specific, differentiable 3D ray traced channels. Evaluated in a high-fidelity digital twin across three target zone geometries, ARCTAN achieves a mean in-zone SINR of approximately 10 dB while reducing mean out-of-zone SINR from 13-16 dB to -3-5 dB, and suppressing signal-leakage ratios from over 93% to below 46% requiring at most 10 of 12 candidate CJs.

Recent theoretical work identified fundamental properties of representation geometry that shape inference ability of deep neural networks. These include signal-noise factorization (SNF), the ability to segregate signal from noise, and signal-signal factorization (SSF), the ability to segregate task-specific and task-irrelevant signals. Here, we built regularizers that reinforce these two properties during training. We compared networks trained with these regularizers to $L_2$-regularized baseline networks on the CIFAR-100 classification task to understand how our regularizers shape representation geometry and impact performance on a well-known computer vision baseline. Enhancing SNF via regularization improved model performance but enhancing SSF did not. Motivated by biomedical applications, we investigated how our regularizers affected performance on the BloodMNIST dataset treated with MedMNIST-C corruptions at five severity levels, and found even larger performance gains using the SNF regularizer. To understand the mechanism by which SNF-regularization produces improved performance, we analyzed the nuisance subspaces across regularization regimes, finding that the SNF-regularized models represent noise in distinct subspaces, separate from class-relevant signal. Because this geometry is explicit, the dominant corruption-induced directions can be estimated on held-out data and projected out of the representations. This manipulation led to a substantial gain in accuracy. These results show that regularizers that enforce signal-noise factorization can produce substantial improvements on computer vision tasks that contain out-of-distribution image distortions at inference time. They also highlight how shaping representations affects model performance: isolating nuisance variables from categorical ones is more important than maintaining factorized representations of categorical variables.

Uncovering new phenomena in nature requires experiments. For centuries, their design has been exclusively a human endeavor. Today, a new paradigm is emerging in which artificial intelligence and computational methods can design experiments themselves. Realizing this form of automated scientific discovery requires simulators that are sufficiently expressive to represent the diverse physical processes and couplings from which new experiments can be constructed. In this direction, we introduce esQueranto, a differentiable software that brings photon-number quantum optics and structured-light propagation into a common description, allowing the spatial evolution of light to directly influence non-classical quantum states and their interference. esQueranto is implemented in JAX, providing automatic differentiation and hardware-accelerated evaluation for optimization and automated search. We demonstrate the framework across a broad range of quantum-optical applications that exercise complementary aspects of its physical description. By combining quantum and structured-light physics within a single differentiable simulator, esQueranto takes an important step towards the dream of a foundational simulator in physics.

Standard 3D Gaussian Splatting (3DGS) learns geometry and appearance jointly from RGB supervision, making it difficult to isolate how luminance and chroma contribute to the learned representation. We study this by training models under different channel supervision, freezing their non-appearance parameters (position, scale, rotation, and opacity), and re-estimating appearance with the same solver before comparing held-out reconstruction. Across eleven benchmark scenes with four independent runs each, geometry learned from luminance alone supports held-out reconstruction 0.085 dB below RGB-trained geometry on average. If chroma is deleted from a trained model, a sufficiently expressive solver can re-fit it on the frozen geometry to the original quality or slightly better. Higher-order spherical harmonics contribute much more reconstruction quality to luminance than to chroma, improving PSNR by 1.44 dB versus 0.19 dB on average, although on mirror-like surfaces hue does still change with viewpoint. The luminance advantage is even larger when geometry is being formed. Chroma-only supervision produces geometry 3.9-5.5 dB worse than luminance-only supervision after the same appearance solve; densification explains part of this gap. Overall, geometry formation in standard 3DGS is strongly luminance-dominated but not luminance-exclusive, and much of the chromatic appearance can be recovered after spatial support has formed.

We propose a calibration-free photonic memory bitcell based on nonresonant variable optical attenuators in the GlobalFoundries 45SPCLO platform. The cell maintains complementary electrical states through electro-optical feedback without cavity-resonance alignment. Wordline-controlled attenuators enable selective writing through differential optical bitlines, while optical readout shares the holding input and output splitters, avoiding additional read attenuators or a separate laser source. Circuit simulations demonstrate writing with $10~\mathrm{ps}$ optical pulses at $2~\mathrm{mW}$ peak power. The cell retains data as long as optical and electrical power are supplied, with retention simulated for $5~\mathrm{s}$ per logic state. The evaluated transition consumes an incremental write energy of $0.562~\mathrm{pJ}$, including circuit electrical energy and the laser contribution at 20% wall-plug efficiency. The bitcell layout occupies $147.92\times219.93~\mu\mathrm{m}^{2}$.

The recent development of wavelength-scale nonlinear waveguides highlights the need for new diagnostic methods that characterize both waveguide dispersion and fabrication-induced spatial inhomogeneities. While tightly confining waveguides enable compact and highly efficient nonlinear devices, nanometer-scale waveguide nonuniformities can limit the effective interaction length and conversion efficiency. In this work, we establish a high-throughput diagnostic method for second-order nonlinear waveguides based on near-degenerate three-wave mixing, hereafter referred to as sum-frequency tomography (SFT). In contrast with conventional diagnostics based on second-harmonic generation, SFT produces two-dimensional transfer functions that allow dispersion orders to be characterized separately from longitudinal phase-mismatch variations. We validate this technique using commercially available PPLN waveguides and show that SFT recovers the group-velocity mismatch between the interacting waves, the group-velocity dispersion of the fundamental, and a low-order polynomial approximation of the longitudinal phase-mismatch profile. SFT avoids the experimental complexities of phase-sensitive or spatially resolved measurements, enabling the characterization of individual waveguides in seconds.

Magnetoencephalography (MEG) is an imaging technique that offers a non-invasive, millisecond-resolution view of human brain activity. The increasing availability of MEG data presents an opportunity to take advantage of a recent advance in artificial intelligence, namely self-supervised foundation models. Existing foundation models for MEG (and electroencephalography) have been built on a transformer architecture. Here, we introduce MEG-Mamba: a generative foundation model for neural activity (source reconstructed, parcellated MEG) built on a Mamba architecture. MEG-Mamba is trained to predict the next token of a discretised MEG signal, conditioned on a brain region and recording session. MEG-Mamba surpasses the generative fidelity of a transformer-based alternative (MEG-GPT) while pre-training in less time (22 vs 400 GPU-hours) and modelling a longer context (4 vs 0.32 s). We evaluate MEG-Mamba by examining the spatio-spectral characteristics of the data it generates and by interpreting its learned embeddings. Furthermore, we demonstrate that lightweight stimulus conditioning (LoRA) can be used to generate realistic task-evoked responses that are not included in pre-training. Our results suggest that Mamba is a promising backbone for scaling neural foundation models.

The study of continuous spin particles, as general low-energy implications of relativity and quantum mechanics, has intensified in recent years with potential experimental signatures. In this note we show that these particles are absent in perturbative super-string theories, generalising a result in arXiv:1302.4771 to the supersymmetric case. This stands out as a general low-energy prediction of all perturbative superstring constructions.

The terahertz (THz) region is highly attractive for imaging because it penetrates many optically opaque materials and can distinguish them by composition. Among the various methods that allow us to access this spectral region, THz time-domain spectroscopy (THz-TDS) is uniquely suited to retrieve broadband information, recording the full waveform of the THz electric field to recover both amplitude and phase with high sensitivity. Translating this capability into fast, wide-field imaging has however remained a major challenge, largely because the low average power of conventional TDS sources forces a trade-off between acquisition speed, field of view, and sensitivity. Here we demonstrate a hyperspectral THz camera in which a high-average-power, pulsed broadband THz source, delivering 38.3 mW at a 400 kHz repetition rate, drives parallel electro-optic detection over a wide field of view at high acquisition rates. The system simultaneously records the full THz waveform, in amplitude and phase, across 145,000 pixels over a ~14.5 mm field of view at an acquisition rate of 2 Hz. These results demonstrate the potential of high-average-power THz-TDS for broadband, high-speed imaging, opening applications such as inline material-resolved inspection and the spectroscopic imaging of irreversible, second-timescale processes such as hydration in biologically relevant systems.

The development of automated cerebral microbleed (CMB) detection models is hindered by the low prevalence of CMBs and the high cost of expert annotation. To address this limitation, we developed a synthetic CMB generation pipeline and investigated the effectiveness of synthetic lesions for training deep learning detectors. Models trained solely on synthetic data achieved substantial detection performance, reaching approximately 87% of the lesion sensitivity of their real-trained counterparts. We further investigated the complementary roles of synthetic and real data under different training paradigms, revealing that a significant performance gap remains. Moreover, we found that successful synthetic-to-real transfer is strongly dependent on the downstream detection architecture, providing new insight into both the potential and limitations of synthetic data for CMB detection.

Every protein has a unique stability landscape, but the physical consequences of mutation are governed by recurring biochemical constraints. We test whether a shared decoder, trained on measurements from diverse proteins, can interpret these constraints in an unseen target, enabling cross-protein transfer for initial experimental round prescreening. We present StabilityArc , which maps frozen ESMC-600M residue representations through a shared RoPE transformer to an Lx20 matrix of substitution effects; a symmetric, contact-aware residual aids in predicting epistasis in simultaneous substitutions. In 66 strict leave-one-protein-out evaluations covering 134,794 ProteinGym variants, StabilityArc achieves 0.7134 Spearman correlation, exceeding the strongest zero-shot baseline, ProSST-2048 (0.6526), by 0.0608. We further explore the utility of this method by providing the score as a prior for Kermut, achieving Spearman correlation of 0.8280 across three supervised split schemes, improving on Kermut's reported 0.8167.

Let $(\xi_t)_{t \geq 0}$ be a continuous-time Markov process taking values in a countable state space $E$. We consider $k$ copies of $\xi$ started from distinct positions, and write $X = (X_t)_{t \geq 0}$ for the resulting process of configurations in $E^k$. We compare two forms of interaction between the particles: the first is conditioning the particles to never collide (noncolliding dynamics), and the second is simply switching off any jumps that would cause two particles to occupy the same position (exclusion dynamics). We show under mild conditions on the tail probabilities of the collision time that the dynamics are related by an explicit exponential martingale change of measure of one another. This generalises a recent result in the integrable probability literature relating TASEP on the ring to asymmetric walkers conditioned to never collide.

Negative Kirkwood–Dirac quasiprobabilities are credited with the quantum advantage of interaction-free measurement and postselected metrology. We test that claim on the task interaction-free measurement was designed for: learning about a sample while it absorbs as little light as possible. The quasiprobability distribution of path and output gives the linear response to a weak sample: its real part responds to absorption, its imaginary part to a phase. For a pure single photon passing once through any interferometer, the Fisher information about transmittance per absorbed photon is the direct-transmission value minus a sum of squared imaginary parts: no interferometer or readout beats direct transmission. Negativity only concentrates information in weakly lit outputs. These imaginary parts depend on the output basis, not the sample, so an optimal basis with all terms real always exists. A single photon beats direct transmission only through a phase response, which for an optical line needs a thick or inhomogeneously broadened sample, or through many passes when absorptions go uncounted, as a critically coupled resonator shows. For detecting a weak absorber, negativity gives a single-shot advantage that does not accumulate over repeated trials. The advantage that accumulates comes from outputs dark without the object, and needs no negativity.

Excited organic dyes chemically bound to an icosahedral virus shell can spontaneously synchronize and emit a superradiant burst of light. The mechanism that allows it to occur at room temperature, despite collisional dephasing and inhomogeneous broadening, is not understood. Few other known low-dimensional condensed matter systems exhibit room-temperature superradiance (SR). In the fluctuation-prone environment of molecular-based systems, electron–vibrational coupling has been proposed to play a central role. Here, we addressed experimentally the existence of underdamped virus capsid phonon-like modes and theoretically their involvement in the transition from uncorrelated to cooperative emission. We find that SR-active dye-labeled capsids exhibit underdamped low-frequency oscillations that modulate the dye excited-state response, and modeling shows how such capsid motion can promote the collective polarization required for superradiant emission.

The Straight-Through Estimator (STE) is a widely used heuristic for Quantization-Aware Training (QAT), but its surrogate gradients can exhibit substantial mismatch with the underlying quantized objective, leading to noisy updates and parameter oscillations, particularly in ultra-low-bit regimes. We propose the Quantization-Aware Minimal-Norm Optimizer (Q-MINO), a temporal bundle method that combines gradient consensus, state-drift regularization, and an alignment constraint to construct stabilized, minimum-norm update directions from recent optimization states. Q-MINO solves the resulting constrained subproblem using a warm-started Frank–Wolfe procedure with a feasible fallback initialization. Theoretically, via a stochastic Lyapunov Kurdyka–Łojasiewicz (KL) framework, we show that Q-MINO achieves asymptotic neighborhood convergence. Moreover, we detail numerical experiments with Q-MINO at various quantizations.

Personalized image generation has remained narrowly focused on conditional synthesis from curated visual exemplars, rather than capturing who a user is. In practice, however, a user's personal context is much richer, comprising reviews, posts, images, captions, and metadata accumulated over time. A truly personalized generator should leverage this history to produce images aligned with the user's lifestyle and aesthetic preferences. To this end, we introduce the first unified benchmark for personalized image generation from user histories. The benchmark comprises two complementary tasks and a multi-axis evaluation protocol that assesses target fidelity, visual quality, user distinguishability, semantic alignment with the user's history, and task-specific utility. Grounded in real-world e-commerce and social media settings, the benchmark includes: (1) Personalized Scene Generation, which places a given object in a scene that reflects a user's preferences and lifestyle, motivated by personalized product presentation; and (2) Personalized Creative Generation, which generates a novel image on a specified topic that is faithful to a user's aesthetic and visual identity, motivated by social media content creation. We further propose PEARL, which couples a multimodal reasoner with a frozen image generator in an interleaved reason-reflect loop optimized with differential data reward. Across both tasks, PEARL outperforms strong baselines, achieving an average improvement of 15% across personalization metrics.

Constraint-preserving mixers keep QAOA within a feasible subspace, but feasibility and global mixer connectivity do not determine which feasible configurations are available to a particular shallow circuit. We study constrained QAOA at the level of the circuit actually executed: a specified feasible initial state, finite depth, and finite schedule of mixer interactions. We define the finite-depth accessible set $R_p(M,x_0)$ and show that the computational-basis support of the ideal QAOA circuit is contained within it. This gives a parameter-independent objective ceiling $f_R^\star=\max_{x\in R_p}f(x)$ and a normalized reachable-quality diagnostic $Q_R$, separating accessible volume from the objective quality contained within it. We evaluate this framework using resource-matched block-local XY mixer schedules for constrained portfolio optimization and synthetic block-constrained quadratic problems. Across 207 favorable-tail CVaR QAOA configurations, $Q_R$ shows the strongest association with realized shallow-QAOA performance among the structural diagnostics considered, with Pearson and Spearman correlations of $0.7003$ and $0.8057$; the relationship remains positive under controlled and dependence-aware analyses. The distinction persists across changes in feasible initialization, portfolio size, and objective family. A compilation-only study further shows that, for the dense portfolio Hamiltonians considered here, phase-separator routing dominates the compiled two-qubit resource scale, while mixer schedules with comparable compiled two-qubit cost can exhibit markedly different finite-depth accessible quality. These results motivate objective-relevant finite-depth accessibility as a complementary diagnostic for resource-bounded constrained QAOA.

Modern neural text-to-speech (TTS) systems achieve remarkable acoustic fidelity but act as black boxes, offering little interpretable control over the vocal tract filter. We propose a controllable source-filter TTS architecture grounded in articulatory kinematics. An Acoustic-to-Articulatory Inversion (AAI) model, enhanced by large-scale pretrained representations, generates kinematic pseudo-trajectories for a large TTS corpus. These trajectories condition the filter response, while predicted pitch and energy contours parameterise the glottal source. Source and filter are predicted independently, and the source is refined by an Optimal Transport Conditional Flow Matching (OT-CFM) module before recombination into the final spectrogram. Our model achieves intelligibility and naturalness competitive with similarly sized baselines, with only a modest spectral fidelity cost in exchange for explicit control. Evaluations reveal clear source-filter disentanglement, enabling stable prosodic scaling and cross-speaker source/filter recombination, where F0 remains tied to the source speaker while vocal-tract characteristics follow the filter speaker. Finally, direct manipulation of articulatory trajectories enables fine-grained control, such as accent modification, offering a new direction for interpretable speech synthesis. Audio samples: https://coding-phoenix-12.github.io/ArticulatorySFTTS/

For a graph G, the Boolean quadric polytope P(G) is the convex hull of the binary solutions of $y_{ij}=x_ix_j$ for $ij\in E(G)$, and Q(G) is its standard linear relaxation. For series-parallel graphs, Q(G) together with the odd-cycle inequalities describes P(G). Lee and Skipper showed vol(Q(G)) is polynomial-time computable for bounded treewidth and gave a closed formula for vol(P(G)) when G is a cycle. We resolve their question of giving an efficient algorithm for vol(P(G)) on series-parallel graphs. With $d=|V(G)|+|E(G)|$, we compute it with $O(d^7)$ arithmetic operations, and $O(d^5)$ if G is a cactus. The algorithm is a dynamic program over the series-parallel decomposition. The same framework computes vol(Q(G)) (hence the number of linear extensions of the incidence poset of G) with $O(d^4)$ operations, and $O(d^3)$ for cacti. With every $x_v$ fixed at 1/2, the recursion reduces to convolutions of univariate polynomials and computes the cut polytope volume of every series-parallel graph with m edges in $O(m^3)$ operations. We also study how much of Q(G) the polytope P(G) occupies. Short odd cycles matter more than long ones: in every graph, the odd-cycle inequalities of a cycle of length $\ell$ cut off at most a fraction $2^{\ell-1}/\ell!$ of Q(G). Hence vol(P(G))/vol(Q(G)) $\ge 1-\sum_C 2^{|C|-1}/|C|!$ for series-parallel G, with C ranging over its cycles. The ratio does not factor over cycles sharing a vertex: in a flower of $k$ copies of $C_\ell$ it decays like $\rho_\ell^k$ for an explicit rational $\rho_\ell$ below the ratio of $C_\ell$. Nevertheless, the small fractions cut off by many long cycles compound: in the worst case, triangle inequalities, or odd-cycle inequalities up to any fixed length, close no fixed fraction of the gap between Q(G) and P(G), and the volume ratio can be exponentially small in d.

We consider multi-dimensional quasiperiodic Schrödinger operators with monotone sawtooth-type potentials in the perturbative localization regime. We prove that a given gap is closed iff a corresponding (via a natural gap labeling) eigenfunction vanishes at the origin. Using this, we show that for a generic choice of the sampling function $f$, the spectrum of such an operator is a Cantor set, with all possible gaps being open. We also provide a class of functions $f$ for which some gaps are closed. Finally, we observe that certain gaps are more difficult to close than the others.

The Navier-Stokes/Mean Curvature Flow system describes a two-phase flow of incompressible, viscous and immiscible fluids separated by a sharp interface, whose evolution is governed by a convective mean curvature flow equation, coupled to a two-phase Navier-Stokes equation accounting for surface tension. Among others, this model describes the dynamics of dry foams and it also arises as a sharp interface limit of Navier-Stokes/Allen-Cahn systems with non-vanishing mobility. Despite this wide range of applications, global-in-time existence of any kind of weak solutions remains a challenging open problem. In this work, we introduce a novel notion of varifold weak solution in two and three ambient dimensions, based on a sharp energy dissipation principle à la De Giorgi, and we show for the first time the unconditional global-in-time existence of such a weak solution. We also show, by means of a relative entropy approach, that any classical solution to the Navier-Stokes/Mean Curvature Flow system is unique in the class of our new weak varifold solutions.

Simulated evaluation is increasingly used alongside real-world evaluation of robot policies because it is cheaper and easier to repeat; however, its value depends on how closely its outcomes track the real robot's. We test whether our reconstruction pipeline, combining metrically scaled object geometry, authored physical parameters and scene reconstruction, reduces disagreement between simulated and real robot scores relative to a default open-source recipe. We constructed two simulated versions of one bimanual robot cell: an authored reconstruction, using object geometry at estimated metric scale, projected textures, authored physics and our own scene splat; and a baseline, referred to as the default reconstruction, using the open-source recipe of a generative single-image mesh, engine-default physics and a Gaussian-splat scene. Both reconstructions use the same object photographs and scene video. Two policies ran five tasks each, giving ten task-policy pairs, which we call cells; each cell was run twenty times in each reconstruction with all other settings held fixed. Both reconstructions were scored against the same real trials, graded by a third-party evaluator. Pearson correlation between the ten simulated and real cell means is r = 0.90 for the authored reconstruction and 0.51 for the default. Mean score error is 6.97 percentage points for the authored reconstruction and 17.54 for the default, a reduction of 10.56 percentage points. These results show that improving the quality of the environment reconstruction through higher visual fidelity, authored physics and metric scale makes the simulation more faithful to the real world and narrows the sim-to-real gap. We release the harness, the per-trial scores, every reported run's configuration, and the assets and scenes of both reconstructions.

Mixed-integer linear programs (MILP) model many real-world decision problems, motivating machine-learning methods that exploit recurring structure to accelerate MILP solving. MILPs can admit many equivalent formulations: integrality-preserving changes of variables and the addition of redundant constraints can alter their formulations while preserving the optimization problem. We leverage these reformulations as a source of self-supervision for learning general-purpose representations of MILP variables and constraints. We characterize the affine reformulations that are valid for every input instance, and distinguish re-descriptions, which leave variables unchanged, from substitutions, which transform them predictably. Building on equivariant self-supervised learning, we introduce ReMILP (reformulation-contrastive MILP representation learning), which jointly trains a graph neural network and a hypernetwork to predict how variable embeddings transform under changes of variables. Without solver-derived labels, ReMILP learns representations that exhibit the intended invariance and equivariance on unseen problem classes. Across binary solution, constraint activity and integrality gap prediction, these representations carry task-relevant information when frozen and provide a useful initialization for fine-tuning.

This paper explores a reward-based policy to achieve zero-shot transfer between source and target environments with completely different observation spaces. While humans can demonstrate impressive adaptation capabilities, deep neural network policies often struggle to adapt to a new environment and require a considerable amount of samples for successful transfer. Instead, we propose a novel reward-based policy only conditioned on rewards and actions, enabling zero-shot adaptation to new environments with completely different observations. We discuss the challenges and feasibility of a reward-based policy and then propose a practical algorithm for training. We demonstrate that a reward policy can be trained within three different environments, Pointmass, Cartpole, and 2D Car Racing, and transferred to completely different observations, such as different color palettes or 3D rendering, or Stretch robot navigation in Habitat-Sim, in a zero-shot manner. We also demonstrate that a reward-based policy can further guide the training of an observation-based policy in the target environment.

Weather reanalysis products rely on computationally intensive numerical weather predictions followed by data assimilation that corrects the forecast toward observations. Recent advances in deep generative models offer a cheaper alternative that shifts much of this cost from inference to offline training. However, existing generative approaches have been evaluated on synthetic observations or under different datasets and evaluation schemes, making it unclear which design choices improve real-world data assimilation. We present the first controlled benchmark of generative data assimilation for single-time near-surface analysis from real weather station observations. Using 11,849 NOAA MADIS stations across the contiguous United States and four near-surface variables, we hold the dataset, observation operator, and deep learning architecture fixed, and measure spatial generalization at held-out stations. The benchmark compares the major design choices proposed for generative data assimilation, including diffusion versus flow matching, pixel versus latent-space formulations, and multiple inference-time conditioning strategies, against a classical 3D-Var baseline. The benchmark reveals three conclusions. First, the best generative methods outperform 3D-Var (35.7% vs. 33.3% RMSE improvement over ERA5), although 3D-Var receives the ERA5 field at the analysis time as its background and the generative methods receive none. Second, full-gradient guidance consistently outperforms stop-gradient and initial-noise optimization. Third, other choices provide little measurable benefit: diffusion and flow matching perform nearly identically under matched conditions, and latent-space variable mixing does not help. Both advantages widen when stations are sparse. Together, these results identify which components of generative data assimilation improve spatial generalization in near-surface analysis.

Controlling an agent with vision requires being able to separate useful information from irrelevant background information. JEPA-style latent world models seem like a natural approach for this, as they do not perform pixel-level reconstruction; however, they are still sensitive to these distractor signals and experience latent collapse. In this work, we introduce Controllability Factorized JEPA (CF-JEPA), a JEPA-style world model which splits the latent space into controllable and uncontrollable subspaces. This factorization allows us to capture all the distractor information into the uncontrollable region, while we use the control-relevant latent information for our task. With this, we show comparable performance across 2D and 3D control tasks under nominal conditions and improved performance under distracted conditions, where CF-JEPA is the only model that does not experience latent collapse. We also validate our model under distracted conditions for a simulated robot task, highlighting the practical application of such a scheme.