Searcharxiv⌕ Search

arXiv subjects

Ran Li

Publications and source records attributed to Ran Li.

At least 19 recordsLinked to original sources
arXiv2026

SKILL2STEER: One-Pass Hypernetwork for Mapping Natural-Language Skills to Low-Rank Post-Block Activation Shifts

Compiling a natural-language agent skill into model parameters removes the skill text from the prompt and leaves behind a small artifact that can be distributed and reused. Existing methods differ in where they inject the compiled quantity, and all of them assume that the artifact carries the content of the skill. We test that assumption. Through three controls on a retrieval agent, we find that a compiled skill shift carries a domain-general behavioural prior, not the content of the skill. A wrong skill document performs on par with the correct one, a single averaged bundle matches per-skill routing under a two-sided equivalence test, and a same-norm random direction collapses to the no-injection baseline, which shows that the trained direction stays load-bearing. We then introduce SKILL2STEER, a one-pass hypernetwork that reads a skill document once and maps it to low-rank activation shifts added at the output of every decoder block. Each skill compiles to 2.36M parameters, 9.25 times fewer than a weight-space adapter, while the base weights stay frozen and the generator runs only at compile time. The shift can be scheduled across retrieval steps and deployed as one routing-free bundle. On SearchQA, a seven-dataset benchmark, SKILL2STEER reaches 37.9 exact match, 9.8 points above the frozen model and 2.3 points above a weight-space method that ships 9.25 times more parameters per skill.

cs.LG↗
arXiv2026

An Enhanced Isothermal Jeans Approach to Constraining Dark Matter Self-Interactions from Galactic Kinematics

We present an improved semi-analytical model to predict density profiles of self-interacting dark matter (SIDM) halos and apply it to constrain the self-scattering cross section using SPARC galaxy rotation curves. Building on the isothermal Jeans approach, our model incorporates (i) velocity-dependent cross sections, (ii) an empirical treatment of core collapse, and (iii) enhanced robustness for identifying solutions. These advances allow us to fit a large sample of galaxies, including systems with baryon-dominated centers often excluded in earlier studies. We find that roughly 1/6 of galaxies admit both a core-growth and a core-collapse solution, while the rest favor a unique evolutionary state. Joint constraints across the sample reveal clear velocity dependence: the allowed parameter space forms an L-shaped degeneracy, where both nearly constant, low cross sections ($σ_0\sim2\,{\rm cm}^2$/g, $ω\gtrsim500\,$km/s) and strongly velocity-dependent models ($σ_0\sim100\,{\rm cm}^2$/g, $ω\sim60\,$km/s) are viable. Adopting the core-growth interpretation yields best-fit values $σ_0\simeq5\,{\rm cm}^2$/g and $ω\simeq250\,$km/s. Our constraints are remarkably consistent with previous results derived from a variety of independent probes. Compared to cold dark matter (CDM) models, SIDM outperforms simple adiabatic-contraction profiles and rivals empirical feedback-based CDM profiles, yet shows no correlation with stellar-to-halo mass ratio, a proxy for feedback strength, offering a distinct explanation for dwarf galaxy diversity. Moreover, SIDM does not affect galaxy-halo scaling relations significantly and makes concentration systematically lower. Our results highlight SIDM as a compelling framework for small-scale structure, while future low-mass kinematic data will be crucial for breaking degeneracies in velocity-dependent cross-section models.

astro-ph.GA↗
arXiv2026

What can strong lensing recover in edge-on disc galaxies? A mock study based on TNG50 simulations

Strong gravitational lensing tightly constrains the projected mass within the Einstein aperture, but what else imaging alone can recover for edge-on disc lenses remains untested against known truth. We construct HST/ACS-like mock observations of 35 edge-on TNG50 disc galaxies selected to resemble lenses from the Sloan WFC Edge-on Late-type Lens Survey (SWELLS), and fit them with a five-stage Source-Light-and-Mass workflow in PyAutoLens; the final model represents the lens mass distribution by scaling the Multi-Gaussian Expansion (MGE) foreground-light template with a global mass-to-light ratio and adding a spherical NFW halo. Decomposed fits succeed for 21 of the 35 systems; in the other 14, coherent MGE lens-light residuals prevent reliable modelling. For the successful subset, the Einstein radius and the total aperture mass are recovered with normalised median absolute deviation (NMAD) scatters of 0.3% and 0.7%. The global projected mass-weighted slope has an NMAD scatter of 10.3%, smaller than that of the mass-sheet-invariant reduced deflection curvature $ξ$ (11.4%) and well below that of the local slope (40.7%), while the stellar and dark-matter aperture masses scatter by 15.1% and 108.8%. A 35-system true-light diagnostic using the exact input light gives component-mass scatters of 0.8% and 6.3%. Under these controlled assumptions, imaging constrains Einstein-scale total mass and global radial structure more reliably than local radial structure or the stellar-dark-matter decomposition.

astro-ph.GA↗
arXiv2026

Latent Harness: Compiling Agent Behaviors into Post-Block Activation Shifts

Harness engineering and self-evolving agents improve language-model agents quickly, yet their gains live outside the model. The harness is re-sent and re-paid on every call, and the standard route to internalization is a weight update that is costly to iterate and fixed once merged. We propose compiling the harness instead. Latent Harness distills an external condition into a rank-8 activation shift. The shift is applied at the output of every decoder block of a frozen backbone, at 0.03-0.05% of parameters. The condition is either an explicit process-level behavior instruction or the behavioral contract of an agentic harness. At inference the compiled bundle carries no harness text and adds no tokens. It exposes three runtime controls for injection strength, injection phase, and shift-space mixing. Compiled behaviors recover the source-harness gain in 8 of 10 preregistered settings. Where the explicit source pays the heaviest per-call token tax, the compiled form exceeds it. The same bundles transfer zero-shot across tasks (5 of 7 pairs positive) and compose in shift space, while stacked text instructions interfere destructively. Used as a search space, they lift budget-matched harness search by 1.45 times. On three agentic benchmarks the operator aligns policy to the harness con- tract. Under matched data, capacity, and effective strength it outperforms same-capacity LoRA, which needs 11.1 times the parameters to reach the same band, and it regresses nothing where the base policy is already strong. Across 1B-to-8B models the method skeleton is invariant while gain locations move with the model-task pair, which delimits where compiled harnesses should be deployed.

cs.CL↗
arXiv2026

Mock Observations for the CSST Mission: CPI-C -- Targets for High Contrast Imaging

We introduce CPISM, a simulation program developed for the Cool Planet Imaging Coronagraph (CPI-C) on the China Space Station Telescope (CSST). CPISM supports high-contrast exoplanet imaging by simulating observational conditions and instrumental effects to optimize target selection and observation strategies. The modular design includes target modeling, imaging simulation, observational effects, detector response, and data product generation modules, enabling flexible and realistic synthetic observations. Validation through simulations of a bright star shows strong agreement with theoretical expectations, confirming the program's accuracy. CPISM's modular design allows flexibility, accommodating different stellar and planetary models, and can simulate instrumental noise, cosmic rays, and other observational effects. This tool aids in data processing, signal-to-noise ratio analysis, and high-contrast photometry, contributing to future exoplanet discovery and characterization efforts. The program's outputs will enhance observation planning and scientific return for the CPI-C mission, providing critical insights into exoplanetary systems.

astro-ph.IM↗
arXiv2026

Testing the isothermal Jeans model for self-interacting dark matter halos in the collapse phase

We benchmark the semi-analytical isothermal Jeans model against a high-resolution isolated N-body simulation that follows a self-interacting dark matter (SIDM) halo into deep core collapse. The model accurately reproduces the density evolution through much of the collapse phase, although it does not capture the sharp rise in central velocity dispersion during collapse. When applied to strong gravitational lensing observables, such as the projected mass and logarithmic density slope of SIDM halos, the Jeans model tracks the simulated evolution more closely than the parametric approach in the deep collapse regime. Our results demonstrate that the isothermal Jeans model provides a reliable and computationally efficient description of SIDM halo evolution.

astro-ph.CO↗
arXiv2026

Mapping Dark Matter in Merging Clusters with Weak Lensing for CSST: Morphology and Filament Recovery

Weak gravitational lensing offers a direct probe of the projected dark matter distribution in cluster-scale structures and the surrounding cosmic web. We investigate the capability of the China Space Station Telescope (CSST) to recover non-spherical dark matter morphologies using mock reduced-shear maps generated from cosmological N-body simulations. We apply the deep-learning-based MIU2Net mass-reconstruction framework to CSST-like wide- and deep-field source densities, focusing on two representative systems: filament-connected halo pairs and elongated halos. MIU2Net accurately recovers halo centers, elongated mass distributions, and the filamentary structures connecting halo pairs in both survey configurations. Halo centers are localized within about 0.1 h^(-1) Mpc, and elongated-halo shapes and orientations are reliably recovered. Filament paths are stable against shape noise, with mean normalized instability parameters I_F, defined as the mean deviation from the true filament locus in units of R_200, decreasing from 0.041 to 0.037 in the low-redshift bin and from 0.051 to 0.045 in the high-redshift bin from the wide to deep field. The recovered filament shapes remain consistent with the true maps at the percent level. Compared with conventional Kaiser-Squires reconstruction, MIU2Net better preserves low-contrast anisotropic structures while suppressing small-scale noise. These results demonstrate that CSST weak-lensing data, combined with machine-learning-based mass reconstruction, will enable statistical studies of cluster assembly, anisotropic matter accretion, and the connection between massive halos and the filamentary cosmic web.

astro-ph.CO↗
arXiv2026

Automated carotid artery atherosclerotic inflammation segmentation in PET using MSR-Net

Objective. To develop an automated method for segmentation and uptake quantification of carotid atherosclerotic inflammation on 64Cu-CANF-comb PET while accounting for partial-volume and tissue-fraction effects. Approach. Synthetic carotid PET images with known fractional-volume ground truth were generated by combining MR-derived high-resolution activity maps with GATE Monte Carlo simulation incorporating major PET physical processes and matched to the Biograph mMR acquisition and reconstruction. A two-stage network first localized the carotid vessel wall from T1-weighted MR and then estimated voxel-wise inflammation fraction from masked PET. Both stages used a multiscale residual (MSR) convolutional backbone. Performance was compared with U-Net, Transformer U-Net, DUCK-Net, and iterative Yang partial-volume correction. Main results. In a two-alternative forced-choice study, five nuclear-medicine physicians identified real images with 54% accuracy (54/100; Wilson 95% CI 0.44-0.63; exact binomial p = 0.48). On 184 simulated test slices from five held-out patients, the proposed method achieved a weighted DSC of 0.79, compared with 0.73 for U-Net, 0.69 for Transformer U-Net, and 0.75 for DUCK-Net, with corresponding HD values of 1.85, 2.29, 2.32, and 2.23 voxels. Patient-cluster bootstrap comparisons favored the proposed method for DSC and HD against all three baselines (Holm-corrected p <= 0.031). Overall ENRMSE was 0.24 versus 0.59, 0.68, and 0.59 for the three DL baselines and 0.57 for iterative Yang. In seven patients, corrected SUV showed a positive association with MR-derived lipid-rich/necrotic core with hemorrhage. Significance. The proposed framework improved carotid PET segmentation and uptake quantification under simulated conditions. The patient analysis is preliminary and requires external validation with biological reference standards before clinical deployment.

physics.med-ph↗
arXiv2026

JevCascade: Readout Stability and Inference-Time Compute Allocation for Prefill-Only Decision Models

We present JevCascade, an inference-time compute allocation framework for prefill-only decision models, built on readout stability and zero-label restricted readout prediction. Prefill-only decision models score every candidate in a menu in a single forward pass and never decode, which makes one call one to two orders of magnitude cheaper than a same-scale generative language model. We show that this read-out structure carries a testable property. When an intervention changes only the menu and leaves the input text fixed, the post-intervention accuracy is already determined by the cached first-pass distribution. The restriction predictor restricts the pass-1 probabilities to the menu, renormalizes and reads off the argmax, with no labels and no second pass. Across six model families, ten core datasets (sixteen in total) and two task types, menu-only interventions are predicted to within 4.2 points, exactly for one family. The property is ordinal, not probabilistic, since a probability-level variant of the same estimator errs by 21.0 points, and class-specific, since same-scale generative models err by 1.6 to 15.8 points and degrade as they grow. It makes inference-time compute decidable before deployment. Uniform extra passes buy calibration but almost no accuracy; at matched cost a confidence cascade beats every scheme that re-asks the same model, and curating the menu beats enlarging the model, a 0.8B model on a curated 5-candidate menu reaching 95.4% on CLINC150 against 80.0% for a 4B model on the full 150-label menu.

cs.CL↗
arXiv2026

Learning to Decide, Not to Reason: Parameter-Efficient Decision Operators via Low-Rank Activation Steering

Injecting skills into a frozen language model currently costs a million parameters and a reinforcement-learning pipeline. We introduce DecSteer, a System-1 decision operator trained by behavior cloning that lowers this cost by roughly two orders of magnitude. The default operator uses 330K parameters to match a 1.33M-parameter operator trained with reinforcement learning, exceeds or achieve comparable performance, while collapsing 3,685-token deliberation into a 6-token decision with no loss in accuracy. A rank-4 variant with 23K parameters, 1/58 of the strongest published skill operator, suffices for SearchQA and near-suffices for LiveMath, where higher rank still helps; the same recipe transfers across five tasks and three backbones, with out-of-distribution gains persisting on LiveMath problems released months after training. The gap to prior work is trainability, and it is set jointly by initialization and architecture. The initialization of prior operators zeroes the gradient of both large factor matrices at the first optimization step, whereas our zero-initialized output projection inside a shared low-rank backbone receives a gradient immediately, which a gradient-flow probe confirms directly. The gain isn't chain-of-thought compression. 23 of 57 LiveMath points beat the base model's best-of-8 sampling, and a logit-lens probe shows the operator amplifies the answer along the model's existing late-layer pathway, not writing it earlier. Gains track the base model's headroom across 13 base-task pairs, and skills compose as approximately linear operators that can be added, interpolated, and hot-swapped at inference time.

cs.LG↗
arXiv2026

Commuting slant Toeplitz operators for polyanalytic and harmonic symbols

We study commutativity and zero products of slant Toeplitz operators on the Bergman space of the unit disk. For bounded polyanalytic symbols and for bounded harmonic symbols with boundary values in $W^{1,2}(\mathbb T)$, we prove that two such operators commute if and only if their symbols are linearly dependent over $\mathbb C$, with both symbols belonging to the same class. In the polyanalytic case, no boundedness assumption is imposed on the individual analytic components. We also show that every finite-rank commutator in this class is zero and that every finite product associated with nonzero polyanalytic symbols has infinite rank. Finally, for $s\ge2$, we prove that a product of $s$ slant Toeplitz operators with bounded harmonic symbols whose boundary values belong to $W^{s-1,2}(\mathbb T)$ is zero if and only if at least one symbol vanishes identically.

math.FA↗
arXiv2026

Diffusion Reward Models

Reward models underpin the alignment of large language models, yet the dominant designs reduce each prompt--response pair to a point estimate or to a distribution from a fixed parametric family. This is at odds with human preference, which is inherently multimodal: the same response can be reasonably judged in many ways, and no single family covers all of them. To better fit this structure, we introduce DRM, a Diffusion Reward Model that recasts reward modeling as conditional density estimation over $p(\mathbf{r}\mid x,y)$. Conditioned on a frozen LLM encoder, a lightweight Diffusion Transformer denoises Gaussian noise into a reward vector, placing no parametric assumption on the output distribution and naturally representing its multimodal structure. A single architecture handles both multi-attribute regression and pairwise preference data, and at inference $N$ samples form an empirical reward distribution that can be aggregated into a scalar, a variance, or quantiles. Across five benchmarks, DRM matches or surpasses baselines under matched data and backbone, stays competitive with much larger discriminative, distributional, and generative RMs despite its modest training scale, and recovers multimodal reward structure where conventional heads collapse to a point. Uncertainty-aware rejection and lower-confidence-bound (LCB) aggregation further demonstrate that DRM can exploit distributional information beyond a scalar reward to improve reward-model decisions. Downstream RLHF experiments additionally show that using DRM as the training-time reward leads to improved policy performance, directly validating the practical benefit of diffusion-based reward modeling for RLHF training.

cs.LG↗
arXiv2026

TokenCast: Forecasting Token Consumption During LLM Agent Execution

When a large language model (LLM) agent executes the same task, token consumption can vary by over an order of magnitude across runs. The agent chooses its next steps based on tool feedback and intermediate results, while the growing context steadily inflates the input size of every subsequent call. The total consumption of a task is therefore hard to predict before execution and the prediction must be revised as the run unfolds. In this paper, we propose TokenCast, which learns a composable cost representation for each execution segment, recording its own consumption and the context growth it introduces. Composing adjacent segments yields a cumulative estimate that captures the extra input cost incurred when context from earlier segments is re-read by every later call. As execution unfolds, newly observed evidence refreshes the forecast, requiring no additional LLM calls and incurring a mean cumulative prediction time of 32.8 ms per run on SWE-bench Verified. Across 4 task suites and 6 agent models, TokenCast's mean absolute error reduction against the strongest comparator averages 14.5% over 96 evaluated combinations. In offline budget-control replay, TokenCast uses 21.3% fewer tokens on average than a fixed-budget policy at matched trace completion. The code is available at https://github.com/DEFENSE-SEU/TokenCast.

cs.LG↗
arXiv2026

Reducing False Positives in Strong-Lens Searches with Generalized-Mean Consensus of Machine-Learning Ensembles in the Kilo-Degree Survey

Context. In wide-field surveys, the main challenge is not just classifier sensitivity, but the overwhelming number of false positives. Searching for strong lenses among millions to bilions of galaxies produces many contaminants, making the bottleneck for follow-up inspection and building statistically useful lens samples. Aims. We aim to improve the purity of strong-lens candidate selection in KiDS DR4 by combining several classifiers. The objective is to retain high completeness for known candidates while substantially reducing the fraction of non-lenses. Methods. We trained convolutional, Transformer-based, and hybrid classifiers, including Li ResNet+, Swin Transformer variants, Swin-MLP, and DemiLensNet. Their probabilistic outputs were combined at score level using averaging and a generalized mean consensus. The models were tested on simulated KiDS-like lens images and then evaluated on real KiDS DR4 lens candidates embedded in a non-lens sample. Results. On the simulated test set, ensembles show no advantage over the best single models. On the mixed real KiDS test set, the arithmetic mean reduces the false-positive rate at 90% completeness from 0.016-0.020 (the range spanned by the two best individual models) to 0.011 for the seven-model ensemble. The generalized mean reduces it further, to 0.007. Applied to the full LRG and BG samples at the same 90% completeness level, the generalized mean reduces returned candidates by roughly 50% for LRGs and 70% for BGs, relative to the best single model. After visual inspection, we obtain 170 new high-quality candidates (24 Class A and 146 Class B), together with 1706 Class C candidates. Conclusions. Our results demonstrate that the generalized mean consensus of an ML ensemble strategy provides a practical route to reducing the visual inspection workload while preserving a high recovery rate of promising strong-lens candidates.

astro-ph.GA↗
arXiv2026

Identification of gravitational lenses obscured by foreground light in the KiDS dataset using U-Nets and ResNets

*Context.* Many lensing images are often obscured by foreground light from the central galaxies, making them challenging to detect. *Aims.* To address the limitations of previous lens search efforts, particularly for samples with smaller $R_E$ or faint lensed images, we developed a composite convolutional neural network framework that utilizes both U-Net and ResNet architectures for feature extraction and classification. *Methods.* We propose a hybrid search method that combines U-Net and ResNet architectures to enhance the detection of foreground galaxy-obscured lenses. Our approach consists of two main stages: first, the U-Net model separates the foreground galaxy light from potential lensing signals, creating residual images that highlight the lensing features. Next, the ResNet module performs binary classification on these residual images to detect lensing signals. *Results.* We evaluated the hybrid search method with real observational data to demonstrate its effectiveness, achieving a recall of 71.5% and a 4.5% false positive rate at a confidence threshold of 0.6. Applying this method to over 638,398 galaxy samples from the Kilo-Degree Survey Data Release 4 and conducting thorough inspections, we identify 88 Class A, 322 Class B, and 1,758 Class C candidates. *Conclusions.* This hybrid approach significantly enhances the completeness of existing strong gravitational lensing searches and shows great potential for improving future astronomical surveys.

astro-ph.IM↗
arXiv2026

LenNet: Direct Detection and Localization of Strong Gravitational Lenses in Wide-Field Sky Survey Images

Strong gravitational lenses are invaluable tools for addressing fundamental questions in astrophysics, from the nature of dark matter to the expansion of the universe. While current sky surveys have successfully identified thousands of lens candidates, the search methods employed face a critical challenge. The conventional approach relies on a "crop-and-classify" strategy, where small images are first cut out around billions of potential host galaxies before being individually classified. This process creates a significant computational and storage bottleneck that is unsustainable for future large-scale surveys. To overcome this limitation, we propose LenNet, an object detection model that identifies lenses directly within large, original survey images. Our method completely bypasses the inefficient cropping step by framing the problem as a direct detection and localization task. We initially train LenNet on simulated data to learn the complex features of gravitational lenses and then use transfer learning to fine-tune the model on a limited set of real, labeled examples from the Kilo-Degree Survey (KiDS). Our experiments show that LenNet performs remarkably well on real survey data, validating its potential as a highly efficient and scalable solution for lens discovery in massive astronomical surveys.

astro-ph.GA↗
arXiv2026

Morphology classification for galaxies in the Kilo Degree Survey using a label-efficient self-supervised learning framework

Galaxy morphology classification is fundamental to understanding galaxy formation and evolution. The advent of large-scale sky surveys has produced an unprecedented volume of galaxy images, making traditional manual classification impractical. Although supervised deep learning can achieve high accuracy, it requires large labeled datasets that are time-consuming to construct. In contrast, unsupervised methods often show limited classification performance. To address this limitation, we propose a label-efficient self-supervised learning framework for galaxy morphology classification. Our method first learns robust morphological representations from 305,583 unlabeled KiDS galaxy images through contrastive learning, and then trains a classifier using only 5,000 human-labeled images. The classifier separates galaxies into five categories: elliptical, spiral, lenticular-disk, irregular, and "other." Using a ResNet-50 model with a crop size of 64x64 pixels, our approach achieves an overall test accuracy of up to 91.0% (90.5% +/- 0.2% on average) on the human-classified catalog. The corresponding F1 scores for elliptical, spiral, irregular, lenticular-disk, and "other" galaxies are 0.96, 0.86, 0.86, 0.95, and 0.92, respectively. We apply this pipeline to the Kilo-Degree Survey Data Release 5 and produce a publicly available morphology catalog of 310,583 galaxies. This is the first morphology catalog for KiDS galaxies and provides a valuable resource for future studies of galaxy evolution. Our results show that self-supervised learning can substantially reduce the need for manual labels while maintaining high classification accuracy, making it a promising and scalable approach for automated galaxy morphology classification in the era of large-scale surveys.

astro-ph.GA↗
arXiv2026

Lower central dark matter densities in nearby galaxies than predicted by simulations

Baryonic feedback in hydrodynamical simulations is typically invoked to alleviate the core--cusp problem in dwarf galaxies. Yet baryonic processes also induce adiabatic contraction of dark matter, producing overly steep density profiles and excessively high dark matter fractions in the inner regions of massive galaxies. The dark matter distribution of galaxies across a wide stellar-mass range is therefore a critical test for such simulations, but a comprehensive benchmark has remained absent. Here, we consistently measure the dark matter distribution from galaxy centres out to radii of 20--50 kpc for 136 nearby galaxies that together span the local mass--size relation over the stellar mass interval $10^9$--$10^{11.5}\,M_{\odot}$. We identify central regions with lower dark matter densities relative to $Λ$CDM simulation expectations---whose extent grows from about 10 kpc to $>50$ kpc as stellar mass increases from $10^{10} M_{\odot}$ to $10^{11.5} M_{\odot}$. Although their physical origin remains unclear, these low--dark matter regions are clearly indicated by the data. Our results provide an important observational benchmark for future hydrodynamical simulations that explore alternative dark matter models and feedback processes.

astro-ph.GA↗