Machine Learning
Papers
Fresh from arXiv, typeset properly. Highlighted words trade on the market.
All of arXiv
Wednesday 9 September
Fine-tuning large language models (LLMs) on domain-specific data is essential for downstream adaptation. In many deployments, a participant cannot hold the complete model locally. This happens because the model owner keeps the full model proprietary, or because the participant lacks sufficient compute resources. Split Learning (SL) addresses this by partitioning the model between the participant and a server so that only a small portion runs locally. When the underlying data is additionally distributed across multiple institutions with privacy requirements, Federated Learning (FL) further enables collaborative training across participants by sharing only model updates instead of raw data. In this combined setting, each client transmits intermediate activations to the server, and for LLM fine-tuning, this exchange poses an inherent privacy paradox. The autoregressive nature of LLMs causes the transmitted activations to leak the input, and existing perturbation-based defenses are fundamentally ineffective in this setting. We address this leakage through a learned obfuscate-and-recover scheme that protects participants' private datasets while still allowing an independently deployable model to be trained on the server side. Experiments demonstrate that our approach achieves strong privacy protection with modest utility loss and system overhead, making split-based federated LLM fine-tuning practically viable.
Cryptography and Security
How Fragile Is Safety Alignment at Frontier Scale? A Single-Direction Attack on a 320B MoE
Directional ablation removes an aligned language model's ability to refuse by projecting a single "refusal direction" out of the weights that write the residual stream. It needs no gradient-based training and no optimization, only a few hundred contrastive prompts, which makes it the canonical white-box attack on open-weight alignment. However, it has been established only on dense models up to roughly 70B parameters. We study whether it survives the shift to frontier mixture-of-experts (MoE) models whose residual streams are no longer a single tensor and whose weights ship quantized. We apply it to GLM-5.3-Flash (320B parameters, 288 routed experts, a four-wide hyper-connection residual, block-FP8). The attack survives the architecture, but what it reaches is no longer where a reader of the original recipe would look for it. Editing the attention, dense and routed-expert writers on their own removes 0.039, 0.016 and 0.148 of refusal respectively; editing all three together removes 0.776. As a result, 74% of the effect exists only under the joint intervention. The part the conventional recipe reaches by module-name matching accounts for 0.066 of that 0.776, which is why it fails silently on an MoE. The effect does not follow from removing just any direction: ablating a random direction orthogonal to it leaves refusal unchanged. A category-concentrated residue survives every edit we tried: subspaces fitted on violence, sexual content and hate leave measurable refusal at every rank from 1 to 12. We report the method, the 41-89 percentage-point reductions it achieves across seven harmful benchmarks with no detected change in capability, and the boundary where it stops.
Read the full abstract
Numerical Analysis
Uniform Chebyshev approximations of functions satisfying a variation-type condition: reconstruction of time-varying signals on graphs
We prove uniform approximation theorems for Chebyshev expansions of continuous functions of two variables. Under a variation-type condition on the square $[-1,1]^2$, a continuous function admits a uniformly convergent Chebyshev expansion in the first variable whose coefficients are continuous functions of the second variable. These results are applied to the spectral graph wavelet transform of time-varying signals on finite weighted graphs: the scaling and wavelet kernels are approximated by expansions with time-varying coefficients whose degrees do not depend on time, the composition of the approximate transform with its adjoint admits the same explicit coefficient formulas as in the time-independent case, and reconstruction by the pseudoinverse is stable, with explicit bounds in terms of the uniform kernel errors. Numerical experiments on a sensor network confirm the convergence and stability estimates. In a denoising problem, soft thresholding of the graph wavelet coefficients with a time-varying transform parameter improves over its time-independent, fixed-parameter counterpart.
Read the full abstract
Computation and Language
MUCnoHARM@GermEval Shared Task 2026: Retrieval-based In-Context Learning for Defamatory Offences, and Where It Falls Short
With hate speech being ubiquitous online, automatic detection is crucial, in particular when it comes to criminally relevant social media posts. We study a variety of retrieval-based in-context learning (RetICL) strategies for detecting defamatory offences under {\S}{\S} 185-187 StGB (the subject of GermEval 2026 Subtask 4). Few-shot prompting beats zero-shot, but retrieval-based approaches offer only marginal gains over random demonstrations, and even fall behind an optimised static set of demonstrations. Providing concrete legal knowledge helps, yet model choice outweighs every other system choice. Models over-predict criminal relevance while still missing 26-57% of criminally relevant posts, suiting them for triage rather than autonomous moderation.
Read the full abstract
Computer Vision
LogiScope-VQA: Benchmarking Vision-Language Models for Logistics Hazard Identification in Industrial Scenarios
Large Multimodal Models (LMMs) large-scale deployment in industrial warehouse settings specifically necessitates that models exhibit human-expert-level hazard-oriented perception, understanding, and reasoning capabilities. However, the scarcity of real industrial data, tightly coupled to commercial terms, significantly hampers further advancement. To bridge this gap, we curate LogiScope-VQA to investigate the practical applicability of mainstream LMMs in real-world logistics operations. LogiScope-VQA comprises 2,476 images and 2,918 videos primarily sourced from real-world logistics parks, along with 10,274 VQAs meticulously curated and validated by human annotators. Grounded in 18 core objects and 20 risk types, we devise 39 subtasks aligned with three principal themes: industrial element perception, warehouse knowledge understanding, and potential risk reasoning. Furthermore, we incorporate dynamic thinking-budget configurations and dual-dimensional risk bias analyses to elucidate the properties of LMMs. Extensive experiments unveil that even powerful proprietary models, including GPT-5.5, Gemini-3.1-Pro, and Claude-Opus-4.7, exhibit a significant gap relative to human performance. The unique challenge of jointly integrating perception, understanding, and reasoning for hazard identification poses substantial headroom for further improvement on LogiScope-VQA. We additionally reveal the pervasive security bias issue that impedes LLMs' practical deployment in real-world settings. The industrial dataset is publicly available under the CC BY-NC-SA 4.0 license.
Read the full abstract
Human-Computer Interaction
Pairit: A Platform for Live Experiments on Human-AI Collaboration
Organizational design in the era of artificial intelligence requires experimental methods that can test how human-AI groups coordinate, delegate, and make decisions. Programmable platforms coordinate live human-to-human sessions or real-time human-AI chat, but researchers cannot easily declare experiment protocols in which AI participants both communicate and act on shared work within one auditable configuration. Here we introduce Pairit, an online platform that facilitates the design, testing, and deployment of experiments that test human-AI organizational designs and interventions. Through a single YAML configuration file, researchers declare an executable experiment graph (pages, routing, randomization, matchmaking, chat, shared workspaces, server-hosted agents, surveys, timers, and custom HTML components) and combine any number of humans and AI agents in live sessions. We have validated the feasibility of the platform through multiple live deployments, including peer-reviewed published studies, capturing high-resolution process traces of communication, negotiation, and collaborative work in live human-AI dyads. By representing complex interactive protocols as standardized, auditable configuration files, Pairit provides reusable infrastructure for specifying, deploying, and sharing live human-AI organizational experiments.
Read the full abstract
Machine Learning
Finite Audits of Model Updating: Coverage and Resolution of Paired Disparity Comparisons
Finite audits can become more precise and decisive without reliably covering the change in subgroup disparity. We study the updating-minus-frozen difference in absolute subgroup gaps, averaged over a fixed history, for true-positive and false-positive rates separately. Controlled paired-count experiments and independent audits of retained model histories distinguish interval coverage, erroneous conclusions and decision resolution. With a zero frozen gap and a five-percentage-point updating gap, bootstrap coverage is 12.0%; a numerical-delta alternative improves coverage to 52.9%, still far below the nominal 95%, with 500 smaller-group and 2,000 other-group observations of the specified outcome per window over ten windows. Unequal absolute-gap biases explain point-estimate displacement, and gap direction changes pairing's variance contribution. Prior strength materially affects sparse, off-centre comparisons. In the retained panel, relatively high mean coverage coexists with substantial undercoverage for individual histories, while larger audits can considerably improve resolution. Conservative projection provides conditional coverage under the sampling assumptions but often leaves comparisons unresolved. These findings show why interval width and conclusive margin decisions cannot, by themselves, establish the reliability of an estimated disparity change.
Read the full abstract
Systems and Control
Spatial LLM Workload Shifting Needs Foresight: Model Commitment for AI Data Center Operation under Power Grid Constraints
AI data centers may face power supply shortages during certain periods, requiring operators to shift large language model (LLM) inference workloads spatially to maintain service rates. However, existing workload-shifting methods typically assume that any data center with sufficient computing resources can immediately serve shifted requests, which may lead to infeasible transfers and unserved demand. This letter proposes model commitment (MC), a mixed-integer linear programming framework that jointly schedules model deployment and cross-site request routing under power constraints and electricity-price signals. First, MC formulates the intertemporal coupling introduced by model replica loading. Second, it translates prefill and decode latency requirements into the amount of demand that each replica can serve. Case studies based on real-world data show that MC enables AI data center operators to achieve a 100% service rate under time-varying grid conditions and reduce total operating cost by 29.0%.
Read the full abstract
Machine Learning
NEXUS-MI: Communication-Aware Federated Personalization for Gateway-Coordinated Motor-Imagery Brain-Computer Interfaces
Electroencephalography (EEG)-based motor-imagery brain-computer interfaces (MI-BCIs) vary across subjects and sessions, complicating personalization from limited calibration data. Federated learning can exploit shared representations without centralizing raw EEG, but existing federated MI studies largely assume regular synchronization. We introduce NEXUS-MI, a gateway-coordinated federated personalization framework that treats synchronization as a coupled learning-and-communication control problem. Raw EEG and classifier heads remain local, while an edge coordinator maintains the shared backbone. We evaluate NEXUS-MI through offline replay using BCI Competition IV Dataset 2a (BCICIV-2a; 9 subjects, 4 classes) and OpenBMI (54 subjects, 2 classes). Session 1 supports backbone learning, and Session 2 provides limited-calibration personalization and held-out testing. An ideal-link reference and six heterogeneous-link policies characterize gateway participation, buffering, stale-update admission, and backbone-download control. The principal comparison holds delayed-update handling fixed while contrasting non-adaptive and communication-aware synchronization. Paired subject-level comparisons use Holm adjustment, and robustness across five matched realizations is assessed by hierarchical bootstrap. Communication-aware coordination reduced server-to-client backbone traffic by approximately 42% on both datasets, while cohort-level accuracy differences were small and realization-dependent. Cohort averages also concealed subject-level vulnerability, with losses reaching approximately 12 percentage points on BCICIV-2a relative to the ideal-link reference. These findings establish gateway synchronization as an explicit design variable in federated MI personalization and motivate joint evaluation of personalized accuracy, communication cost, update freshness, and subject-level reliability.
Read the full abstract
High Energy Physics (Phenomenology)
More QCD Masterclass Lectures on Jet Physics and Machine Learning
These lectures were presented at the 2026 QCD Masterclass in Saint-Jacut-de-la-mer, France. They supplement the published notes from the 2024 school with presentation of a few new topics that have become popular in recent years. These include event metrics, anomaly detection, scaling laws, the limits of classification performance, and regression that respects spacetime symmetries. Many of the presented results are known, but some new results are derived, especially related to anomaly detection. As was the case in those original notes, while all these problems are motivated from machine learning applied to QCD, no discussion of details of machine learning is presented in these notes. All results follow from application of fundamental principles of QCD. I begin these notes with an apology for humanist physics.
Read the full abstract
Signal Processing
Cooperative LEO-Terrestrial Multistatic ISAC: CRLB Analysis, Scaling Laws, and Satellite Selection
Low Earth orbit (LEO) satellites provide elevated and spatially diverse viewpoints for enhancing three-dimensional (3-D) sensing in integrated satellite-terrestrial networks (ISTNs). This paper investigates a LEO-assisted terrestrial multistatic integrated sensing and communication (ISAC) network for 3-D target localisation, where multiple LEO satellites act as cooperative sensing illuminators and provide additional bistatic observations to distributed terrestrial radar receivers. We first model cooperative satellites as a homogeneous Poisson point process (PPP) and derive a tractable approximation of the average hybrid Cramér-Rao lower bound (CRLB). The resulting scaling laws show that the root-CRLB decreases with the inverse square root of the mean number of cooperative satellites for a fixed cooperation region, while increasing the cooperation radius at fixed satellite density yields logarithmic diminishing returns. We then develop an Earth-curvature-aware Walker model incorporating orbital structure, satellite motion, visibility, and time-varying sensing geometry, and derive a tractable approximation of the corresponding hybrid CRLB. Analytical bounds on the marginal gain and a sufficient condition for ordering candidate satellites are obtained. Based on these results, a CRLB-oriented greedy satellite-selection strategy is proposed to account for SCNR-dependent reliability and geometric complementarity with the terrestrial sensing configuration. The proposed strategy consistently outperforms benchmarks and approaches exhaustive-search performance with substantially lower complexity. Monte Carlo simulations validate the analytical approximations for both models.
Read the full abstract
Machine Learning
BRACE: Anchored Bellman-Residual Correction for Stale Critics in Asynchronous RL
Asynchronous reinforcement learning has become the standard way to scale training for large language models (LLM), but the resulting policy lag biases the critic toward the stale behavior policy. Existing work on asynchronous LLM training corrects the actor and leaves this bias unaddressed, while the off-policy value correction of classical RL does not carry over to long-horizon agentic tasks, since a short correction horizon leaves the regression target free of the reward and a long one lets the product of importance ratios drift exponentially with the trajectory length. We propose BRACE, an anchored Bellman-residual correction for stale value models. BRACE bounds the correction horizon to a prefix of policy tokens and anchors a constant-weight Monte-Carlo tail beyond it, which separates policy correction from reward propagation. BRACE delivers a $9.8\%$ relative improvement in mean@1 on BrowseComp-Plus over the strongest baseline, runs $2.46\times$ faster per step than synchronous training, and remains stable $50$ updates off-policy.
Read the full abstract
Information Theory
Channel Estimation for OFDM via Delay-Doppler Refinement
In this paper, we propose a novel channel estimation (CE) algorithm for orthogonal frequency division multiplexing (OFDM) systems that exploits the unique characteristics of the delay-Doppler (DD) domain channel. Specifically, the time-frequency (TF) domain input-output relationship (IOR) is derived in a compact form by focusing solely on the non-zero elements of the TF domain channel matrix. Based on this compact IOR, a coarse TF domain CE is first performed using a linear minimum mean square error estimator. Then, the resultant TF domain estimates are transformed to the DD domain through a unitary transformation for further refinement. We reveal that the effective DD domain channel matrix can be viewed as an aggregation of multiple DD domain channel responses with different phase shifts. This allows us to devise a threshold-based estimation for DD domain channel parameters with high accuracy. The estimated DD domain channel parameters are then applied to form a refined estimate of TF domain channel. Our numerical results demonstrate that the proposed method can achieve substantial performance gains over conventional OFDM channel estimation techniques under the same pilot deployment.
Read the full abstract
Strongly Correlated Electrons
Semi-Dirac States and Quantum Linear Magnetoresistance in Helimagnetic Pnictide MnP
Large linear positive magnetoresistance (LPMR) in topological and magnetic materials remains a subject of intense debate, particularly in noncollinear spin systems where spin-dependent scattering complicates charge transport. Manganese phosphide (MnP), a helimagnetic binary pnictide with multiple field-induced magnetic transitions, provides a useful platform to investigate the interplay between complex magnetism and electronic topology. Here, we present a comprehensive experimental and theoretical investigation of phase-dependent magnetotransport in high-quality MnP single crystals. Hall measurements reveal an anomalous Hall effect dominated by skew scattering at high temperatures and a finite topological Hall effect in the noncollinear fan (FAN) and low-temperature screw (SCR) phases. At low temperatures, we observe a large, non-saturating LPMR reaching nearly 800 percent at 4 K and 15 T, with a pronounced linear field dependence in the field-polarized ferromagnetic (FM2) state. First-principles calculations reveal a strongly anisotropic semi-Dirac-like band at the Y point that progressively approaches the Fermi level from the SCR to FAN and FM2 states. Our analysis indicates that the resulting small Fermi pocket can access the extreme quantum-limit regime at experimentally accessible fields, providing a microscopic framework for the observed LPMR within Abrikosov's quantum magnetoresistance theory.
Read the full abstract
Solar and Stellar Astrophysics
The changing optical and X-ray emission of the dormant $\gamma$ Cas star HD 45314
$\gamma$ Cas stars are Oe/Be stars that exhibit bright and hard X-ray emission. HD45314 belonged to this category, but lost its $\gamma$ Cas characteristics as its circumstellar disc started to fade away. The star was monitored in optical spectroscopy, optical photometry, and X-ray spectroscopy to follow its variability on timescales of hours to years. Time series of magnitudes, colours, line equivalent widths and violet over red peak intensity ratios were analysed with Fourier methods to uncover possible periodicities. For five years, HD45314 exhibited oscillations of its magnitude and line strengths on a timescale of 230 d. It then declined towards a nearly disc-free stage and is now slowly rebuilding its disc. During these phases, the violet over red peak intensity ratio exhibited a modulation on a timescale of about 1190 d. Over the whole campaign, prominent short-term photometric variations, notably at a frequency of 3.369 d$^{-1}$, were probably due to $\beta$ Cep-type pulsations. The amplitude of these pulsations was stronger during phases of overall brightness changes. In parallel, the X-ray spectrum switched to a low state at the onset of the oscillation phase and has remained in this low state since then. Some residual hard emission is nonetheless observed. HD45314's long-term optical variations bear resemblance with the predictions of smooth particle hydrodynamics calculations, although a close look does reveal some inconsistencies. Unlike some other $\gamma$ Cas stars, where the X-ray emission remained nearly constant despite major changes of the circumstellar disc, HD45314's X-ray properties changed dramatically as the disc begun to dissipate. These differences might arise from a wider orbital separation between the Oe star and its putative white dwarf companion or from the ablation of the disc by the radiation field of the Oe star.
Read the full abstract
Quantum Physics
Split-Step Dirac Cellular Automata for Continuous-Time Dirac Dynamics on Finite Spatial Lattices
Dirac Cellular Automata (DCA) provide a framework for simulating Dirac dynamics, yet the rigid coupling between spatial and temporal resolutions can introduce artificial phase-matching symmetries on finite grids that suppress interference phenomena such as Zitterbewegung. In this work, we propose a Split-Step Dirac Cellular Automaton (SDCA) that enables continuous-time Dirac evolution at fixed spatial discretization. By employing a Trotterized fractional-step scheme in the momentum representation, SDCA breaks the phase-matching cancellation present in the standard DCA and recovers the interference dynamics of the continuous-time limit. We benchmark the SDCA through analytical and numerical studies and demonstrate its implementation on IBM Quantum processors. Despite the increased circuit depth required for finer temporal resolution, the NISQ implementation reproduces the characteristic velocity oscillations and entanglement-entropy dynamics of the continuous-time model. We further investigate hardware-topology trade-offs and dynamic circuit implementations of the Quantum Fourier Transform (QFT), highlighting the competing effects of gate errors, measurement, and feed-forward latency. These results demonstrate that SDCA provides a practical framework for improving temporal resolution while maintaining a fixed spatial quantum register, enabling the exploration of relativistic quantum dynamics on near-term quantum devices.
Read the full abstract
Computation and Language
ROAM: Robust Organization of Atomic Memories for Agents through Semantic Relations
Long-term language-model agents rely on external memory across interactions. Atomic memories are particularly useful: their fine-grained semantic boundaries enable precise retrieval and direct comparison between observations. Yet accumulating atoms inevitably become redundant, overlapping, or conflicting. Existing methods often ask an LLM manager to add, update, delete, or rewrite memories directly, coupling semantic interpretation, storage decisions, and content generation in one error-prone operation. We introduce ROAM, a relation-guided framework that uses atomicity for management while allowing richer answer-time representations. ROAM classifies incoming–stored atom pairs as independent, equivalent, directionally subsuming, or conflicting, then organizes observations into active Primary and supporting Evidence roles. Fusion subsequently combines complementary details and temporal changes into compact, potentially non-atomic views. Only Primary views are retrieved for answering, preventing redundant or outdated atoms from competing independently. Across models and evaluation settings, ROAM improves answer accuracy by up to 29.8 percentage points. Ablations show complementary benefits from different relations and consistent gains from fusion beyond role organization. Mechanism analysis further finds 15.6-point higher answer-critical source recall and an 11.5-point lower confounder-token share. ROAM remains robust across manager scales.
Read the full abstract
Signal Processing
Multi-Stream Spatiotemporal Channel Coding for MIMO Systems: Transmission Scheme Design and Achievable Rate Optimization
Spatiotemporal channel coding (STCC) can improve the achievable rate over traditional temporal channel coding (TCC) by leveraging spatial degrees of freedom to extend the codeword length. Although several information-theoretic foundations on STCC have been established, the investigation of transmission schemes from a communication-theoretic perspective remains in its early stages. This paper proposes a multi-stream over multi-subchannel STCC (STCC-MSC) under full channel state information assumption and optimizes its achievable rate in the finite blocklength regime. We first formulate the transmission architecture of STCC-MSC in a point-to-point MIMO system, which introduces a stream-subchannel matching mechanism. We then maximize the achievable rate of STCC-MSC by jointly optimizing the subchannel assignment and power allocation strategies, which is formulated as a mixed-integer-nonlinear-programming problem. Next, a penalized alternating convex approximation (PACA) algorithm is proposed to solve this problem. Subsequently, we extend the point-to-point STCC-MSC designs to the more general multi-user MIMO systems, including both uplink and downlink scenarios. Finally, simulation results indicate that the PACA algorithm achieves a 9.85% rate improvement over the benchmark algorithm within the STCC-MSC scheme. Furthermore, the joint STCC-MSC-PACA scheme improves the achievable rate by 28.68% over TCC scheme.
Read the full abstract
Artificial Intelligence
Proof-Carrying Cognition: Closing the Verification Gap with Reality-Settled Reward
Frontier gains in language-model reasoning come from reinforcement learning on reasoning traces and are concentrated in domains with a cheap, sound verifier. We argue the field's binding constraint is the verification gap: no scalable, incorruptible reward for reasoning outside formal domains. We make four contributions. (1) Theory: in a joint-Gaussian model of best-of-N selection, verifier-gold correlation rho is the exact exchange rate between test-time compute and capability, and an unsound verifier pays a polynomial penalty N^(1/rho^2); a margin-free copula form predicts realized soundness of real LLM judges to 4% median error. (2) Demonstration: in program-synthesis testbeds with executable ground truth, including a pre-registered scaled replication, unsound verifiers lose Soundness-under-Pressure as optimization grows (0.94 to 0.32 at N=4096) while a sound verifier improves monotonically; reality-anchored settlement beats a frozen verifier under i.i.d. and adversarial pressure, driving the hacking gap from ~0.27 to ~0; soundness scales log-linearly with settled labels, with on-policy settlement ~10x more label-efficient than random labeling. With real LLM judges and unit-test execution as gold, a weak judge loses soundness under best-of-N (p<0.001), a stronger judge is more robust, and selection alone manufactures +0.53 hacking gaps from honest samples. Under real GRPO training, a frozen reward model traces the full overoptimization curve (executed reward collapses 90%) while the same model refit on a 10% settlement stream preserves 6x the executed reward. (3) Paradigm: proof-carrying cognition, where reasoning steps are typed probabilistic claims priced by a self-built world model trained only on held-out reality and settled by proper scoring rules. (4) Benchmark: we specify Soundness-under-Pressure as the headline metric for a reality-settled reasoning benchmark.
Read the full abstract
Functional Analysis
Power quasinormal operators and the root problem
In this paper, we construct an $n$-power quasinormal operator $T$ such that $T^n$ is not quasinormal for some positive integer $n$, thereby providing a counterexample to \cite[Lemma 3.1]{ko-filomat-2023}. We then investigate the relationships among the $n$-power quasinormality of $T$, the normality of $T^n$, and the quasinormality of $T^n$, and show that these three conditions are equivalent in finite-dimensional spaces. We also provide a new proof that $n$-power quasinormal operators have the single-valued extension property \cite[Theorem 3.2]{ko-filomat-2023}; unlike the original proof, our argument does not rely on \cite[Lemma 3.1]{ko-filomat-2023} and thus closes the gap in the original argument. In addition, for a fixed operator $T$, we characterize all positive integers $n$ for which $T$ is $n$-power quasinormal. As consequences, several results of Sid Ahmed \cite{ahmed-bmaa-2011} are extended. Finally, we prove that every paranormal $n$-power quasinormal operator is quasinormal. Closely related to this, we also give an affirmative answer to the root problem of Stanković and Kubrusly \cite[Question 2.11]{stankovic-afa-2025}.
Read the full abstract
Artificial Intelligence
Procedural Memory Under Change: Reuse and Interference in Controlled Web Tasks
Procedural memory lets language agents reuse successful routines, but reuse presumes that a stored routine remains applicable. We study what happens when that presumption is deliberately violated. The study combines a retrospective, human-assisted interface-adaptation case from BrowserGym TimeWarp with controlled frozen-memory comparisons on synthetic shopping decisions. During the documented WebShop V1-V6 development path, interface-specific code was adapted while the separately stored high-level procedure was not reported to change; this phase does not constitute an autonomous memory-agent evaluation. In the controlled phase, an early pilot produced one task on which two memory conditions selected a more expensive item while the no-memory condition selected the reference minimum. Follow-up probes did not establish a recurring row-order or identity-binding pattern. We then tested four forms of mismatch: changed quantities, a different evidence representation, a conflict between local and global optimization, and distributed promotion evidence, across 32 formal cells. Each cell used one temperature-0 generation with the same local qwen3:8b configuration and no adaptive retry. Across these pairs, none of the predefined diagnostic interference signatures appeared on the tasks for which they were defined when current-task evidence was explicit and sufficient. The result identifies a tested region of non-interference: a procedural memory can be mismatched without becoming behaviorally disruptive. It does not establish general safety or a mechanism. The remaining question is which additional conditions turn applicability mismatch into observable, memory-caused error.
Read the full abstract
Computational Physics
Nonlinear flame describing function and mean shift kinematics of slit flames under combined axial-transverse forcing
This study investigates the nonlinear kinematics of a premixed slit flame using a two-dimensional $G$-equation level-set framework. Results show that combined forcing induces nonlinear saturation in the FDF, characterized by early gain flattening and premature phase drops, which intensify with the transverse forcing amplitude. Kinematic analysis reveals that this geometric nonlinearity manifests as a reduction in the time-averaged flame height, defined as the mean shift. In the quasi-steady limit, this mean shift is analytically quantified via a multivariate asymptotic expansion, where fourth-order terms successfully capture the saturation mechanism at elevated amplitudes. By introducing a scaling parameter to account for transverse dominance, the frequency-dependent decay of the mean shift in the compact limit collapses onto a single master curve, enabling the derivation of a unified theoretical model that integrates this asymptotic response with a second-order low-pass filter. Furthermore, because the mean shift reduces the physical extent of the flame, it alters the wrinkle propagation time. Correcting the Strouhal number using the measured mean shift collapses the dispersed nonlinear FDF curves onto the linear theory prediction. The analysis is further extended to disturbances convected at a finite speed, for which the linear transfer function is derived analytically and the correction with the measured mean shift continues to collapse the nonlinear FDF. These findings establish that the nonlinear FDF behavior under multidimensional forcing is fundamentally governed by the kinematic mean shift, providing a theoretical baseline for decoupling geometric nonlinearities from other thermo-diffusive or hydrodynamic instabilities in turbulent flames.
Read the full abstract
Computation and Language
SymbolicLight V2: Hybrid Neuromorphic Architecture and Sparse Execution for Low-Energy Language Inference
SymbolicLight V2 combines sparse event computation with continuous-state processing in a hybrid neuromorphic language architecture. Extending V1's spike-gated dual paths, it adds graded signed events at further projections and softmax-free local attention. We implement the 194M-parameter model on an Alveo U50C FPGA using digital fixed-point arithmetic and on an ARM CPU using sparse integer execution. Across three same-checkpoint FPGA implementations at 175 MHz, active-row weight gathering and valid-state KV loading raise decode throughput from 474.6 to 643.2 tokens/s for a 32-token prefix and 128 outputs. Estimated gross card energy falls from 0.06087 to 0.04407 J per generated token, a 27.6% reduction. Complete-request energy, including prefill, falls by 24.4-27.7% across three prefix lengths. An independent idle split attributes 82.8% of gross card energy to loaded idle, explaining the benefit of shorter token latency. Against the recorded RTX 5090 compiled-FP32 baseline, integer FPGA execution uses 89.1% less estimated card energy during short-context decode; arithmetic precisions differ, and the GPU baseline is not the lowest-energy tested configuration. On four Cortex-A76 cores of a ROCK 5T, complete requests reach 65.4 tokens/s at 9.80 W and 0.151 J per generated token at the adapter's AC input. These results connect event sparsity to omitted computation and data movement. The mechanisms also support other dedicated V2 implementations: increasing throughput by a greater factor than active power lowers energy per generated token. Evaluation holds the deployed checkpoint fixed; its quality trails a same-budget dense control, so the results do not establish equal-quality efficiency.
Read the full abstract
Materials Science
Facile hBN-hBN Interfacial Overlap Engineering for Enhanced Quantum Emitter Formation
Quantum emitters in two-dimensional materials, particularly hBN, are promising platforms for quantum technologies. However, achieving high-density emitters at predetermined locations while preserving optical quality remains challenging. Here, we introduce a facile, cost-effective double-layer all-dry transfer approach to deterministically create overlap regions between hBN flakes. These pre-defined capped regions exhibit a significantly enhanced emitter density, with up to a 15-fold increase compared to uncapped areas. Importantly, this method does not compromise emitter quality: emitters within overlap regions demonstrate excellent optical performance, including high signal-to-background and signal-to-noise ratios, large Debye-Waller factors, high brightness, and strong spectral stability. Possible defect configurations are also discussed to contextualize the observed emission characteristics. This scalable strategy enables preferential formation of quantum emitters in targeted regions, achieving higher densities than simple treatments such as plasma irradiation while avoiding the complexity of advanced fabrication techniques. The approach provides a practical pathway for integrating high-quality quantum emitters into scalable quantum photonic platforms.
Read the full abstract
Classical Analysis and ODEs
Optimal dimension-dependent $\ell^p$ and $\ell^{1,\infty}$ estimates of the second-order discrete Riesz transforms
In this paper we investigate the optimal dimension-dependent estimates of the second-order discrete Riesz transforms \[ R_{\mathrm{dis}}^{(jk)}f(n) = c_d\sum_{m\in\mathbb Z^d\setminus\{0\}} \frac{m_jm_k}{|m|^{d+2}}f(n-m), \qquad c_d=\frac{\Gamma\left(\frac{d+2}{2}\right)}{\pi^{d/2}}. \] For $j\neq k$ and every fixed $1<p<\infty$, we prove that \[ \|R_{\mathrm{dis}}^{(jk)}\|_{\ell^p\to\ell^p} = c_d\left[ \frac{2}{2^{d/2}} + \left(\frac83+o(1)\right)\frac{d}{3^{d/2}} \right] \] and \[ \frac{2c_d}{2^{d/2}} \leq \|R_{\mathrm{dis}}^{(jk)}\|_{\ell^1\to\ell^{1,\infty}} \leq c_d\left[ \frac{2}{2^{d/2}} + \left(\frac83+o(1)\right)\frac{d}{3^{d/2}} \right]. \] Since $ c_d\sim \sqrt{\pi d}\left(\frac{d}{2\pi e}\right)^{d/2} $ by Stirling's formula, thl $\ell^p$ estimates give a negative answer to the conjecture proposed by Bañuelos and Kim in \cite{BK2}. The diagonal case exhibits quite a different phenomenon: for every $j$ and every $1<p<\infty$, $R_{\mathrm{dis}}^{(jj)}$ is neither bounded from $\ell^p(\mathbb Z^d)$ to $\ell^p(\mathbb Z^d)$ nor of weak type $(1,1)$. Cancellation is restored for the operators $R_{\mathrm{dis}}^{(jj-kk)} =R_{\mathrm{dis}}^{(jj)}-R_{\mathrm{dis}}^{(kk)}$. For every fixed $1<p<\infty$, \[ \|R_{\mathrm{dis}}^{(jj-kk)}\|_{\ell^p\to\ell^p} = 4c_d\left[ 1+(1+o(1))\frac{d}{2^{d/2}} \right], \] and \[ 4c_d \leq \|R_{\mathrm{dis}}^{(jj-kk)}\|_{\ell^1\to\ell^{1,\infty}} \leq 4c_d\left[ 1+(1+o(1))\frac{d}{2^{d/2}} \right]. \]
Read the full abstract
Software Engineering
XAgent: eXecution-guided Agentic AI for Effective Localization and Resolution of GitHub Issues
Agentic AI has enabled capabilities in leveraging Large Language Models (LLMs) to autonomously resolve repository-level GitHub issues. However, due to the reliance on limited static description of issues, existing agentic approaches suffer from incorrect localization and incomplete validation. Solely relying on this information can bias LLM reasoning toward the narrow scope of the issue description, leading to incomplete patches that fail to address the underlying issue. In this paper, we present XAgent, an execution-guided agentic framework that analyzes dynamic behavior and additional program context to localize and validate issues. The experimental results on the SWE-bench-lite dataset demonstrate that XAgent outperforms other existing approaches, achieving a resolve rate of 62.0% and a function localization accuracy of 72.8%, while maintaining cost efficiency. Our analysis further shows that XAgent successfully resolves 7 additional issues that the top existing baselines fail to address. This work highlights a shift from static, description-oriented patch generation toward dynamic execution-guided issue resolution, opening new opportunities for LLM-based coding agents to achieve more robust and generalizable software maintenance.
Read the full abstract
Machine Learning
Fine-Tuning a KV Cache Concatenation-Aware Model or Recomputing KV Caches? Why Not Both?
In Retrieval-Augmented Generation (RAG) systems, a large number of retrieved chunks are concatenated to form the input context so that users can receive high-quality responses based on external knowledge. As a result, the input context length increases substantially, leading to a larger prefill workload and, in turn, a longer time to first token (TTFT). While previous works that reuse precomputed key-value (KV) caches effectively reduce TTFT for long-context inputs, it remains unclear whether response quality is preserved when the input context becomes very long. In this paper, we propose a combined approach that (i) fine-tunes the model while taking KV cache concatenation into account and (ii) selectively recomputes a subset of the KV caches. By applying both techniques, we demonstrate improved accuracy for long-context inputs. Experiments on the RULER benchmark show that, for a 124k-token input, our method improves the RULER score by 9.7 point over the baseline that recomputes KV caches only. Moreover, TTFT is reduced by 80% compared with full attention.
Read the full abstract
Solar and Stellar Astrophysics
Connecting Dynamo Theory with DNS Data: A Computational Analysis of $\alpha$ and $\beta$ effects
We investigate the influence of current helicity on the turbulent magnetic diffusivity $\beta$ using three complementary derivations of the $\alpha$ and $\beta$ coefficients, based on the large-scale magnetic field $\overline{\mathbf{B}}$, the turbulent velocity $\mathbf{u}$, and the turbulent magnetic field $\mathbf{b}$. Applying these coefficients to raw DNS data, we reconstruct $\overline{\mathbf{B}}$ and compare the results with the original simulations. In the kinematic regime all models agree well with the DNS data. In the nonlinear regime, however, $\beta_{\mathrm{vv-vw}}$ alone produces unbounded growth of $\overline{\mathbf{B}}$. Including the contribution from turbulent magnetic fields ($\beta_{\mathrm{bb+jb}}$) suppresses this unphysical growth and restores agreement with the DNS results. We find that kinetic helicity drives $\beta$ more negative, while current helicity shifts it back toward zero. Weighted combinations of the coefficients further show that the $\beta$ effect dominates the evolution of $\overline{\mathbf{B}}$ throughout, whereas the $\alpha$ effect becomes important mainly for sustaining the field in the nonlinear regime. The corresponding IDL analysis scripts are provided to facilitate practical implementation of the theoretical models.
Read the full abstract
Computation and Language
CARRE: Counterfactual Action Retrieval and Reason Evaluation for Explainable Churn Prescription
Churn models typically identify high-risk customers but do not specify which feasible retention action should be considered or why that action is appropriate. We present CARRE (Counterfactual Action Retrieval and Reason Evaluation), a three-stage framework that combines retrieval-augmented candidate generation, cost-aware counterfactual scoring, and large language model (LLM) reasoning. CARRE retrieves a predefined catalog of retention actions, estimates model-predicted churn-risk changes under explicit feature transformations, and generates a structured churn reason and a profile-grounded explanation for the selected action. On the IBM Telco Customer Churn dataset, CARRE achieves 79.8% greater mean model-predicted risk reduction than the plain SHAP baseline and 80.4% greater reduction than the cost-controlled SHAP+Cost baseline across 313 high-risk test cases; its cost-normalized efficiency is 10.5% higher than that of plain SHAP. On a 136-case reason-stratified evaluation sample, diagnosis-driven prompt refinement increases weak-label agreement from 79.4% to 90.4%, with no auxiliary-plan constraint violations; because the same sample was used for error diagnosis and re-evaluation, the post-refinement result is not an independent estimate of generalization. For 135 explanations generated using the pre-refinement v2 reason outputs, two cross-vendor LLM judges assign mean scores ranging from 4.02 to 5.00 out of 5, although one judge saturates on actionability, and a deterministic audit finds no contradictions among 66 verifiable profile claims. Retrieval ablations show that k=5 provides the best evaluated compromise between high candidate coverage and downstream reasoning agreement in this dataset. These results illustrate how retrieval, model-based counterfactual scoring, and language generation can be separated and jointly evaluated in a prototype churn-prescription pipeline.
Read the full abstract
Astrophysics of Galaxies
Tracing M22's origins: Spatial and chemical constraints on its formation history
Globular clusters (GCs) are well known to host stellar populations characterized by light-element variations. A subset of Galactic GCs, beyond such 'canonical' populations, contains an additional group enriched in iron, s-process elements, and total C+N+O abundance (i.e., the anomalous stars). We combine multi-facility photometry with APOGEE spectroscopy to investigate the spatial distribution, chemical properties, and formation history of the stellar populations in M22, with particular focus on its anomalous component. We trace the canonical and anomalous populations, together with their subpopulations, from the cluster center out to eight half-mass radii. The anomalous population becomes increasingly dominant in the outermost regions, whereas no significant radial gradients are detected among its subpopulations. Our chemical analysis reveals light-element anticorrelations within both the canonical and anomalous components, although the latter are shifted toward higher C, N, and Al abundances. For the first time, we show that the Fe, s-process, and C+N+O enhancements among anomalous stars are not uniform but correlate with their light-element composition: the most chemically extreme anomalous stars are also the most Fe-, Ce-rich, and C+N+O-poor. We identify a distinct red overdensity on the horizontal branch, likely populated by the most He-poor stars, and tentatively associate the extreme horizontal branch with the most chemically enriched anomalous population. These observations are difficult to reconcile with M22 being a merger between two GCs. Instead, they qualitatively favor a self-enrichment scenario regulated by dilution, similar to that recently proposed for $\omega$Centauri, with their close chemical correspondence suggesting that they experienced analogous formation histories.