Machine Learning
Papers
Fresh from arXiv, typeset properly. Highlighted words trade on the market.
All of arXiv
Tuesday 1 September
Looped Transformers increase effective depth by iterating a shared block of layers, but most evaluations compare at fixed model size, conflating architectural advantage with extra FLOPs. We study looping on Mixture-of-Experts Transformers while closely matching per-token FLOPs, total non-embedding parameters, and KV cache. Through a series of ablations, we arrive at a recipe we call SMELT (Sparse MoE Transformer, middle layers Loop Twice), which loops the middle half of layers twice while matching the unlooped Baseline on all three budgets. We scale SMELT across four sizes up to 54B non-embedding parameters and fit a separate Chinchilla-style scaling law for each architecture. SMELT's loss drops faster with compute, saving 6.8–18.0% of training FLOPs on the compute-optimal frontier. The advantage transfers to downstream benchmarks beyond what validation loss predicts, is largest on Code, and grows with sample length and the number of in-context examples. Mechanistic analysis shows that the second visit reduces the attention sink and redirects mass toward content-relevant tokens, an inductive bias that may underlie the observed performance gains. These results show that looping can improve Transformers even under budget matching, offering a practical recipe that turns depth reuse into measurable gains.
High Energy Physics (Phenomenology)
On the Nature of $X(2370)$
Recently, the BESIII Collaboration claimed that their observed $X(2370)$ ($J^{PC}=0^{-+}$) is predominantly a glueball state based on several arguments. However, we find these arguments to be flawed and the conclusion unjustified. Instead, we demonstrate that all observed properties of the $X(2370)$ can be naturally explained by a $\bar\Sigma\Sigma$ molecule dominated state.
Read the full abstract
Computation and Language
Probing Factual Knowledge Transfer with Training Data Interventions
Do multilingual language models transfer factual knowledge across languages during continued pretraining, or do they mostly recall facts learned directly from the target-language data? To answer this question more reliably, we propose an intervention-based framework: starting from an English-pretrained model, we continue pretraining on Persian data from which specific facts have been systematically removed at varying levels of granularity. We construct SIFT, a resource of 500 triples across 20 relations, stratified by the cultural origin of each fact's subject into general (globally prominent) and Persian-related entities, designed for both systematic fact removal from training data and evaluation, with natively written Persian cloze templates. Our results show that fact transfer is very limited: under the strictest removal condition, a large majority of English-acquired facts fail to transfer into Persian. We further show that sentence-level co-occurrence removal is insufficient to eliminate fact signal, and that easier (randomly selected) negative candidate sets substantially inflate apparent transfer by rewarding shallow associative heuristics, while performance on a harder candidate set that allows for less reliance on heuristics is much lower. Finally, we show that source-language entity frequency has a large influence, with Persian-related facts, which are orders of magnitude rarer in the English corpus, hardly transferring.
Read the full abstract
Quantum Physics
Quantum phase transitions of the Cavity Heisenberg spin-chain
The interplay between quantum criticality and ergodicity breaking constitutes a central challenge in complex quantum many-body systems. Here, we investigate ground-state quantum phase transitions (QPTs) and excited-state quantum phase transitions (ESQPTs), as well as ergodic-nonergodic transition in the cavity Heisenberg spin-chain (CHS) model. By combining quantum information measures with semiclassical fixed-point analysis, we identify a deformed phase that interpolates between the normal and superradiant phases, featuring a logarithmic nonanalyticity in the density of states (DoS) and two additional jump discontinuities. We further elucidate the spectral-structure mechanism underlying the connection between ESQPTs and the ergodic–nonergodic transition (ENET) via level statistics, participation ratios, and multifractal analysis. Our results provide a general framework for characterizing phase structures and spectral features in light–matter quantum many-body systems.
Read the full abstract
Robotics
Integrating Traffic Noise Emission Modelling into Variable Speed Limit Control
Road traffic noise remains a major environmental challenge, yet most speed management strategies are static and do not respond to short-term variations in traffic noise emissions. Although variable speed limit (VSL) systems are widely deployed for safety and congestion mitigation, traffic noise is rarely treated as an explicit operational control objective. This paper proposes a noise-aware VSL framework that integrates aggregated traffic-state estimation with a simplified CNOSSOS-EU-based emission indicator. A stage-based controller with time-varying reference thresholds dynamically adjusts discrete speed-limit levels in response to estimated emission conditions. The framework is evaluated using microscopic traffic simulation calibrated with empirical motorway data and replicated across multiple stochastic realisations. Over a 24-hour evaluation period, the adaptive strategy reduces the receiver-based equivalent sound level by 2.9 dB(A) relative to unrestricted traffic conditions, while maintaining an average vehicle speed approximately 11.3 km/h higher than a permanently imposed low-speed regime. Period-wise analysis shows that speed reductions are activated selectively when emission levels approach calibrated targets, rather than enforcing a constant intermediate limit. Traffic stability indicators reveal moderate increases in speed variability compared with unrestricted operation, but substantially lower braking intensity than under uniform low-speed enforcement. These results demonstrate the feasibility of integrating environmental performance indicators into operational speed control, providing a practical complement to conventional infrastructure-based noise mitigation measures.
Read the full abstract
Operating Systems
mzCache: On-Device LLM Memory Management under Multitasking
On-device mobile Large Language Model (LLM) inference is gaining significant attention. However, mobile devices operate in highly dynamic multitasking environments where users frequently switch between applications. This creates memory pressure, forcing LLM memory (model weights and KV cache) to be evicted by the operating system. When a new inference request arrives, the inference system must restore the evicted memory through slow storage reads or recompute the entire KV cache, severely degrading responsiveness. To address this, we present mzCache, an on-device LLM inference system with specialized memory management for multitasking environments. Under unpredictable memory pressure, mzCache elastically evicts LLM memory and leverages the unified memory of mobile SoCs to enable zero-wait inference on the GPU with concurrent CPU-side restoration. mzCache realizes this through restoration-oriented memory management: LLM memory is partitioned into fine-grained shared buffers to enable partial eviction and restoration with concurrent cross-processor access, while hybrid swap and backward-out eviction policies ensure low-latency restoration from any eviction state. Implemented on llama.cpp and deployed as an Android application, mzCache achieves 2.1-5.5$\times$ reduction in Time-to-First-Token compared to storage-backed partial offload and demonstrates its effectiveness in real multitasking scenarios.
Read the full abstract
Artificial Intelligence
LEAP: Likelihood Elicitation and Aggregation for LLM-based Probabilistic Forecasting
LLM-based forecasting systems have improved on real-world tasks such as financial markets and sports outcomes, largely through stronger search and tool use. Many systems still ask an LLM to read all collected evidence together and produce the final forecast. We call this design Monolithic Prediction. It can obscure how individual evidence items affect the result and collapse uncertainty across competing outcomes. We propose LEAP (Likelihood Elicitation and Aggregation for Probabilistic forecasting), which reorganizes how collected evidence is used in the prediction stage. LEAP examines each evidence item separately and elicits likelihood parameters that describe its implications for the target. An explicit prior and a deterministic probabilistic model then combine these likelihoods into a posterior distribution. This procedure supports continuous, single-choice, and multi-choice forecasts while preserving reproducible evidence contributions. We build a benchmark covering forecasting, information-seeking, and browsing tasks, and evaluate LEAP on our own agent loop and several agent CLI frameworks. Given the same evidence, LEAP improves most prediction and calibration metrics across models and remains stronger under controlled comparisons of prior access, inference budget, and aggregation.
Read the full abstract
High Energy Physics (Experiment)
Validating a BDT-based Electron-Positron Identification Algorithm at CLAS12 with Experimental Data
This article presents a machine-learning, particle-identification algorithm for electrons and positrons in the CLAS12 experiment. The main objective was to minimize charged-pion contamination for both experimental and simulated data. We developed, evaluated and validated two BDT models with different input features, both trained on simulated samples, and conducted rigorous validation using both simulated and experimental data to ensure reliability. Our results show the effectiveness of the applied models in mitigating background contamination. By retaining more than 90% of leptons in the simulated samples, the charged-pion background is largely reduced. Performance was evaluated on experimental data, providing a framework to validate such approaches at CLAS12 and future electron-beam facilities.
Read the full abstract
Machine Learning
Bandits in Prod: Hyperparameter Optimization at Inference Time
Many production systems can assess a configuration only by using it on live requests and observing noisy feedback. Modern agentic systems are a prominent example, with inference-time choices such as model selection, retrieval depth, prompting strategy, and decoding temperature, yet often with no representative validation data. We formalize this setting as Online Hyperparameter Optimization (OHPO) and cast it as an infinitely many-armed bandit over mixed and conditional search spaces. We introduce IMABO, a general framework that combines any bandit policy for choosing among already sampled configurations with any oracle for proposing new ones. We instantiate it with IMOSS, a restart-free anytime policy whose active set grows as $t^{\beta}$, and prove an expected cumulative quantile-regret bound of $O(p_\rho^{-1/\beta} + T^{(1+\beta)/2})$, where $\beta\in(0,1)$ controls active-set growth and $p_\rho$ lower-bounds the probability that a proposed configuration falls in the top-$\rho$ fraction of the search space. We combine IMOSS with three practical oracles: a Tree-structured Parzen Estimator, an incumbent-mutation oracle driven by a per-coordinate bandit, and a pretrained tabular foundation model, all three improving over the uniform random oracle baseline. IMABO outperforms all baselines in terms of regret across diverse OHPO settings, from tuning classical machine-learning models to configuring LLM-based agents. Our implementation is available at https://github.com/Tiime-Software/IMABO.
Read the full abstract
Quantum Physics
Fast Microwave-free State Preparation and Measurement of Superconducting Qubits
Fast, high-fidelity, scalable state preparation and measurement is critical to the realization of a quantum computing system. The state-of-the-art methods for preparation and readout of superconducting qubits require finely tuned microwave signals and ~100 ns of measurement time, which are major obstacles to the scalability and performance of superconducting quantum computers. Here, we have demonstrated novel, microwave-free methods for both preparation and readout of superconducting qubits with >99% fidelity in only 10 ns for either operation while maintaining qubit coherence. This technology is compatible with scalable superconducting digital control systems, and using quantum flux parametrons for amplification, we demonstrated full quantum-to-digital conversion in only 15 ns, which is an order of magnitude faster than state-of-the-art microwave-based techniques.
Read the full abstract
Analysis of PDEs
Delaunay-type interface in a screened model of diblock copolymer melts
A diblock copolymer is a soft-matter composed of two chemically distinct block of repeating monomers covalently bonded together at an end-to-end junction to form a single polymer chain. In this paper, we establish the existence of infinitely many smooth periodic unbounded domain patterns of Delaunay-type in $\mathbb{R}^3$ that optimize the energy distribution in diblock copolymer melts. We emphasize that pattern domains at the equilibrium correspond to stationary sets of the screened Ohta–Kawasaki free energy functional \begin{align*} \mathcal{P}_\gamma(\Omega) := |\partial\Omega| + \gamma \int_{\Omega}\int_{\Omega} G_{\kappa}(|x-y|) \,\mathrm{d}x\mathrm{d}y, \end{align*} where $\gamma>0$, $\kappa>0$ and $G_\kappa(r)=\frac{1}{r} e^{-\kappa r}$ is the repulisive Yukawa potential. Equivalently, these equilibria satisfy the corresponding Euler–Lagrange equation \begin{align*} \mathcal{H}_\Omega (x):= H_{\partial\Omega}(x) + \gamma \int_{\Omega} G_{\kappa}(|x-y|) \mathrm{d}y = Const \quad \text{on } \partial\Omega, \end{align*} where $H_{\partial\Omega}$ denotes the mean curvature of the surface $\partial\Omega$. By analyzing the linearization of $\Omega \mapsto \mathcal{H}_\Omega$ around flat cylinders and applying the Crandall–Rabinowitz bifurcation theorem, for any $\kappa > 0$ and sufficiently small $\gamma > 0$, we prove the existence of non-trivial, $2\pi$-periodic Delaunay-type equilibrium cylinder interfaces with shapes close to a Delaunay unduloid surface of constant mean curvature.
Read the full abstract
High Energy Physics (Theory)
Perturbative Double-Scalar Hair on Planar Black Holes in AdS$_{4}$-Einstein-Scalar Gravity
We construct a two-scalar hairy planar black hole in four-dimensional Einstein gravity coupled to two scalars whose potential descends from the ABJM consistent truncation of eleven-dimensional supergravity. Starting from the known single-scalar hairy black brane, we switch on the second scalar $\phim$ perturbatively in its dimensionless hair parameter $\eps$, which at fixed source is proportional to $\lambda_{-}T$, and solve the coupled Einstein-scalar system to order $\eps^{2}$. The linearized $\phim$ profile is an exact hypergeometric function, independent of the $\pp$ hair, whose horizon-regular branch always carries a source, so there is no spontaneous $\phim$ hair. Because the dual operator is irrelevant ($\Delta_{-}=4$), the second-order response develops a resonant logarithm: the condensate runs with a closed-form coefficient fixed by the cubic vertex. The metric and $\pp$ back-reaction is obtained in closed form; for arbitrary hair it diagonalizes into two Pöschl-Teller problems and locks the $\pp$ correction to the transverse metric correction with ratio $6/\sqrt7$. On AdS-Schwarzschild the first law and holographic renormalization agree and give an analytic equation of state, $T/T_{0}-1=-\tfrac{5\sqrt3}{8\pi}\eps^{2}$ at fixed entropy and $c_{s}^{2}=\tfrac12-\tfrac{5\sqrt3}{8\pi}\eps^{2}$, with $\eta/s=1/4\pi$ and $\zeta/\eta=\tfrac14\eps^{2}$. We then add the $\mathbb{Z}_{2}$-odd third scalar $\chi$ of the truncation, dual to $\Delta_{\chi}=5$. On the hairy background its linear equation is a Heun equation with an apparent singularity, solved by a Clausen function. At second order $\chi$ necessarily sources $\phim$, yet neither cross-coupling makes either condensate run, a mutual non-renormalization that follows from a one-step resonance identity and the relation $2\Delta_{\chi}-\Delta_{-}=2d$.
Read the full abstract
General Relativity and Quantum Cosmology
Gravitational-wave signatures of primordial black hole clusters and the imprint of an early dense-core collapse
Primordial black holes (PBHs) are expected to be born essentially non-spinning and, if the primordial curvature fluctuations possess non-Gaussian tails, strongly clustered. Any spin measured in a PBH population must therefore be dynamically generated. Using direct $N$-body simulations with nbody6++gpu, modified to treat nearly parabolic dynamical-capture mergers, we quantify the conditions under which an initially non-spinning PBH cluster develops the spin, mass and stochastic-background signatures that gravitational-wave observations are now sensitive to. We find a sharp tension: producing a root-mean-square spin $\chi_{\rm rms}\sim0.1 - 0.2$ in the $30 - 50\,M_\odot$ range requires initial Plummer radii of order $10^{-5}\,{\rm pc}$, whereas survival over cosmological times requires parsec-scale clusters. This tension is resolved naturally if PBHs form with a compact, short-lived core embedded in an extended halo, and we therefore simulate the two regimes as independent blocks. The core block ($4000$ PBHs, Plummer radii $3\times10^{-6} - 10^{-4}\,{\rm pc}$) collapses in days to years, producing a burst of mergers whose orbits are strongly non-circular ($e\to1$) and of low dimensionless orbital angular momentum, so that remnant spins are systematically below the quasi-circular value $\chi\simeq0.7$. Redshifted from their formation epoch, the core mergers deposit a stochastic background peaking in the $10^{-6}$–$10^{-4}\,{\rm Hz}$ band accessible to LISA. The picture predicts a specific division of labour between observables, spins set in the primordial collapse, detectable mergers set by the long-term halo, and a two-component stochastic background, and, in particular, that both the spin and the mass data favour a primordial mass function truncated near $50\,M_\odot$.
Read the full abstract
Computational Engineering
Connectome-Based Modelling Reveals Orientation Maps in the Drosophila Optic Lobe
The ability to extract oriented edges from visual input is a core computation across animal vision systems. Orientation maps, long associated with the layered architecture of the mammalian visual cortex, systematically organise neurons by their preferred edge orientation. Despite lacking cortical structures, the Drosophila melanogaster brain contains feature-selective neurons and exhibits complex visual detection capacity, raising the question of whether map-like vision representations can emerge without cortical infrastructure. We integrate a complete fruit fly brain connectome with biologically grounded spiking neuron models to simulate neuroprocessing in the fly visual system. By driving the network with oriented stimuli and analysing downstream responses, we show that coherent orientation maps can emerge from purely connectome-constrained dynamics. These results suggest that species of independent origin could evolve similar visual structures.
Read the full abstract
Statistical Mechanics
Pole-Zero Geometry, Model Reduction, and Identifiability in Sensory Adaptation
Sensory adaptation provides a concrete setting in which low-order system identification can fail qualitatively. We show that one fixed higher-order adaptive system composed entirely of real first-order relaxation modes can be reduced to opposite sides of the second-order pole boundary: low-frequency moment matching gives $\rho_{\rm moment}=4.50$, whereas finite-window fitting gives $\rho_{\rm window}=3.31$, and the inferred pole class changes further with sampling protocol. Thus the real-versus-complex classification of a reduced model is not itself reduction invariant. We then use the general two-state spectrum to connect stochastic identifiability to adaptation: for nontrivial coupling and one-state observation, cross diffusion drops out of the scalar spectrum when the hidden state has no self-relaxation. In the adaptive model, this condition is precisely the integral-memory limit that produces exact adaptation, while leaky memory restores spectral sensitivity. For the exact-adaptation model, the Gaussian path-space irreversibility nevertheless depends on the hidden cross-diffusion channel. Hence $\{H,S_x\}$ does not determine the irreversibility rate. Independently, for a specified all-even reduced two-state drift with $\rho<4$, the drift-only lower bound is $\sigma \ge \tau_x^{-1}(4/\rho-1)$. Published E.~coli and C.~elegans responses provide biological examples of these limits. The distinction established here between transfer-function invariants, reduction-dependent properties, and hidden-state quantities provides a concrete framework for evaluating the limitations of low-dimensional models of adaptive biological dynamics.
Read the full abstract
Fluid Dynamics
Multiphase Reactive Transport in a Heterogeneous Flow Field: Channel Formation in Hydrocarbon-Bearing Carbonate Rock
Carbon capture, utilization and storage (CCUS) is a key approach for reducing anthropogenic CO2 emissions. Currently the vast majority of CO2 stored is injected into depleted hydrocarbon reservoirs. Reactive transport in hydrocarbon-bearing carbonate reservoirs is controlled not only by the balance between reactant delivery and surface reaction, but also by the heterogeneous pore-scale flow field created by pore structure and fluid distribution. Here, we used time-resolved micro-CT imaging, pore-network analysis, and direct numerical simulation to investigate channelized dissolution during injection of CO2-saturated brine into oil-bearing Ketton limestone at 0.5 mL/min. The system remained in a high Pe, low Da regime throughout the 180 min injection. Despite the advection-dominated flow regime, dissolution became strongly localized. Preferential flow pathways were already present originally because the heterogeneous pore structure and remaining-oil occupancy restricted brine flow to a subset of the connected pore space. Dissolution progressively amplified these pathways. The associated reduction in hydraulic resistance further focused flow, producing a continuous tortuous channel by 180 min. Meanwhile, the effective reaction rate was only 1.6*10^5 mol/m2s,, approximately one order of magnitude lower than the batch reaction rate, showing that rapid advective transport did not translate directly into rapid overall dissolution under multiphase conditions. These results demonstrate that channel formation arose from the coupling between strong reactant delivery and pore-scale flow-field heterogeneity, rather than from the bulk Pe Da regime alone. Accounting for flow-field heterogeneity is therefore important for predicting reactive transport and pore-structure evolution during CO2 injection into hydrocarbon-bearing carbonate reservoirs.
Read the full abstract
Number Theory
An improvement on the largest prime factor of $n^2+1$
The study of the largest prime factor in polynomial sequences can be traced back at least to the late 19th century in the work of Störmer. Mahler (1933) and Chowla (1934) proved that the largest prime factor of $n^2+1$ grows at least as fast as $\log_2 n$. In 2023 we improved this to $(\log_2 n)^2/\log_3 n$. In this note we show the lower bound $(\log_2 n)^2/\log_4 n$, and that when this bound is nearly sharp it also holds for many prime factors of $n^2+1$.
Read the full abstract
Cryptography and Security
Hidden Services Protocol for Mixnets
Mix networks (mixnets) provide network-level privacy by routing each communication packet through a sequence of intermediaries, called mixnodes, that randomly delay and cryptographically transform packets before forwarding them, making it difficult for observers to link mixnet entries to exits. While this mechanism protects sender privacy from both external adversaries and the receiver, existing mixnets lack a secure and practical protocol that simultaneously protects receiver (destination) privacy, particularly from the sender. We close this gap. We introduce the first practical and secure hidden-service protocol for mixnets, providing receiver privacy alongside sender anonymity. Our design builds on Single-Use Reply Blocks (SURBs), which enable anonymous replies without revealing the receiver's address. We show, however, that existing approaches to using SURBs expose two practical attacks that can compromise sender or receiver anonymity when the opposing party controls only a single mixnode. We develop defenses against both vulnerabilities. Building on these defenses, we introduce NymHS, a secure and practical hidden-service protocol for mixnets that supports anonymous service discovery, authenticated bidirectional sessions, and asynchronous SURB replenishment. We implement NymHS on the open-source Nym codebase and evaluate its practicality through web-browsing experiments across 118 websites, measuring each of the 27 configurations three times (9,558 page loads). Our results demonstrate that hidden services can be deployed efficiently over mixnets. In particular, increasing the Sphinx payload from 2 KiB to 10 KiB reduces mean page-load latency by approximately 5.2x and decreases communication overhead from 21.7% to 4.3% relative to the current Nym baseline.
Read the full abstract
Computation and Language
VerTox: Verifiable Reward-Guided Corpus Poisoning Against Neural Ranking Models
Neural ranking models have become core components of modern information retrieval systems and important building blocks of AI systems such as retrieval-augmented generation (RAG) pipelines. However, their robustness remains insufficiently understood in the presence of large language models (LLMs), which can generate fluent and deceptive content at scale. This work investigates the vulnerability of neural ranking models to corpus poisoning attacks, in which an adversary injects a small number of maliciously crafted documents into the corpus to distort ranking behavior. We propose VerTox, the first framework to formulate corpus poisoning as a verifiable reward-guided reinforcement learning (RLVR) problem. By explicitly coupling ranking distortion with factual corruption through specialized reward shaping, we fine-tune compact LLMs into adversarial generators. Experiments demonstrate that our method achieves near-perfect attack success rates, producing adversarial documents that frequently rank higher than target documents across major neural ranking architectures, as well as a proprietary commercial embedding model. The generated adversarial documents are fluent and exhibit low perplexity, making them difficult to detect. Furthermore, by explicitly encouraging factual corruption, our adversarial documents significantly degrade the performance of a downstream RAG application.
Read the full abstract
Differential Geometry
Mass Bounds for Confined Area-Minimizing Minimal Surfaces
We establish codimension-independent mass bounds for geometrically confined area-minimizing rectifiable currents. In Euclidean space, we combine the confined-volume doubling theorem of Colding–Minicozzi with a current-theoretic squashing argument. This gives an affirmative answer to Lin's interior mass-bound problem for every algebraic projection multiplicity $Q$: the interior mass is bounded by $C(n)Q$, without an a priori mass bound at a larger scale. This result yields a degree-one Bernstein-type rigidity result under sublinear confinement. For the hyperbolic application, we make the curvature modification of the fixed-scale argument needed in a thin tubular neighborhood of a totally geodesic copy of $\mathbb{H}^n$. Combining this auxiliary estimate with a localized squashing estimate removes the doubly exponential local mass-growth condition from the boundary regularity results in [13].
Read the full abstract
Computational Finance
Adaptive singular-point method for pricing and hedging surrenderable equity-linked contracts
We propose a deterministic numerical method for pricing and hedging surrenderable equity-linked life-insurance contracts with periodic premiums and fund contributions, maturity and death guarantees, and Bermudan surrender under correlated stochastic volatility and stochastic interest rates. The main computational challenge is the non-recombining accumulated fund, which couples with multiple stochastic factors and early exercise. Our key idea is to avoid a full multidimensional fund lattice: variance and interest-rate factors are discretized on recombining lattices, while at each factor node the contract value is represented as an adaptive one-dimensional function of the fund. Periodic contributions then act as translations of the fund argument, whereas surrender is handled directly through a backward obstacle condition. Piecewise-cubic representations propagate payoff and exercise singularities and are compressed by continuous pruning criteria that control the representation error. We establish weak convergence of the financial chains, convergence of the adaptive valuation under vanishing representation error, and Delta consistency on regular fund regions. For the strict binomial scheme, additional regularity yields first-order weak accuracy and a Talay-Tubaro expansion supporting Richardson extrapolation. Numerical experiments show compact representations, favorable cost–accuracy, and close agreement with independent Monte Carlo and cross-fitted least-squares Monte Carlo benchmarks. Hedging results further show that contracts with similar values can generate materially different exposures to equity, volatility, and interest-rate risk.
Read the full abstract
Computation and Language
Exploring Sparse Autoencoders in Text-Based Causal Confounding Adjustment
In many settings, studying causal questions based on text data requires adjusting for confounding information within texts. Yet there is a tradeoff in constructing text representations for adjustment: they must be sufficiently large and/or dense to preserve the confounding variables necessary for unbiased effect estimation, but sufficiently small and/or sparse to satisfy finite-sample overlap and yield low-variance estimates. To address this tradeoff, we turn to sparse autoencoders (SAEs), and propose a novel causal adjustment pipeline that iteratively selects a minimal set of SAE features via conditional independence tests. We find that SAE representations achieve better adjustments (lower bias and and higher coverage) than alternative representations in standard semi-synthetic evaluations with binary confounders, and their interpretability offers opportunities for falsification. We also introduce a more realistic semi-synthetic evaluation that uses multi-label data as the unobserved confounders and find off-the-shelf adjustment methods require increased investigation for these more complex settings. Code: https://github.com/mianzg/sae-text-confounder
Read the full abstract
Representation Theory
Total positivity in coordinate algebras
Artificial Intelligence
Automated Event Log Generation from Unstructured Text Using Finetuned LLMs
Process mining (PM) provides a powerful framework for discovering and optimizing operational processes from event data. However, the efficacy of PM techniques is strictly predicated on the availability of structured event logs. Thus far, event logs have often been laboriously created by domain and process mining experts. This costly effort causes large portions of organizational knowledge, including incident tickets, manuals, and textual reports, to remain underutilized. We address this bottleneck by investigating the efficacy of Large Language Models (LLMs) as automated data translators. We propose a scalable framework that leverages LLMs as data translators to bridge the gap between unstructured textual resources and structured event data. We finetune LLMs on a newly created text-to-log dataset, demonstrating that the resulting models can extract high-fidelity event logs from unstructured resources. Our results show that this finetuning approach outperforms few-shot or zero-shot prompting by a large amount, highlighting finetuning as a necessary pre-condition for generating reliable event data. We conclude that our method provides a promising pipeline for making previously unused data available to process mining ecosystems, effectively expanding the possibilities of using PM to further investigate organizational workflows.
Read the full abstract
Machine Learning (Statistics)
Matched Queries for Curvature and Density at Branching Junctions
At a junction, a score field can reveal weighted tangent rays, yet these first-order quantities do not determine how individual branches bend or how their densities change away from the center. Recovering this missing information is necessary for describing local continuation beyond a single point, but finite observations must separate branchwise second-order effects while allowing error in the estimated center. We address this inverse problem using matched score queries at noise scales $\sigma$ and $\lambda\sigma$. For a finite union of $C^{2,\alpha}$ half-branches in $\mathbb{R}^D$, the normalized score has the expansion $F_\sigma=F_0+\sigma G+O(\sigma^{1+\alpha})$. Matched subtraction cancels the tangent contribution and exposes $G$, which depends linearly on branchwise curvature and log-density slope. Given tangent directions and weights on distinct rays, $G$ uniquely identifies all $sD$ branch parameters, and $sD$ scalar component observations are necessary. An $O(\sigma^2)$ center error introduces $D$ translation modes, leading to $(s+1)D$ observations under full-rank calibration, except for a translation-invariant full line. We also establish a perturbation bound and a conditional kernel-density-estimation rate. Experiments reproduce the predicted population and $N^{-1/5}$ trends and remain full rank up to $D=20$ with 16 supplied branches. In end-to-end tests for $D=3$–$5$, a known-count first-order frontend yields full rank in all 135 population systems and a median relative jet error of 0.132. With strong first-order error, matched responses reduce median parameter error by a factor of 49.4 relative to naive tangent subtraction.
Read the full abstract
Computer Vision
Reliability Challenges in Diffusion Vision-Language Models
Diffusion-based Large Vision-Language Models (dLVLMs) have recently emerged as a compelling alternative to autoregressive (AR) LVLMs, offering advantages in parallel decoding, bidirectional context, and controllable generation. Despite rapid progress, their reliability properties remain largely uncharacterized. We present the first systematic reliability evaluation of hallucination and bias in dLVLMs, benchmarking six diffusion models against competitive AR baselines across four dimensions. Our key findings are: (1) dLVLMs reverse the yes-bias of AR models in binary visual queries; (2) they achieve competitive hallucination rates yet exhibit degraded linguistic quality; (3) they collapse to near-zero accuracy on underrepresented racial groups with opposite-polarity gender bias; and (4) they exhibit accuracy collapse in multiple-choice settings when the correct option is shorter than its distractors, associated with a length prior that emerges at the first denoising step. Tokens committed at late denoising steps with low confidence further correlate with hallucinated content, pointing to a mechanistic signal unique to diffusion generation. These patterns vary across model families, suggesting reliability is shaped by the generative paradigm together with training data.
Read the full abstract
Optimization and Control
Consensus-based optimization for linearly separable functions
Consensus-based optimization (CBO) is an efficient metaheuristic for global optimisation with attractive mathematical properties, allowing global convergence results even in non-convex settings. In practice it suffers greatly from the curse of dimensionality, as do most particle-based optimisers. Different strategies have been proposed to apply CBO even for high-dimensional optimisation problems, the most prominent being the so-called anisotropic noise model. However, a recent work by Bonandin et al. highlighted that this method performs well primarily on separable objective functions, and even small coordinate rotations severely degrade the performance. Motivated by this observation, in this work we study the case where the objective function is separable only in a linearly transformed coordinate system. To leverage the efficiency of anisotropic CBO we propose a reparametrisation scheme, which provably estimates such a coordinate transformation and then applies CBO in the new variables. Furthermore, we show how our scheme can be interpreted as noise adaptation in CBO. Numerical examples highlight the efficacy of the method, demonstrating significant performance improvements on challenging benchmark objectives.
Read the full abstract
Information Retrieval
MIDR: Enrichment-Augmented Indexing for Multimodal Document Retrieval
Retrieval over visually rich documents has a representation problem: important content often lives in tables, charts, figures, and layout relations that plain OCR linearizes, corrupts, or omits. ColPali-family visual retrievers address this with patch-level multi-vector indexes and late-interaction scoring, keeping image-derived retrieval on the query-time serving path. We introduce MIDR (Multimodal Indexing for Document Retrieval), a training-free framework for enrichment-augmented indexing that shifts multimodal reasoning to index time. During ingestion, a multimodal LLM converts rendered pages into verified textual fields that are indexed with BM25F and optionally fused with dense retrieval, enabling text-centric serving over multimodally grounded evidence. On ViDoRe V3, MIDR Hybrid achieves 0.6219 average nDCG across five English domains, a 23.0% relative gain over BM25, remaining competitive with ColQwen2.5. On two French-document domains, enrichment bridges English queries and French page text, lifting BM25 from 0.1532 to 0.5448 nDCG and outperforming ColQwen2.5. Across all seven domains, MIDR leads ColQwen2.5 on four while using approximately 9x smaller index memory and approximately 2x lower query latency. These results establish index-time multimodal reasoning as a compelling accuracy-deployment alternative to serving-time visual late interaction.
Read the full abstract
Artificial Intelligence
A Composable Evaluation System for Reproducible Omni-Modal Foundation Model Evaluation
Building an omni-modal foundation model means evaluating it across text, image, video, and audio. Excellent evaluation toolkits exist for each modality, but their inference engines, prompt conventions, and metric implementations are mutually incompatible, so practitioners end up maintaining separate environments for every toolchain and still struggle to compare results across them. OmniEvaluator grew out of this need in our own model development: rather than reimplementing benchmarks, it connects existing inference engines and curated evaluation libraries at a higher level, exposing four inference backends, four evaluation frameworks, and over a thousand benchmarks through a single interface. Every run is recorded as an artifact capturing the full configuration for exact reproduction, and results flow into a shared dashboard for cross-model comparison. A federated mode shares GPU inference servers across concurrent evaluations, and a built-in verifier, small enough to run on CPU, keeps its score stable across engines and prompts where rule-based scoring fluctuates under configuration mismatch, matching cost-efficient commercial LLM judges without their recurring API cost. The system, demo video, and dashboard are publicly available. (https://github.com/naver-ai/omni-evaluator)
Read the full abstract
Astrophysics of Galaxies
Collisional excitation of cyclopentadiene by helium. A complete set of rate coefficients and astrophysical applications
Complex organic molecules, including large cyclic species, are prevalent in interstellar space and play a key role in various astrochemical processes. Cyclopentadiene (c-C5H6) is a five-membered cyclic hydrocarbon recently detected in TMC-1 and some other interstellar molecular clouds. While accurate spectroscopic data were available, collisional rate coefficients for its rotational transitions were missing so far, introducing a potential limitation in the interpretation of the observations. This study aims to provide a comprehensive set of state-to-state thermal rate coefficients for the rotational excitation of c-C5H6 due to collisions with helium, crucial for non-local thermodynamic equilibrium (non-LTE) radiative transfer modelling in astrophysical environments, and to examine how far the molecule is from thermalisation under the physical conditions of cold molecular clouds. The research employed accurate quantum scattering calculations using the close-coupling (CC) and coupled states (CS) methods, based on a highly-correlated three-dimensional potential energy surface for the [c-C5H6 + He] collisional complex. Calculations were performed for a wide range of rotational states (from j = 0 to j \leq 25) and kinetic temperatures (from 10 to 50 K). We calculated a complete set of thermal rate coefficients for both ortho- and para-C5H6. Radiative transfer simulations demonstrated that most rotational levels of cyclopentadiene are fully thermalized under typical cold cloud conditions and exhibit minor non-LTE effects. Nevertheless, this study is the first to utilise accurate state-to-state rate coefficients for radiative transfer simulation of a large, five-membered cyclic species detected in space. This allows to draw some general conclusions and paves the way for future studies of complex astromolecules that will enable a more precise interpretation of upcoming detections.