Papers

Drug development is an expensive and lengthy process requiring developers to select drugs for human trials and diseases to target. Computational tools that can estimate before testing which drug-disease candidates are most likely to reach FDA approval could reduce development time and costs through early program advancement and disease prioritization. Yet existing approaches predict outcomes at a single level or stage or do not distinguish between disease outcomes for the same drug. We ask (1) whether FDA approval can be forecast for a drug-disease pair using only pre-clinical trial information, and (2) whether these forecasts can help prioritize diseases for the same drug. We built a cohort of 14,134 drug-disease pairs by reconstructing programs from ClinicalTrials.gov, extracting disease-level approvals from FDA label documentation, and enriching each pair with molecular, disease, ADMET, target, and pathway features. We divided the cohort into pre- and post-2019 programs and trained a classifier on pre-2019 data to predict FDA approval. On the post-2019 cohort (2,637 programs), the model would have prioritized a program that ultimately gained approval over one that did not approximately three out of four times. For drugs tested across multiple diseases, selecting the model's highest-ranked disease would have identified an ultimately approved disease 61.3% of the time, compared with 40.7% under random selection. These results show that public pre-trial information contains signal for forecasting FDA approval at the drug-disease level and can help prioritize diseases for the same drug. We provide a public web tool for SMILES- and ICD-10-based drug-disease approval forecasting.

Many systems read the latent space of a variational autoencoder (VAE) as discrete codes, and rely on each latent being read back as the code it was written as. Shannon (1948) took such a shared codebook as given; inside a VAE nothing enforces it. We prove that no item of a standard training report can reveal whether the written and read codes agree on encoded data. We construct two VAEs that reconstruct every input perfectly, infer exactly, and match on every per-example ELBO and every code histogram, yet agreement is one in the first and zero in the second. We then show that the ELBO's own terms bound how far a measured agreement can move: the variational gap bounds the shift to exact inference, and the rate bounds the shift to generation from the prior. For Gaussian posteriors, a mean-field encoder pays an exact price in the gap that depends on its choice of axes. On probabilistic PCA the gap certifies a trained encoder in all 75 settings we test, at agreement between 0.76 and 0.97. On 240 trained VAEs, a floor set by the rate and the dataset size rules out rate-side certification in advance for 66 models, and none of them certifies. The results yield a simple protocol, Certify-or-Measure, for any pipeline that reads a VAE as codes.

Modern experiments increasingly produce rich post-treatment objects—such as text, images, and video—that may provide early surrogate signals for outcomes that are delayed, expensive, or observed only for a subset of participants. A pretrained encoder can turn such an object into a strong generic embedding, but the embedding is not trained for the estimation tasks required by a causal estimator and may emphasize information irrelevant to the causal target. We propose SIRNET, a Surrogate-Informed Representation adaptation framework for neural-NETwork-based causal estimation in the presence of unstructured post-treatment surrogates. Starting from pretrained surrogate embeddings, SIRNET learns a shared adapter jointly for outcome prediction, treatment propensity, and arm-specific surrogate marginalization, beyond the generic pretraining objective. The resulting covariate and surrogate representations provide a common interface for these nuisance tasks. The final treatment-effect estimate and confidence interval are obtained from cross-fitted orthogonal scores that attain the semiparametric efficiency bound under the stated nuisance-rate conditions. We also validate SIRNET's performance in synthetic and semisynthetic experiments, as well as the Anti-Amyloid in Asymptomatic Alzheimer's Disease (A4) trial, using frozen pretrained brain-MRI embeddings as surrogates for a long-term outcome.

We give the geometric classification of complex 3-dimensional Novikov-Poisson and Gelfand-Dorfman algebras. In particular, we prove that the variety of complex 3-dimensional Novikov-Poisson algebras is 12-dimensional; it has 4 irreducible components and has no rigid algebras; the variety of complex 3-dimensional Gelfand-Dorfman algebras is 11-dimensional; it has 10 irreducible components and has 3 rigid algebras. As an additional result, we have the geometric classification of complex 3-dimensional Novikov algebras: it is 9-dimensional, and it has $4$ irreducible components with one rigid algebra.

Large AI models increasingly rely on distributed GPU execution, yet it remains unclear whether established anomaly detection methods transfer effectively to this setting. We introduce GF-Bench, a PyTorch-based framework for generating controlled failures in distributed GPU workloads, and a dataset produced with it. The framework supports extensible workload and fault-injection logic; based on this design, the current release targets failures visible in multi-GPU communication logs and provides 220 labeled two-GPU runs covering normal behavior, CUDA out-of-memory fail-stop faults, and persistent-straggler fail-slow faults. Using GF-Bench, we evaluate three sequence-based methods (DeepLog, LogAnomaly, and LogBERT) and the large language model (LLM)-based LogPrompt with in-context learning (ICL) and chain-of-thought (CoT) prompting strategies on parsed logs, and additionally evaluate LogPrompt on raw logs. On parsed inputs, no method reliably detects both failure classes, and the highest overall F1 score is 0.667. Raw inputs raise the overall F1 score of ICL-LogPrompt to 0.983, at 9.0x its average input-token usage and 8.9x its per-run cost. These results reveal a trade-off between effectiveness and efficiency: parsed inputs reduce LLM token use and inference cost but can discard the distributed execution context required for accurate failure detection. GF-Bench provides a reproducible basis for studying cost-efficient anomaly detection methods and can be extended to additional GPU workloads and failure types.

Let $p$ be a complex polynomial of degree $n\ge2$, and let $\sigma_2(p)$ be the root mean square distance of its zeros from their centroid, including multiplicities. We prove that every zero of $p$ lies within distance $\sigma_2(p)$ of a critical point. This proves Borcea's quadratic conjecture and, in particular, gives a new proof of Sendov's conjecture. Two polynomial integration identities prove the result analytically for $n\ge2048$. For the remaining degrees, we derive explicit necessary inequalities, including a variable integration endpoint and an exact centered coefficient identity. Directed integer arithmetic certifies a finite covering of the resulting parameter domains. Degrees $4$–$26$ use $150\,845$ certified three-dimensional regions; degrees $27$–$2047$ use a further $27\,429$ regions, with degree intervals allowed. The cases of degree 2 and 3 were already covered by Khavinson–Pereira–Putinar–Saff–Shimorin.

We study foliations which force every transverse flow to have a closed orbit. We fully characterize this property among $C^2$ finite depth foliations on atoroidal 3-manifolds. We also show that for many foliations, this property can be broken by a $C^0$-small modification of the foliation. We produce closed orbits using the theory of endperiodic maps, and we construct $C^1$ flows without closed orbits by inserting dynamical plugs transverse to foliations along curves of attracting holonomy.

Keith numbers are integers that reappear in sliding-window recurrences generated from their own digits. We introduce Keith mirrors, a subclass defined by saturation at the center of the recurrence's incidence profile. For an even index m, let C_d(m) denote the sum of the two central incidence counts. We prove that the central deficit m-C_d(m) is exactly the distance from m to the interval [2d,2d+2], which forces every Keith mirror of order d>=3 in base B>=3 to return at 2d+2. A digit identity and analytic capacity bounds then reduce the classification to finitely many cases. Exact enumeration, independently checked against the original recurrence and incidence definition, yields precisely 35 mirrors of order at least three, while order two consists of two explicit infinite families. We also classify exactly eight cross-base triple alignments satisfying R=C=x_1+x_d and record the conditional reconstruction property induced by incidence symmetry.

Cascaded recommender systems use the same user history for retrieval and ranking, while the two stages have access to different information at different points of an interaction. We propose the Lifecycle-aware Interaction Factorization Transformer (LIFT), which decomposes each interaction into ordered Request, Item, Context, and Action states and models them as a causal sequence. Retrieval reads the Request state, while ranking reads the Context state, allowing both tasks to share history modeling while preserving stage-specific information. LIFT instantiates this representation with Role-Conditioned Attention and a lightweight Pre-LN Bias. On ML-20M and Taobao, LIFT achieves the highest Joint Score among the evaluated joint models, improving over the strongest baselines by 4.9% and 3.6%, respectively. Loss-weight sweeps show favorable retrieval–ranking trade-offs, while ablations and scaling analyses further examine lifecycle sequence construction, model components, and capacity settings.

Longitudinal medical imaging analysis is a cornerstone of modern medical practice and patient care. Deep learning applied to longitudinal imaging offers wide potential to enhance diagnosis and track disease progression by capturing spatial changes over time. With major advances in single-timepoint deep learning for imaging, there has been growing interest in longitudinal image analysis, given its increased clinical relevance, though technical challenges remain. Several recent innovations may lead to a new era of multi-timepoint image evaluation, yet the scientific landscape, recent progress, and areas of need remain under-characterized. To address this gap, we conducted a scoping review of deep learning methodologies applied to longitudinal medical imaging, yielding 102 studies published between 2018 and 2025. Neurological disorders (48%) and ophthalmic conditions (12%) were the most common clinical applications, with MRI serving as the predominant imaging modality (67%). Sequential feature modeling approaches combining convolutional neural networks (CNNs) with temporal models (LSTM/RNN) were the most frequent methodology (40%), followed by direct feature aggregation across timepoints (23%). Most studies targeted classification tasks (56%), while external validation was performed in only 24% of studies. Our findings highlight that deep learning-based longitudinal imaging analysis remains a promising field, though newer temporal architectures and larger datasets may improve success and clinical adoption of these tools.

Surface braids are a generalization of braids to surfaces in the 4-disk. They are described by planar graphs called charts, and sequences of braid words called braid systems, that represent monodromies of branch points in projections. Yang-Baxter operators (YBOs) have been used extensively in knot theory for constructions of Jones and other polynomials, defined by braid group representations. We propose to use YBOs for constructing surface braid invariants through braid systems. The intersections of eigenspaces are used for this construction. We further use 3-cocycles of the Yang-Baxter cohomology theories via charts, in an analogue of quandle cocycle invariants but more generally utilizing tensor products of modules. An approach towards defining knotted surface invariants through closures of surface braids is discussed.

One of the key unresolved questions in modern cosmology is the Hubble tension, which arises from the inconsistency between determinations of the Hubble constant ($H_0$) based on nearby observations and those predicted by early-Universe data under the standard $\Lambda$CDM paradigm. Recent advances have identified the new Dark Energy Spectroscopic Instrument (DESI) survey as a promising observational tool for probing late-time cosmology. In this paper, we use the latest DESI DR2 baryon acoustic oscillation (BAO) measurements, combined with nonparametrically reconstructed Type Ia supernovae and cosmic chronometer data, to constrain the Hubble constant using three complementary machine learning methodologies: Gaussian process regression (GPR), artificial neural networks (ANNs), and long short-term memory (LSTM) networks. We perform individual and joint constraints on $H_0$ at the effective redshifts of DESI DR2 tracers, yielding $H_0 = 61.8 \pm 5.3\,\mathrm{km\,s^{-1}\,Mpc^{-1}}$ (GPR), $66.3 \pm 5.2\,\mathrm{km\,s^{-1}\,Mpc^{-1}}$ (ANN), and $68.7 \pm 6.0\,\mathrm{km\,s^{-1}\,Mpc^{-1}}$ (LSTM). Our results exhibit excellent agreement with the recent Planck measurements within $1\sigma$, which highlight the potential of data-driven, model-independent methods to address the Hubble tension. Finally, we further test the sensitivity of our results to potential systematics in the DESI data, emphasizing the importance of understanding systematic uncertainties in BAO surveys.

The spread of fake news may cause severe social consequences. Existing fake news detection methods mainly focus on stylistic variations or incorporate external information such as explanations. However, news articles are often rewritten under different emotional backgrounds while preserving their underlying factual claims, which may affect the robustness of detection models. In this work, we investigate fake news detec- tion under fact-preserving emotional variations. To study this problem, we construct emotion-rewritten test sets and generate explanations from the original news articles as stable background knowledge. We then propose a Gated Cross Attention (GCA) framework that adaptively integrates emotionally rewritten news with the corresponding explanations, enabling the model to focus on informative explanation content while reducing potential mismatches caused by emotional reframing. Experiments on PolitiFact, GossipCop, and LUN demonstrate that the proposed method achieves notable improvements under multiple emotional conditions on PolitiFact and LUN, while maintaining competitive performance on GossipCop. We further analyze the effects of explanation guidance and gating mechanisms under different emotional conditions. Our code and data are available at: https://github.com/Flulike/fakenews gca .

We present JoyAI-Voice~2.0, an end-to-end anthropomorphic speech generation model built upon a fully continuous, dual-encoder architecture. Raw speech is encoded into continuous latents and partitioned into patches. Each patch is decomposed by a semantic-acoustic dual encoder into a semantically purified representation and an acoustic representation, which are fused and jointly fed to a causal autoregressive Transformer for planning. The Transformer predicts the conditioning for the next patch, and a local diffusion Transformer renders its full latents for 48\,kHz synthesis. The model is trained with a joint flow-matching and stop-prediction objective, followed by supervised fine-tuning and reinforcement learning with DiffusionNFT to improve model performance. It achieves the lowest average word error rate of 2.51% on Seed-TTS, a 14.9% relative reduction over the strongest baseline, and state-of-the-art attribute fidelity on InstructTTSEval in both Chinese and English, leading on 5 of 10 perceptual dimensions with the highest overall score of 0.893 on MDVD-Eval.

Let $m_T[0,1)$ denote the number of Laplacian eigenvalues of a tree $T$ that are strictly less than $1$. Guo, Xue and Liu \cite{GXL} proved that every tree $T$ of diameter $d$ satisfies $m_T[0,1)\ge \lceil(d+1)/3\rceil$, and that this bound is sharp when $d\equiv2\pmod3$. It is classical that every graph of diameter $d$ has domination number at least $\lceil(d+1)/3\rceil$. In this paper, for any diameter $d$, $m_T[0,1)=\lceil(d+1)/3\rceil$ if and only if $\gamma(T)=\lceil(d+1)/3\rceil$; we characterize all trees satisfying this necessary and sufficient condition, and also prove that almost all trees satisfy $m_T[0,1)\ge\lceil(d+1)/3\rceil+1$.

As GUI-based mobile agents rapidly progress, rigorous safety evaluation of their autonomous decision-making in realistic app interfaces becomes increasingly critical. Existing benchmarks mainly focus on execution-level anomalies using task success or hijack rates, but fail to capture the in-task goal deviation risk in multi-candidate selection tasks, where the decision may be steered toward an attacker-specified target, even in violation of instruction-implied constraints (e.g., cheapest/highest-rated), without any overt execution anomalies. We present DIBench, a decision integrity benchmark for measuring this risk in mobile agents. DIBench covers 7 commercial and 3 simulated apps with 5 task types. Under a threat model restricted to non-privileged UI content, we construct 8 deceptive injection probe instantiations that can steer critical selections without overt anomalies. The benchmark includes 1,000 clean and 36,672 injected instances, with a unified protocol and integrity metrics for comparison. Experiments spanning 4 agent frameworks and 7 base models show that completion-based evaluation can overestimate agent trustworthiness and miss decision-integrity risks: deceptive injections steer selections and shift early action policies, inflating completion rates and creating a misleading illusion of safety. Common defenses, including detection, image preprocessing, and prompt reminders, yield inconsistent integrity gains. Overall, DIBench provides a unified, reproducible benchmark to quantify the risk of in-task goal deviation in mobile agents and enable comparable evaluations of safety defenses.

We revisit the Reissner-Nordström (RN) geometry within the coherent quantum black hole framework, introducing a version that remains well-behaved at short distances while preserving the expected classical behavior at large scales. The key ingredient is a smooth Gaussian regularization, which replaces the usual sharp cutoff and {suppresses} the unphysical oscillations. This construction introduces a characteristic core radius $R_{\rm s}$, constrained relative to the mass $M$ and charge $Q$ to ensure either a {Ricci-regular} center or a softened integrable singularity. Near the origin, the spacetime exhibits distinct behaviors depending on the choice of parameters. When $R_{\rm s}$ satisfies the critical relation $R_{\rm s}^* = \sqrt{\pi}\,R_{\rm Q}^2 / R_{\rm M}$, the geometry is {Ricci-regulFar}, while deviations from this value produce a milder, integrable singularity. {We clarify that for generic parameters, the Kretschmann scalar retains a $1/r$ divergence even at $R_{\rm s}^*$, so the center is not fully regular in the sense of finite curvature invariants.} The effective stress-energy tensor can be interpreted as an anisotropic fluid, whose density and pressure smoothly interpolate between the quantum-dominated core and the classical RN region at large radii. The Misner-Sharp mass accumulates gradually from the center, recovering the ADM mass asymptotically. The quantum scale also modifies the horizon structure, giving rise to standard, extremal, or horizonless configurations depending on the ratios $\alpha=R_{\rm Q}/R_{\rm M}$ and $\beta=R_{\rm s}/R_{\rm M}$. Photon motion is correspondingly affected: the photon sphere moves inward, light deflection and Shapiro delay are slightly reduced, and the resulting shadow appears more compact and smoother, reflecting the quantum-corrected geometry.

The mechanical wave equation, derived from Newton's laws, takes its simplest form in the rest frame of the medium, and appears to lose that form entirely once expressed in the coordinates of another inertial frame — raising the question of whether, and how, it can be written consistently for an arbitrary inertial observer. This puzzle, often glossed over in undergraduate instruction, is typically set aside by invoking the rest frame of the medium as “preferred,” without clarifying what this means for the principle of relativity underlying physical law. We revisit the wave equation from first principles and demonstrate that this apparent non-invariance stems from its expression in Eulerian coordinates. We present two reformulations: a manifestly invariant form using Lagrangian derivatives applied at the level of microscopic particle force laws, and a factorized Eulerian form separating left and right-moving wave solutions. In this context, we explore limitations of Galilean transformation as the low-speed limit of Lorentz transformation, and discuss potential observational connections, particularly regarding wavelength changes in the Doppler effect. Our derivation provides both conceptual clarity and pedagogical value for intermediate and advanced undergraduate physics.

For multivariate polynomial bases of total degree at most n, a column of the Vandermonde matrix is identified by a multi-index of exponents. In this work, we give a direct, recursive map from the scalar column index to this multi-index, based on a prescribed ordering in which exponents of the last variable are increased at the block level and the same rule is applied recursively to the remaining variables. The map is first derived explicitly in the bivariate case, where it yields closed-form column indices for the univariate Vandermonde matrices, and is then extended recursively to arbitrary dimension by means of a binomial counting function, with a proof of correctness by induction. The procedure provides random access to any total-degree column without forming the full tensor-product matrix, and is particularly advantageous for streaming construction and matrix-free evaluation. It is basis independent and applies directly to Chebyshev–Vandermonde matrices. For the full dense matrix, nested loops over the exponent multi-indices achieve the same asymptotic cost; the gain of the present approach lies in exact indexing and the avoidance of superfluous intermediate arrays.

Soap films afford gentle production of air-water interfaces and underlie numerous practical applications. Recently these interfaces have been shown to exhibit unexpected redox capability, wherein Triton X soap films formed with water or HAuCl4 solutions (i) oxidize water up to 1.67 mM H2O2 and (ii) reduce aqueous Au3+ ions into gold nanoparticles, without external energy or catalyst. These reports tend to align with the broader claims of the redox nature of the air-water interface, which has been intensely debated. Here, we re-examine these claims via complementary application of a broad range of techniques, including NMR spectroscopy, fluorescence, colorimetric, TEM, DLS, ICP-OES. Experimental results reveal no evidence for H2O2 formation despite repeated formation of over 300 cycles of soap films, were probed via NMR spectroscopy and other analytical methods. Next, we interrogated soap films for their ability to spontaneously reduce Au3+ ions. TEM and time-resolved DLS revealed that the formation of Au nanoparticles was already formed spontaneous in bulk solutions, and casting soap films (300 cycles) had no measurable effect on particle size or density. This observation was confirmed via complementary quantitative ICP-OES. Our investigations showed that Au nanoparticles formation in bulk HAuCl4 solution without or with triton soap was accompanied by H2O2 generation, and we provide mechanistic insights into these processes. Collectively, these results demonstrate that amplification of the air-water interfacial area through soap-film formation is insufficient, by itself, to induce detectable H2O2 generation or enhance Au-ion reduction. More broadly, our findings highlight the need to distinguish reactions that are driven by purely air-water interfacial effects from those that originate from bulk-solution chemistry and at the solid-water interface

Soft tactile sensors elevate robotic touch through enhanced flexibility and adaptability, yet most existing designs depend on embedded electronics that are susceptible to interference and environmental limitations. In this work, we leverage fluid-solid interactions to develop a class of soft tactile sensors that operate entirely without electronics at the sensing site. The sensor comprises a fluid-filled elastomeric channel connected to only two external pressure sensors. Touching different regions of the elastomeric surface displaces the viscous fluid, producing distinct pressure patterns that encode both touch position and force. These signals are decoded through a machine learning framework that integrates feature extraction, soft clustering, and adaptive neuro-fuzzy inference to achieve accurate localization and force estimation. We validate this concept through single-point touch localization and force estimation in a linear (1D) sensor and extend the same sensing principle to 2D tactile mapping by routing the channel across the surface using space-filling curves, while maintaining the same minimal hardware setup. This simple approach remains effective in environments where conventional electronic sensors often fail, such as underwater or in the presence of magnetic interference.

The rapid growth of deep learning has substantially increased the energy consumption associated with model training, making energy efficiency an increasingly relevant design criterion. This study empirically measures the energy variation of training seven modern computer vision architectures, MobileNetV3-Small, MobileNetV3-Large, EfficientNet-B0, EfficientNet-B1, ViT-B/32, ConvNeXt-Tiny, and ViT-B/16 for the ImageNet-1k classification task, using a homogeneous 10,000-image subset (ImageNet-10k) and a uniform 40-epoch baseline configuration executed on two NVIDIA Tesla P100 GPUs at the Bioinformatics and Computational Biology Center of Colombia (BIOS). Energy was recorded directly via NVML and contrasted with the computational complexity of each model. The results show a Pearson correlation of 0.85 between floating-point operations (GFLOPs) and energy consumption in kWh, indicating that computational complexity is a strong but imperfect predictor of energy expenditure: architectures with comparable GFLOPs exhibited consumption differing by up to 3.1x due to differences in the hardware efficiency of their dominant operations. The EfficientNet variants offered the best balance between classification performance (Val Top-5 up to 97.15%) and energy efficiency (0.457-0.611 kWh), while Vision Transformers exhibited the highest relative energy consumption and lower classification performance under the evaluated configuration. These findings guide architecture selection in energy-constrained computing environments.

We propose a fine-tuning method for flow-matching diffusion models aimed at realistic artificial light modeling without the need for a large training dataset. We address the task of controllable interior image editing, where the goal is to turn artificial light sources on or off while preserving the scene geometry, object placement, materials, and visual identity of the original image. To achieve this, we decompose the task into two independent formulations. We introduce the LoRA Direction Training Method, which extracts the pure direction of the LoRA adapter effect in the diffusion model flow field, and we also introduce specialized loss functions to ensure the realism of the inverse transformation. Additionally, the resulting increment map is used for more precise adjustment of the lighting color and temperature.

A connected graph is called $\ell_1$-embeddable if its shortest-path metric admits an isometric embedding into the $\ell_1$-space; equivalently, in the finite case, it admits a scale embedding into a hypercube. This paper gives a complete parameter-level classification of $\ell_1$-embeddability for regular quadrilateral maps on the torus and the Klein bottle. We treat the two toroidal families $Q_{k,m,r}$ and $Q_{k,m,e}$, obtained from suitable identifications of finite square grids, and, using the known isomorphisms among Klein-bottle quadrilateral maps, the three essential Klein-bottle families $Q_{k,m,a}$, $Q_{k,m,b}$, and $Q_{k,m,h}$. The $\ell_1$-embeddable toroidal maps are exactly $Q_{k,m,0}, Q_{k,2,1}, Q_{k,1,e}, Q_{5,2,e}, Q_{2i,2,e} (i\ge 2)$, and the $\ell_1$-embeddable Klein-bottle maps are exactly $ Q_{2,2,a}, Q_{4,2,a}, Q_{2,m,b}, Q_{1,2,h}.$ All other members of these five families are not $\ell_1$-embeddable.

Multi-agent LLM systems pair a sender with broad context and an executor with a limited local view. We study when a short message improves the executor's next decision, when raw context is preferable, and when a stronger sender helps. Our framework, receiver-relative bounded coordination, expresses message utility as receiver gain minus protocol tax. Compression beats raw context when tax savings exceed losses from omitted information and decoder mismatch. Even Bayes-sufficient compression can fail when a bounded executor cannot use its surface form. A three-stage decomposition separates externalization, absorption, and action closure, explaining how errors remain after the correct content reaches the receiver. Under a single-crossing condition, sender upgrades help above a receiver-burden threshold. Across six benchmarks, the same Qwen protocol raises ContextBench joint accuracy from $0.633$ to $0.775$ but lowers ToolSandbox from $0.889$ to $0.653$. Fixed-message replay reveals closure failures despite correct artifact recovery. These results guide an inference-time selector that improves the accuracy-cost frontier on the evaluated communication regimes.

We examine the Casimir interaction between planar massive Dirac materials coupled to a Lorentz-violating Bumblebee-CFJ vector sector. The standard renormalized one-loop polarization tensor of the Dirac quasiparticles is used as the nonlocal surface response in the Lifshitz scattering formalism. For the normally oriented Lorentz-violating background, the surface response remains parity-even. Hence, the repulsive regimes do not originate from a Hall-like conductivity. Instead, they arise from the CFJ-induced mixing of TE and TM bulk modes, combined with the spatially dispersive response of the Dirac sheets. Furthermore, we show that the Dirac mass suppresses the interaction and changes its large-distance scaling, whereas the CFJ parameter drives transitions between attractive and repulsive regimes.

Hambleton and Kreck described the group of based homotopy self-equivalences of a closed oriented $4$-manifold by a commutative braid of exact sequences built from bordism groups and the Postnikov $2$-type. We construct the analogous braid for closed oriented $BO\langle n+1\rangle$-manifolds $M$ of dimension $2n$ or $2n-1$ whose fundamental group is finite or infinite cyclic and which satisfy one condition on the action of $\pi_1$ on the homotopy of the fibre of $M\to P_n(M)$. We compute $\operatorname{Aut}_\bullet(S^1\times S^{m-1})\cong(\mathbb{Z}/2)^2$ for $m\ge4$, and $\operatorname{Aut}_\bullet(S^3\times S^3)\cong(\mathbb{Z}/12)^2\rtimes SL_2(\mathbb{Z})$ with the standard action modulo $12$. Further, $\operatorname{Aut}_\bullet(S^3\times S^4)\cong(\mathbb{Z}/2\oplus\mathbb{Z}/24)\rtimes(\mathbb{Z}/2)^2$ has order $192$, that $\operatorname{Aut}_\bullet(S^4\times S^4)\cong D_4\times\mathbb{Z}/2$ has order $16$, and that for connected sums of $r\ge2$ copies of $S^3\times S^3$ the kernel contains a subgroup $(\mathbb{Z}/12)^{2r}$ on which the permutations of the summands act non-trivially. For $S^p\times S^p$ the kernel is a sum of two copies of $\pi_{2p}(S^p)$ modulo Whitehead products with $\iota_p$, and the image is determined by an obstruction on the Whitehead square. In particular $\operatorname{Aut}_\bullet(S^5\times S^5)$ is isomorphic to the theta subgroup of $SL_2(\mathbb{Z})$ and $\operatorname{Aut}_\bullet(S^7\times S^7)$ is an extension of $SL_2(\mathbb{Z})$ by $(\mathbb{Z}/120)^2$.

We show that cohomological and homological goodness are equivalent. We prove that if $G$ is a coherent even Artin group or an even Artin group whose underlying graph does not contain 4-clique or a LERF Artin group then $G \otimes G$ and $G \wedge G$ are residually finite. Furthermore for $G$ a RAAG both $G \otimes G$ and $G \wedge G$ are residually $p$-finite for any prime integer $p$. For $G$ a limit group or a Coxeter group we show that $G \wedge G$ is residually finite. For $G$ a soluble group of type $FP_{\infty}$ we prove that the groups $\X(G)$, $\nu(G)$, $G \wedge G$ are virtually residually $p$-finite for almost all primes $p$. For $G$ a generalised Baumslag-Solitar group of rank $n \geq 1$ that is residually finite, we show that $G \wedge G$ is residually finite. For $G = SL_n(\Z)$ and $G = GL_n( \mathbb{Z})$, $n \geq 3$, we show that $G \wedge G$ is residually finite.

Chemical reaction-condition optimization – choosing the catalyst, ligand, solvent, reagent, temperature, time, and atmosphere that jointly maximize yield and stereoselectivity – is a central, judgement-laden subtask of organic methodology research that large language models are increasingly expected to support. Yet existing chemistry benchmarks evaluate reaction-class labelling, retrosynthesis, or SMILES manipulation, and do not ask models to read a real condition-screening table and pick the best set. We introduce RxnOptBench, a benchmark whose every option and precedent is a real wet-lab entry mined from the optimization tables of organic-methodology papers published in 2025, graded by a continuous relative score derived from a declared headline utility that combines reported yield with enantiomeric excess (ee), diastereomeric ratio (dr), and regioisomeric ratio (rr), and equipped with a paired precedents-vs-no-precedents design that isolates in-context use of literature evidence from parametric memorization. Across nine frontier LLMs and three Chemistry LLMs, even the best models leave substantial headroom: chemistry-specialized models fall to the random-baseline floor on multi-axis selection, while open-weight models have closed most of the gap to proprietary frontier models. We release the final human-reviewed benchmark test set and evaluation code.

Mixture-of-Experts (MoE) architectures allow frontier language models to scale to trillions of parameters, but their deployment is constrained by massive memory footprints and memory-bandwidth limitations. Although modern accelerators provide Sparse Tensor Cores (SpTCs) that reduce weight storage and increase throughput through low-precision semi-structured sparsity, exploiting them for MoEs remains challenging because of substantial model-quality degradation and the lack of grouped sparse GEMM primitives. We present an end-to-end hardware-software co-design framework that compresses expert weights into hardware-native, low-precision sparse representations and accelerates their execution on SpTCs. Algorithmically, our framework relaxes discrete semi-structured support selection through continuous reparameterization, enabling differentiable joint optimization with quantized weights under a router-weighted reconstruction objective and scalable expert-parallel compression. Systemically, we develop a custom grouped sparse GEMM kernel tailored to low-precision sparse MoE inference on SpTCs. Across MoE models ranging from 30 billion to one trillion parameters, our framework improves state-of-the-art joint sparse-quantization accuracy by up to 4.35 percentage points while preserving 96.09% of the original model's performance. On NVIDIA B200 GPUs, our kernel outperforms the vendor baseline by up to $1.65\times$, increasing serving throughput by $1.18\times$ and reducing end-to-end latency by up to $4.03\times$. These results establish hardware-software co-design as a practical path toward scalable and efficient MoE deployment.