Papers

Training large language models (LLMs) entails a fundamental trade-off: memory-efficient optimizers such as Adam discard cross-parameter curvature, whereas full-curvature methods such as SOAP can accelerate convergence at prohibitive memory costs. We introduce Clean, a memory-efficient and full-curvature optimizer designed to resolve this bottleneck. Clean leverages the randomized Nystrom method to accurately approximate the left and right preconditioners in SOAP, and to reduce the optimizer's memory complexity from quadratic to linear in terms of model dimensions. We subsequently reintegrate the off-subspace components to capture curvature information beyond the low-rank approximation, preserving rich curvature at minimal memory cost. We further propose Q-Clean, a low-precision variant that aggressively compresses optimizer states. Q-Clean reduces optimizer memory consumption by over 50% compared to Muon when pre-training a LLaMA-1.3B architecture, all while maintaining strong and competitive predictive performance. Notably, Clean operates with a smaller optimizer-state footprint than standard AdamW while reaching AdamW's final performance 26% faster in wall-clock time. Furthermore, our methods uniquely enable the pre-training of a 13B-parameter model on a single 80GB GPU, providing a scalable, efficient, and accessible approach to large-scale model optimization.

Surface reconstruction under sparse-view settings remains challenging due to limited geometric cues. Volume rendering methods based on signed distance functions often produce over-smoothed surfaces, while 3D Gaussian Splatting (3DGS), though time-efficient, suffers from incomplete geometry due to the lack of reliable depth supervision and the limitation of being optimized only from given input views. In this paper, we present Sparse-GS2Mesh, a stereo-aware framework for surface reconstruction from sparse views. While 3DGS and stereo matching have been leveraged for surface reconstruction under dense view settings, we extend them to operate effectively under sparse view conditions by first initializing 3DGS using epipolar depth priors to mitigate the 3DGS overfitting problem, followed by our three key components: (I) adaptive baseline selection, (II) fine-tuning with a stereo matching network, and (III) 2D/3D co-regularized fine-tuning. Given a warmed-up 3DGS initialized with epipolar depth, the adaptive baseline selection automatically determines a baseline to synthesize for each sparse view. We then fine-tune 3DGS by backpropagating depth-refining gradients from the stereo matching network, effectively specializing the 3DGS for stereo matching. The 2D/3D co-regularization further helps obtain stable reconstruction, addressing weak geometric cues in close stereo views. Sparse-GS2Mesh achieves a 15% improvement over state-of-the-art methods in little-overlap settings and comparable results in large-overlap settings. Project page: https://yhnoh8623.github.io/projects/s-gs2mesh/.

We propose a phenomenological toy framework promoted to a minimal three-state internal sector in $(3+1)$ dimensions and realized by composite strong dynamics. It is in a genuine three-body holonomy which originates in a relativistic $(1+1)$-dimensional Dirac system. The hypothesis is that the number three is not imposed first as a gauge symmetry, but is inherited from the primitive closed cyclic structure of the three-body holonomy. This threefold structure subsequently reappears as a three-state internal space, the three-site loop, a three-flavor hidden sector, and a cubic determinant interaction. We introduce an asymptotically free hidden $SU(3)_H$ gauge theory with three Standard-Model-singlet Dirac flavors. Strong hidden dynamics generates a composite condensate, while the inherited holonomy selects its cyclic internal orientation. The anomalous cubic determinant interaction produces a holonomy-dependent potential, and dimensional transmutation generates the characteristic composite scale. The radial composite excitation along the cyclic direction, denoted by $S_3$, couples to the Standard-Model Higgs doublet through a holonomy-dependent CP-even portal. The resulting condensate shifts the Higgs quadratic term and can trigger electroweak symmetry breaking. In the classically scale-invariant reference limit, the electroweak scale is induced from the dynamically generated hidden-sector scale through dimensionless portal couplings.

Deep learning precoders, particularly those based on Transformer architectures, offer superior spectral efficiency for Massive MIMO systems but may incur high computational cost. While existing mixed-precision quantization studies on convolutional neural networks establish important energy baselines, traditional discrete search methods fail to scale to the large parameter spaces of modern Transformers. To bridge this gap, we apply a differentiable precision learning technique to massive MIMO precoding. This approach autonomously and jointly optimizes weights, activations, quantization step sizes, and layer-wise bit-widths in a single training loop, directly optimizing for energy efficiency. We further expand this approach by incorporating Neural Architecture Search across various Transformer model sizes and evaluating the impact of random versus floating-point trained initializations on resolution tuning. Ultimately, we demonstrate that the proposed training method enables the deployment of highly compact Transformer precoders, improving energy efficiency by up to 288\(\times\) compared to the classical Weighted Minimum Mean Square Error algorithm at equal sum-rate performance in a dense downtown environment.

Concurrent learning (CL) has recently been extended to Lyapunov-based deep neural networks (DNNs) to obtain parameter convergence under finite rather than persistent excitation. However, the existing first-order law converges slowly, especially for high-dimensional systems and overparameterized networks. To address this problem, this paper develops a Nesterov-accelerated concurrent learning (NACL) adaptation law for uncertain second-order control-affine systems. The developed law drives momentum dynamics with recorded state-input data, so that all hidden layers of the DNN are updated online under a verifiable finite excitation condition. A dynamic state-derivative observer is designed to reconstruct the recorded input without acceleration measurements. The momentum variable introducesa cross term coupling the momentum lag error to the state-dependent DNN Jacobian and the tracking error, and a state-dependent normalization is introduced into the momentum filter to dominate it. A Lyapunov-based analysis establishes exponential convergence of the tracking, momentum, observer, and weight estimation errors to a bounded residual ball. Simulations on a six- degree-of-freedom unmanned underwater vehicle demonstrate a 24.5% and 62.9% reduction in the root-mean-square function approximation error relative to first-order CL and to a higher-order tuner without CL, respectively.

A robotic microscope watching living cells cannot afford to look as often as it would like. Every volume it acquires costs photons the specimen does not get back, and time owed to other wells. What such a platform exists to produce is a record of individual cells through time: which cell is which from one volume to the next, and which cell divided into which two. Sampling sparsely breaks that record exactly where it matters, and the fault lies in the acquisition schedule rather than in the analysis software. We fill the gaps by reconstructing them, fitting a 4D Gaussian model to whatever volumes the hardware could afford. The model is a cloud of light-emitting blobs, each carrying a position, a shape and a lifetime. Being continuous in time, it renders any missing volume on demand, decoupling how often the robot analyses from how often it can afford to look. The microscope's point-spread function is measured from the data rather than inherited from acquisition metadata or left to the optimizer, because metadata inflates it and the optimizer cannot recover it at all: a wider blur around a smaller blob fits the images equally well. Each blob's lifetime is stored in frames rather than as a fraction of the recording, so that it denotes a fixed duration on any sequence. Unmeasured timesteps are supervised at coarse scale by a 3D U-Net that predicts the intermediate volume directly and estimates no motion field, since a dividing cell becomes two and no motion describes that. On two Cell Tracking Challenge sequences, a C. elegans embryo and a Chinese Hamster Ovarian (CHO), with fidelity scored per cell nucleus, our reconstruction holds the highest nucleus fidelity at every distance from an acquired frame, has the flattest decay across the gap, and best recovers focal planes it was never shown with graceful degradation across the gap.

Diffusion language models (DLMs) enable fast generation by predicting multiple tokens in parallel, but their practical adoption remains limited by a persistent quality gap relative to comparably sized autoregressive (AR) models. We attribute this gap to a computation-difficulty mismatch: within a partially observed sequence, some unknown tokens are easy to predict, while others require substantially more computation. Existing DLMs nevertheless apply uniform computational depth to all unknown positions at each denoising step. We introduce ALoDLM, which replaces uniform computation with token-adaptive latent recurrence. At each denoising step, ALoDLM iteratively refines latent representations and allocates computation according to token difficulty. Tokens ready to commit are fed back as discrete context, while unresolved tokens retain and further refine their latent states through additional recurrent passes. To learn token prediction and computation allocation jointly, we formulate token-wise computation schedules as latent variables and derive a conditional negative evidence lower bound (NELBO). We train ALoDLM at 1.7B and 8B parameter scales. Across eleven benchmarks, ALoDLM outperforms all evaluated DLMs and the corresponding AR baselines in average benchmark score at both scales. ALoDLM also retains fast parallel decoding, yielding a strong quality-efficiency trade-off among evaluated autoregressive and diffusion models under optimized inference engines.

The decay of out-of-time-ordered correlators (OTOCs) to a small residual value signals the scrambling of quantum information. This behavior is generic: it occurs in both chaotic and interacting integrable many-body systems, and therefore cannot by itself distinguish between them. In this paper we analytically study the late-time saturation value of a closely related observable, operator purity, and show that its scaling with system size robustly distinguishes chaotic from integrable systems. Our approach is based on the Mazur bound, which provides concrete scaling estimates and makes explicit why the locality of integrable charges matters. We support our findings with numerical simulations of one-dimensional spin chains.

Synchronous policy updates can limit the efficiency of post-training for long-horizon LLM agents. When trajectories span multiple interaction turns, completion times can vary substantially, and synchronous update barriers leave faster workers idling, wasting compute resources. Asynchronous reinforcement learning addresses this inefficiency but introduces policy lag and off-policy optimization. Evolution strategies (ES) are a backpropagation-free alternative to gradient-based training, yet require many perturbed-policy evaluations and have remained largely synchronous in current LLM post-training. We introduce bounded-staleness asynchronous ES and investigate its training behavior under stale fitness measurements. On WebShop, we systematically vary the lag and fraction of stale fitness measurements and additionally introduce two stress tests to probe extreme staleness cases. Natural asynchrony closely tracks synchronous ES, and degradation in evaluation accuracy remains small even under extreme staleness, including when all fitness measurements are stale. These findings suggest that ES can tolerate substantial policy staleness without explicit off-policy correction, supporting its use in asynchronous post-training for long-horizon LLM agents.

Personal AI agents in enterprise multi-tenant deployments share a common vector store for long-term memory. Shared embedding spaces create a surface for cross-user memory leakage: a user's query can retrieve semantically adjacent memories belonging to another user through ordinary cosine-similarity retrieval, without any exploit. We formalize this as cross-user admissibility failure and evaluate it across six experiments, plus follow-up ablations, under both sparse (TF-IDF) and production-faithful (MiniLM-L6-v2) retrieval. Non-adversarial, incidental leakage reaches 70–100% under pooled {same-team} retrieval; adversarially crafted memories achieve 90–100% top-$k$ placement, exceeding weaker keyword-based attacker baselines, with score lifts of $+0.416$ to $+0.511$ under production-faithful dense retrieval (Config B); and end-to-end response contamination reaches 5.00/5 under a production retrieval path and 4.67/5 with Claude Sonnet~4.5, with contaminated responses often scoring as helpful or more helpful than clean ones, a gap validated against human judgment. Among three architectural mitigations, only hard post-retrieval ownership gating consistently restores the clean baseline (1.00/5) across {two generation models, at a measured latency overhead of roughly 1.4~ms per query.

We study the largest size $C(n)$ of a subset of the $n\times n$ integer grid containing no isosceles triple, including equally spaced collinear triples. We prove $C(n)=\Omega\left(n\sqrt{\log\log n/\log n}\right)$ by applying the sparse hypergraph coloring theorem of Cooper and Mubayi. Elementary geometry and primitive-direction counts give maximum degree $O(n^2\log n)$ and pair-codegree at most $5n$ for the forbidden-triple hypergraph. The resulting bound improves the explicit guarantee in PatternBoost by a factor of order $\sqrt{\log\log n}$. Recovered interaction records document how a Codex literature proposal, human route selection, finite verification, and author-relayed review informed proof synthesis and repair. We connect these actions to successive proof artifacts and extract two case-derived checking practices: tracking objects, parameters and conclusions in theorem applications, and seeding a definitional omission to test a finite verifier. Separately implemented enumerators agree on complete edge sets for every $2\le n\le12$; suppressing the degenerate-case branch loses exactly the equally spaced collinear triples. These records document assistance under human direction, and finite checks support implementation consistency rather than the asymptotic theorem itself.

A molecular crystal structure is jointly described by its space group symmetry, unit cell, and the molecular alignment within the asymmetric unit. Concurrently predicting all three variables is a daunting task, as it mixes discrete symmetry choices with a high-dimensional search in the continuous space. To address this challenge, we decouple these variables using a three-step generation process. Specifically, we train a flow model to learn the conditional distribution of invariant lattice descriptors (e.g. direct- and reciprocal-lattice successive minima and Selling scalars) from a molecular graph, a Hall setting, and the number of molecules in the asymmetric unit ($Z'$). Using a two sequential quasi-random sampling processes, we first reconstruct the cell parameters that match the predicted lattice invariants and density requirements, and then conduct a molecular packing search within the give symmetry and unit cell constraint. On 84 single-component systems with $Z' \le 1$, our approach reproduces experimental matches for 83 systems; the remaining failure stems from force-field limitations in preserving the experimental structure. These results demonstrate that learned cell proposals can effectively support crystal structure prediction (CSP) for a given Hall setting and $Z'$, which may be extended to fully blind prediction with variable symmetry and $Z'$ settings in the future.

Opacity is a widely used channel in data visualization, but it remains less well understood compared to channels such as color, length, size, etc. Recent work from Meng et al. investigated the impact of opacity across competing color schemes, finding that certain color schemes were associated with better participant accuracy. We examine these effects further in a controlled two-alternative forced-choice setup to determine whether opacity differences are truly equal across possible opacity comparison ranges. In a within-subjects study with 96 trials, including two competing color schemes (best and worst from Meng et al.) and 48 opacity pairs, we find little differences between color schemes but larger individual differences in accuracy. Further, results show stable performance in middle opacity ranges, with more errors occurring when comparing extreme values. We discuss potential implications for design guidelines and further study and make our study materials, analysis scripts, and data available at https://osf.io/zv9dx/overview?view_only=38d03cea1b3d42788593e3e6b1016cfd.

We establish statistical consistency for the penalized maximum-likelihood estimator underlying CPLASS, a method for detecting changes in velocity in $d$-dimensional time-series data. The signal is modeled as a continuous piecewise-linear trajectory observed with independent Gaussian noise. Unlike classical change-in-mean models, continuity across adjacent segments couples their parameters and prevents direct application of standard segmentation arguments. Under compact parameter spaces, minimum segment-length and velocity-jump conditions, and a strengthened Schwarz information criterion penalty $\rho_k(\log n)^\gamma$ with $\gamma>1$, we prove joint consistency of the estimated number of segments and all changepoint locations in the high-frequency regime $n\to\infty$. The maximum changepoint-location error is $O_{\mathbb{P}}\{(\log n/n)^{1/2}\}$. The proof first uses empirical-process theory to establish convergence of the fitted signal, variance, and likelihood for each fixed model size. It then combines overfitting and underfitting arguments with a two-stage geometric localization analysis, yielding an initial $(\log n/n)^{1/3}$ rate that is sharpened by exploiting the local two-segment continuous piecewise-linear structure.

Quantum kernel methods are promising for near-term quantum machine learning, but their training costs are driven by the quadratic scaling of Gram-matrix construction and the measurement cost of estimating individual kernel values. In this work, we investigate one-class anomaly detection as a resource-efficient operating regime for training non-variational quantum kernel models. Using the ULB Credit Card Fraud Detection dataset, we evaluate three quantum feature maps across training-set sizes and qubit counts. Power-law saturation models reveal strong diminishing returns with increasing training data, with 500-sample models retaining approximately 90–93% of their fitted asymptotic average precision (AP). We further observe no severe exponential concentration among the evaluated feature maps over the investigated qubit range, thereby avoiding the exponentially increasing measurement requirements associated with that regime. Finally, Nyström approximation reduces kernel construction from $\mathcal{O}(N^2)$ to $\mathcal{O}(Nm)$ evaluations, with QSVR retaining 94–97% of its exact-kernel AP using only five landmarks. For ZZ-QSVR, downsampling from 4,000 to 500 training samples and applying a five-landmark approximation decreases the number of required kernel evaluations from approximately $8.0$ million to $2,485$, a reduction of more than 3,200-fold, while AP decreases from $0.812$ to $0.737$. Under the minimum IonQ pricing assumption considered here, this corresponds to an illustrative reduction in quantum kernel training cost from approximately \$1.34 billion to \$417{,}480. These results demonstrate that training-set reduction and low-rank approximation can substantially reduce the resources required to train non-variational quantum kernel anomaly detectors while retaining meaningful anomaly-discrimination performance.

Let $Q$ be a bipartite Dynkin quiver of finite simply-laced type and $\mathscr C_Q$ the corresponding Hernandez–Leclerc category. We study a two-parameter specialization of the Fedele–Hernandez toroidal Grothendieck ring and identify its generic positive part with a Cartan-type multiparameter quantum group. The structural input is the identity \[ 2D^{(2)}-D_0=\omega_Q\circ(\deg_Q,\deg_Q), \] where $D_0,D^{(2)}$ are the two effective toroidal commutation forms and $\omega_Q$ is the antisymmetric Euler form of \(Q\). It realizes the specialized character algebra, after an auxiliary scalar extension, as a graded bicharacter twist of the one-parameter algebra at $v=t_0t^{1/2}$. This yields the required root grading and PBW graded dimensions. For every source–sink edge $r\to s$, these dimensions combine with the local \(A\)–\(Y\) commutation formula to give a quadratic annihilator for \(\operatorname{Ad}_{Z_r}\). Consequently \[ \operatorname{Spec}_{Z_r}(Z_s)=\{t_0t,t_0^{-1}\}, \] without thinness or type-specific character formulas. The two spectral values yield the oriented multiparameter Serre relations. With \[ q_{ii}=t_0^2t,\qquad q_{ij}=t_0^{-1},\qquad q_{ji}=t_0^{-1}t^{-1}\quad(i\to j), \] and \(q_{ij}=1\) on nonedges, we obtain \[ U_{\mathbf q}^{+}(\mathfrak g) \simeq \mathscr K_\infty(\mathscr C_Q)\otimes_RK \] for $\mathfrak g$ of type $A_n,D_n,E_6,E_7,E_8$.

Recent studies show that artificial intelligence (AI) with language and vision capabilities still experiences limitations in spatial reasoning. In this paper, we have studied the spatial capabilities of advanced generative AI to understand the rotations of objects in 3D space, utilizing AI's image processing and language processing features. We trained and examined the spatial intelligence of a generative Agentic AI model (GPT-5.6) to understand the spatial rotation process with rotation diagrams based on the revised Purdue Spatial Visualization Test: Visualization of Rotations (Revised PSVT:R). We improvised the Revised PSVT:R by superimposing additional graphical and contextual features to evaluate how different Chain-of-Thought (CoT) reasoning strategies influence model performance. The results indicate that structured CoT reasoning improves the spatial reasoning performance of the base GPT-5.6 model in both datasets (PSVT:R and PSVT:R with coordinate system). We used three CoT approaches - (1) Structured CoT, (2) few-shot Structured CoT, and Structured CoT with Self-optimized Prompt. The three CoT approaches evaluated in this study showed no significant performance difference. Results showed that combining structured CoT reasoning with relevant contextual information leads to considerable improvements in VLM performance on 3D rotation tasks, demonstrating the potential of agentic AI for more effective spatial reasoning. However, when contextual information is removed, structured CoT reasoning alone provides limited improvement, and the models continue to exhibit notable difficulties in understanding spatial transformations. These findings suggest that effective spatial reasoning in VLMs relies on the integration of visual, textual, and reasoning-based information in future agentic AI systems for spatial intelligence.

Autoware is an open-source autonomous driving software platform widely adopted by researchers and industry developers. Originally developed primarily for public-road applications, including passenger vehicles, taxis, and buses, Autoware is increasingly being extended to off-road environments such as construction and agricultural sites. Construction sites, however, differ fundamentally from public roads and challenge many assumptions underlying conventional autonomous driving systems. They are characterized by unstructured terrain, airborne dust, continuously evolving site conditions, and construction-specific objects. This paper presents lessons learned from ongoing efforts to extend an Autoware-based autonomous driving system to large dump trucks operating at construction sites. We identify gaps between public-road and off-road construction applications across the sensing, mapping, localization, perception, planning, and control modules of the Autoware stack. We then discuss potential solutions for addressing these gaps, with particular emphasis on ongoing LiDAR-based perception pipeline development and findings from field testing. Finally, we propose a roadmap toward end-to-end autonomous driving for construction vehicles.

Multi-pass underfill dispensing couples cycle time with process reliability because the sequence determines the inter-pass flow times. This paper formulates a tri-objective multiset-permutation routing model that minimises nominal cycle time and additive duration variance while maximising reliability under hard- and soft-dwell policies; reliability combines pass, schedule-dependent flow and travel terms. Under independent durations and deterministic dwell thresholds, the variance measure equals the cycle-time variance under soft dwell and bounds it from above under hard dwell. A multiset-permutation binary-addition-tree algorithm enumerates the feasible routes without duplication and prunes prefixes by dominance. Six instances, including three BeagleBone Black benchmarks, yield reference Pareto fronts, the largest search space of $1.47 \times 10^{12}$ routes in 38 minutes on 20 CPU cores. Among simulated candidates, the nominally cost-selected routes incur maximum estimated excess costs of USD 0.0014 per board at the modelled duration variability and USD 0.0035 with it doubled. Round-robin dispatch is within 3.7 % of the reference optimum at the tested costs; NSGA-II and simplified swarm optimisation show at most 0.53 % observed cost regret while recovering about 30 % of the largest hard-dwell front. The framework quantifies sequencing trade-offs and heuristic decision regret; physical reliability prediction requires inspection-based calibration.

Referring multi-object tracking (RMOT) takes a video and a language expression as input and tracks all referred objects. Many tracking requirements involve how an object moves rather than how it appears. However, in a video captured by a moving camera, a parked vehicle may appear to move, while a moving vehicle may show little displacement. Existing RMOT methods relate motion with text but do not explicitly remove camera-induced motion. In this paper, we propose extracting residual motion across frames by estimating camera motion in driver-view videos and compare motion characteristics with the query expression. We consider the motion-matching extent and integrate it with the RMOT method's prediction result through late fusion. In the evaluation, we verify the performance gain of taking the motion compensation module as a plug-in across different RMOT hosts.

Recursive self-improvement (RSI) relies on evaluation feedback to assess progress and guide further research, yet repeatedly running complex benchmarks is costly and slows iteration. Human experts reduce this cost by selecting benchmark subsets or designing compact suites. We ask whether AI agents can automate this design process and introduce EvalResearchBench (ERB), a benchmark for autonomous evaluation research. Given target materials, development references, candidate APIs, and fixed time and API budgets, an agent called the researcher selects or synthesizes tasks, implements graders, and revises them in pilot tests before freezing an executable evaluator for coding, co-work, and reasoning. We study 9 researchers and 13 candidate models and compare each frozen evaluator with 14 target benchmarks on score concordance and pairwise agreement. The best evaluators order about 75% of candidate pairs as the targets do, below the 91% ceiling set by disagreements among the targets. No researcher leads on every metric, and the best evaluator on development targets is not the best on sealed targets hidden from the researcher. A human-designed sample of public tasks remains a strong baseline, and the evaluator with the lowest recorded execution cost attains the highest pairwise agreement. Agents repair tasks and graders through pilot feedback, yet their evaluators can still truncate answers, exhaust the evaluation budget, or let a few questions dominate a domain score.

Language models exhibit remarkable robustness, continuing to produce coherent text even when their activations are perturbed by interventions like linear steering. We hypothesize that this robustness is a result of passive dynamics, i.e., constraining mechanisms in the forward pass that funnel activations toward "good" regions that produce coherent outputs. To investigate these hypothesized error-correcting mechanisms, we probe the geometry of language model activation space by observing the action of model layers on low-dimensional curves. In doing so, we discover that model activations occur within a cluster of distinct attracting basins, which differentiate natural activations geometrically from distributionally similar synthetic activations. Applying this lens to language model steering, we observe feature-specific basins along semantic steering directions, and find that steering moves activations between these basins. To demonstrate the active role of this geometry in neural computation, we show that adaptively modulating steering strength to transport activations across basins improves inter-language steering, significantly increasing the probability of sampling tokens from the target language compared to fixed-strength steering. Our findings establish analysis of activation space geometry as a promising approach to interpreting and controlling language models.

What role do journalists play in determining the political slant of the news they produce? We develop and estimate a model where journalists and newspaper outlets contract over both slant and wages. The model implies a set of conditions under which we can consistently estimate the role of journalist preferences in driving the observed variation in slant across outlets by leveraging journalist transitions between outlets. To measure slant, we train a transformer-based, machine learning model using articles tweeted by politicians and apply it to a full-text database of 9+ million newspaper articles published in the US between 2013 and 2018. Under our model and identifying assumptions, our estimates (a) reject the hypothesis that journalists have zero ideological preferences over the content they produce and (b) imply that 10% of the variation in slant can be explained by journalists. When journalists move across outlets, their average slant shifts by 77% of the gap between their destination and origin outlet averages. In counterfactuals that reweight journalist party composition, moving from the observed left-leaning distribution to an even distribution of Democrats, Republicans, and non-partisans shifts average article slant in a more conservative direction, but only by 0.04 standard deviations in slant.

We propose confidence intervals with near-nominal coverage of variance-covariance parameters in linear mixed models, even when interest or nuisance parameters are near the boundary, that is, when some random effects have small variances or are strongly correlated. In such settings, our simulations show standard Wald, likelihood ratio, and score intervals with nominal level 95% can have coverage as low as 45% and as high as 99%, while the proposed intervals are near nominal. The proposed intervals invert a modified profile score statistic based on the restricted likelihood, with nuisance parameters estimated on an extended parameter set. In a longitudinal study of socialization in children with autism, where the estimated correlation between the random intercept and slope is near minus one, our intervals show that children who start higher improve more slowly. In a genetic study of white blood cell count in mice, they distinguish the contributions of variant classes even though several intervals include zero. Compared to universal inference intervals, which are valid in finite samples, ours are 1.7 to 3.1 times narrower, and finite where one universal interval is not, at a forty-seventh of the computing time. We establish asymptotically correct coverage uniformly in the parameter under a condition known to hold with independent clusters and crossed random effects.

Auditory attention decoding (AAD) is often evaluated on static, simplified speech scenes that poorly match everyday listening. We introduce MOV-AAD, a large-scale dataset for studying auditory attention under moving, naturalistic conversations. MOV-AAD combines 64-channel EEG with synchronized physiological recordings, including eye tracking, respiration, galvanic skin response, heart rate, peripheral oxygen saturation, body temperature, body motion, and photoplethysmography, enabling analysis of cross-modal neural and physiological markers of attention and listening effort. This dataset uses a more ecologically valid listening paradigm with dynamically moving conversational speech sources and behavioral measures of attentional engagement. MOV-AAD supports research on robust AAD in realistic spatial dynamics, multimodal attention modeling, listening effort, and intersubject neural responses, providing a resource for benchmarking selective auditory attention in naturalistic listening.

We study diffeomorphisms of Banach spaces satisfying generalized hyperbolicity: the stable and unstable subspaces are invariant only up to inclusion, need not depend continuously on the point, and expansivity may fail. On closed manifolds, this condition is equivalent to Axiom~A with strong transversality. We introduce a continuous-selection method that converts uniform orbitwise estimates into bounded continuous solutions of cohomological equations for generalized-hyperbolic cocycles. For diffeomorphisms in the uniform global \(C^1\) category, this yields continuous semiconjugacies in both directions under sufficiently small bounded Lipschitz perturbations, and topological stability in the sense of Walters. For the infinite product of one-dimensional Morse-Smale systems on \(\ell^p(\mathbb Z)\), \(1\leq p\leq\infty\), we construct compatible forward and reverse cohomological solution operators and prove structural stability, including for perturbations that couple coordinates, although the product is not expansive. For \(p=\infty\) this gives a structurally stable diffeomorphism with uncountably many hyperbolic fixed points, in contrast with the finitely many basic sets of Axiom~A diffeomorphisms. The cocycle analysis yields Lipschitz shadowing for the induced operator on bounded continuous vector fields.

Agent harnesses are systems that coordinate model calls, tool use, and task execution to help large language models complete complex tasks. To meet task requirements and address failures, these systems are often iteratively refined by amending and patching their instructions, tools, and workflows, continuously increasing harness complexity. It is therefore unclear whether some resulting harness modules are redundant, introducing substantial token overhead with little, if any, performance gain. Inspired by neural network pruning, in this paper, we study harness pruning as a means of striking a better balance between task performance and token cost. We propose SHarP (Saliency-based Harness Pruning), a simple yet effective pruning strategy based on the saliency of each harness module with respect to performance and efficiency. Specifically, we first identify tools, instructions, and supporting mechanisms as components that can be individually disabled. We then estimate the saliency of each module by ablating it and assessing its task performance and token cost relative to the full set of single-module ablations. Modules with the smallest contribution to performance or largest computational overhead are subsequently pruned. Our evaluation across various harnesses on held-out validation sets reveals a surprising finding: most harnesses that we studied are highly redundant and can maintain comparable performance and efficiency even after a substantial portion of their modules are pruned. Our pruning approach and empirical findings provide new perspectives on agent harness design and optimization.

Euler's number $e$ is often introduced through compound interest, an infinite series, an area, or the slope of an exponential function. The last viewpoint is especially attractive: choose the base $b\in (2,3)$ so that the graph of $y=b^x$ has tangent slope $1$ at $(0,1)$. However, a curious student may raise several questions that a brief introduction necessarily leaves implicit. What does $b^x$ mean when $x$ is irrational? Why is $b^x$ differentiable at $x = 0$? How can one show that the slopes corresponding to bases $2$ and $3$ lie on opposite sides of slope $1$? Finally, why is there a unique base whose slope is exactly $1$? The paper provides an answer using tools from a first course in real analysis, including limits, continuity, derivatives, and the Intermediate Value Theorem. Its primary audience is undergraduate students familiar with standard concepts from an introductory analysis course who wish to apply what they have learned to explore questions left open in earlier courses. A secondary audience is instructors who may use selected portions as supplementary material in an introduction-to-proofs course. The mathematical ingredients are classical; our aim is to organize them into a self-contained exposition and to make explicit the logical steps that are often compressed in standard treatments.

Inspired by classical Markov Chain Monte Carlo (MCMC) methods, dissipative quantum algorithms are an emerging class of methods for preparing quantum states on a quantum computer. They operate by simulating a dissipative process governed by a mathematical operator known as the Lindbladian. The performance of these algorithms can be limited by the sequential implementation of the Lindbladian's constituent jump operators, which leads to high circuit depth. To address this issue, this work analyzes a parallelization scheme to implement jump operators simultaneously. We leverage the structure of geometric locality to show that individual jump operators remain spatially localized. This exponentially reduces the circuit depth per time step, resulting in an overall polynomial improvement. Numerical experiments demonstrate a 53x reduction in query depth for a 1D benchmark system of 100 qubits. The reduced circuit depth makes dissipative algorithms more practical for near-term quantum computing.

Rydberg atom-based electrometry is a pivotal technology that enables radio-frequency (rf) field measurements traceable to the International System of Units. This is typically accomplished by converting the direct rf power measurement of classical probes to a frequency measurement referenced to two well-defined spectral features, a commercial wavemeter, or narrow cavity lines produced via clock disciplined electro-optic modulators, although there exist cases where none of these are available. In this article, we demonstrate that a known rf frequency scale can be used to infer a laser's frequency scan using Autler-Townes (AT) splitting. In other words, we show that frequency-axis calibration can be achieved using the avoided-crossing behavior of near-resonant rf radiation. The relative separation between AT-split features is measured as a function of an applied rf field's power and detuning from atomic resonance. These peak separations are then fit to a simple, two-level model using a weighted nonlinear least-squares regression analysis that utilizes three fitting parameters, one of which is the desired calibration factor. While we do not benchmark this method against an absolute frequency reference, we do show that this method is (1) consistent with calibration via known fine-structure separation to within approximately 1.11% or better, and (2) robust against substantial experimental deviation from the system's known rf power dependence.