Searcharxiv⌕ Search

arXiv subjects

Zhao Chen

Publications and source records attributed to Zhao Chen.

At least 19 recordsLinked to original sources
arXiv2026

Online Label-Free Gradient Matching for Continual Language Model Training

Language models in customer-facing products must be continually updated, since the data they serve shifts over time. As retraining on the full accumulated data is impractical, each update should instead run on a small set of examples carrying the training signal of all data seen so far. The strongest current methods build this set by gradient matching, which synthesizes examples whose training gradients match those of the full data. However, they require ground-truth labels, run once on a fixed snapshot rather than on a growing stream, and are slow on large data. We introduce \textbf{OLGA}, which lifts these limitations. It distills unlabeled data by drawing the gradient matching targets from the model's own next-token predictions. Across five text classification benchmarks with phi-1.5, it improves accuracy by 10 percentage points over the strongest selection baseline using as few as five examples. OLGA also distills online data streams by keeping a running rehearsal memory of synthetic examples, which is re-distilled together with each new batch as it arrives. In the online setting, it outperforms the strongest selection-based continual baseline by 6.6 stream-AUC percentage points. Finally, our method scales to large data by pairing distillation with selection in a hybrid pipeline, letting selection handle the mild stage of compression and distillation the extreme one. It cuts runtime by 30\% versus distillation alone at similar accuracy, and beats the best selection-only pipeline by up to 12.6 percentage points.

cs.AI↗
arXiv2026

Fragility Is Not Reducible to Hardness: Conditional Causal Difficulty for LLMs

Large language models show large and seemingly arbitrary performance drops when a task is rewritten in ways that leave its requirements untouched. We recast task difficulty as a causal quantity: the treatment effect of meaning-preserving perturbations on a model. A meaning-preserving rewrite can only be observed to hurt a task the model can already solve, so we place the estimand on the principal stratum where the degradation occurs: the conditional causal difficulty (CCD), the probability that a randomly drawn twin breaks a solvable task. Because harder tasks tend to be more phrasing-sensitive, complexity enters as an effect modifier rather than a confounder, and we residualize it out to obtain complexity-residualized difficulty (CRD); zero for a task exactly as fragile as its complexity predicts, positive for anomalously fragile ones. Current benchmarks do not supply the data CCD requires, so we introduce GlassGSM and GlassMATH, augmenting GSM8K and MATH with twins built from perturbation families, raising fragility on GlassGSM by $\approx 12\times$ over GSM-IC-style distractors on the same problems ($3.7\%\to44.9\%$) while keeping the answer key unchanged. Our predictor Frail learns the per-perturbation break probability and reaches $0.613{\pm}0.023$ of the suite reliability ceiling on GlassGSM and $0.557{\pm}0.023$ on GlassMATH, against $0.171$/$0.401$ for a surface-feature baseline and $0.444$/$0.463$ for supervised baselines such as VAP3, and, at a $5\%$ curation budget on GlassGSM, doubles random recall of fragile items, matching an oracle ranking. Additionally, cross-fitted residualization leaves CRD uncorrelated with the declared proxy in practice. Code is available at https://github.com/upwork/Conditional_Causal_Difficulty_for_LLMs

cs.LG↗
arXiv2026

Deformed Q-learning for High-Dimensional Dynamic Treatment Regimes

In high-dimensional dynamic treatment regimes, nonlinear terminal outcome models complicate upstream conditional-mean modeling and inference based on recursively estimated continuation values. We introduce a Deformed Q-learning procedure for general action spaces that accommodates a terminal-stage GLM, random designs, conditional heteroscedasticity, and fine-grained action discretization. The procedure changes the recursive criterion by propagating the terminal deformed optimal Q-function through sparse block-wise linear conditional-mean models. Under the stated models and regularity conditions, we establish consistent and asymptotically normal estimators, accounting for downstream estimation uncertainty in inference for optimal Q-values, deformed optimal Q-values, and their contrasts. We bound population regret relative to the original terminal mean-outcome objective without explicit stage-wise accumulation. Estimation and inference also extend to a growing number of stages under conditions uniform across stages. Simulation studies demonstrate accurate estimation and variable selection. In an empirical application to sepsis management using the MIMIC-III database, the proposed framework supports stage-specific value inference, while an auxiliary action-alignment analysis provides an exploratory signal that early-stage concordance with the learned recommendation may be associated with improved survival.

stat.ME↗
arXiv2026

High-Frequency Estimation of the Leverage Effect under Rough Volatility with Microstructure Noise

This paper develops feasible high-frequency inference for the leverage effect when the spot variance follows rough Volterra dynamics, leverage loadings depend on endogenous rough states, and observed prices may be contaminated by microstructure noise. We define an integrated leverage parameter through a scaled covariation between log prices and spot variance that accounts for their different local fluctuation scales. The scaled realized covariation admits a stable functional central limit theorem with an explicit pathwise centering correction arising from within-interval variation in the leverage loading. A loading regularity condition renders this correction asymptotically negligible, and we verify the condition for direct Volterra loadings and for smooth functions of closed rough states for every Hurst index below one half. We construct a block-based estimator with an exact finite-window correction for the leading proportional bias, together with a nonnegative single-scale estimator of its asymptotic variance. Pre-averaging extends feasible inference to observations with independent and identically distributed microstructure noise. Simulations under the rough Heston model demonstrate accurate finite-sample calibration with and without noise. An empirical application to fifteen years of one-second data on nine U.S. equities and ETFs finds negative median leverage estimates for every asset at weekly and monthly horizons.

stat.ME↗
arXiv2026

Quantiles, Not Points: Streaming Quantile Matching for Ordinal-Aware LLM-as-a-Judge Calibration

LLM-as-a-judge scores are often distributionally shifted from human ratings on any given evaluation task, and post-hoc calibration (mapping LLM to human scores in a small labeled anchor set) has become a popular method to correct for this shift. In this paper we derive the calibration from two positions: that LLMaJ scores are ordinal ratings drawn from a distribution rather than isolated points. We present streaming quantile matching (SQM), a version of a classical method not previously bench marked for LLMaJ calibration, which is built on the two positions by computing the optimal-transport map between the LLM and human score distributions. We develop an evaluation framework and within it we show that unlike point-wise calibrations, SQM respects the ordinality of the human anchor set and achieves the best distributional fidelity and the top overall rank across models and data. We also show that in general, post-hoc calibration narrows the raw error gap between smaller and larger judge models, which opens the possibility of more efficient LLM judges.

cs.AI↗
arXiv2026

Prompt Lottery in LLM-as-a-Judge: Measuring and Mitigating Sensitivity to Syntactic Rephrasing

LLM-as-a-Judge has become a standard for evaluating large language models, yet most approaches implicitly assume that a single judge prompt functions as a stable evaluation instrument. Prior work has primarily explored deliberate prompt modifications such as scoring rubrics, role descriptions, or chain-of-thought reasoning in order to optimize accuracy. In contrast, we investigate whether validated prompt variants intended to preserve evaluation semantics while differing in surface realization (\textit{paraphrastic rephrasings}) yield consistent judgments under a fixed evaluation task. We introduce \textit{Prompt Lottery}, a phenomenon in which paraphrastic rephrasings produce different LLMaJ judgments. Across a large-scale study spanning diverse evaluation tasks, judge families, and model backbones, we find that these variations can alter evaluation outcomes and, in some settings, reverse relative model rankings. A user study with 10 human annotators shows that LLMaJ judges exhibit 6-9$\times$ larger agreement degradation under prompt rephrasings than humans, falling below the generally accepted threshold in usability studies. This indicates that single-prompt evaluation is insufficiently robust for model ranking. Finally, we show that multi-prompt aggregation substantially improves robustness by emphasizing judgments that remain consistent across prompt variants.

cs.CL↗
arXiv2026

SWE-Clar: Can Coding Agents Clarify Incomplete Client Requirements and Deliver Greenfield Software?

Greenfield software development in freelance settings requires coding agents to turn incomplete client requests into working software. This makes requirement clarification essential, yet clients may be unable to provide continuous feedback over long development cycles, and receiving their answers does not guarantee correct understanding of complex requirements. These challenges motivate two benchmark design choices: clarification precedes code implementation, and its effectiveness is assessed through explicit tests of requirement understanding. Guided by these choices, we introduce SWE-Clar, a benchmark grounded in real, high-value client commissions to evaluate requirement clarification and code implementation in this setting. It comprises 100 tasks based on real software development jobs on Upwork, each with at least US\$4,000 in total payments. Each task begins with multi-round interaction between a coding agent and a client agent, followed by separate runs for code implementation with acceptance testing and clarification assessment. For the latter, we construct 512 questions paired with reference answers from real client-freelancer interactions, covering both explicitly stated requirements and those inferred from the conversational context. Both runs begin from the same clarification context but remain isolated, preventing assessment questions from guiding code implementation. Experiments with nine frontier models show that even GPT-6 Astra, which has the highest observed implementation score, achieves a task pass rate of only 67%, underscoring the challenge of independent software delivery in this setting. The dataset and evaluation code are available at https://github.com/upwork/SWE-Clar.

cs.SE↗
arXiv2026

InsightRAG: Aggregating Distributed Scholarly Judgments for Scientific Survey Generation

The rapid growth of scientific literature makes survey writing increasingly challenging, motivating automatic survey generation. Yet, existing systems must search an open space of papers and an open space of concepts with no trace of how the community itself has navigated them, leading to incomplete scope, weak taxonomies, and unreliable citations. Our core insight is that author-written Related Work sections contain the human judgments needed to address both limitations: authors curate the citations relevant to their contributions and explain how the cited works relate, yielding paths for scope reconstruction and relational evidence for organization. We propose InsightRAG. Its insight-guided retrieval follows author-supported citation relations to reconstruct a verified research scope and structure the resulting relational evidence. It aggregates these judgments under a minimum-conflict objective into a taxonomy that reflects recurring author-written perspectives in the retrieved literature, guides generation, and bounds each section's evidence. Results on ten topics show gains in scope reconstruction, taxonomy quality, citation support, and reference reliability.

cs.AI↗
arXiv2026

QuAD: Quantitative Assessment via Distribution-Alignment for Qualitative Data Quality Enabled by a Content-Aware Multi-Agent System

Assessing whether qualitative responses collected from semi-structured questions are ready for a particular analysis goal is a persistent challenge. Manual qualitative analysis is interpretive and labor-intensive, while existing automated methods often lack a dataset-level perspective for deciding whether a response set supports a researcher-selected analytical lens. We introduce QuAD, a framework for quantitative pre-analysis assessment of qualitative response datasets. QuAD compares a reconstructed intrinsic cluster distribution reconstructed from responses with an extrinsic, researcher-defined analytical lens derived from metadata. QuAD is structured into a four-quadrant assessment whose metrics are read jointly, from lens-level relations to within-cluster variation, to provide fitness-for-purpose signals for revealing general patterns and unique insights. Because this assessment relies on lens-conditioned inter-response dissimilarity, we propose a content-aware multi-agent system with two-stage instruction generation and adaptive anchors to improve scoring consistency. We validate QuAD through controlled experiments on synthetic data, including robustness and baseline comparisons, and a real-world social survey case study. The case study shows how QuAD quantitatively assesses whether a selected analysis axis is promising and also surfaces representative or unusual responses for human interpretation.

cs.AI↗
arXiv2026

CytoCrowd: A Multi-Annotator Benchmark Dataset for Cytology Image Analysis

High-quality annotated datasets are crucial for advancing machine learning in medical image analysis. However, a critical gap exists: most datasets either offer a single, clean ground truth, which hides real-world expert disagreement, or they provide multiple annotations without a separate gold standard for objective evaluation. To bridge this gap, we introduce CytoCrowd, a new public benchmark for cytology analysis. The dataset features 446 high-resolution images, each with two key components: (1) raw, conflicting annotations from four independent pathologists, and (2) a separate, high-quality gold-standard ground truth established by a senior expert. This dual structure makes CytoCrowd a versatile resource. It serves as a benchmark for standard computer vision tasks, such as object detection and classification, using the ground truth. Simultaneously, it provides a realistic testbed for evaluating annotation aggregation algorithms that must resolve expert disagreements. We provide comprehensive baseline results for both tasks. Our experiments demonstrate the challenges presented by CytoCrowd and establish its value as a resource for developing the next generation of models for medical image analysis. The dataset is publicly available at https://github.com/YHSI5358/CytoCrowd-Datasets.

cs.CV↗
arXiv2026

Improved Cosmological Constraints from Morphology-Based Marked Correlation Functions

The cosmic web contains morphology-dependent information that is not fully captured by standard two-point statistics. We construct morphology-based marked correlation functions (MCFs) by assigning marks to halos according to the cosmic-web morphology identified with the \textsc{Nexus} algorithm. Using the \textsc{Kun} simulation suite, which spans 129 $w_0w_a$CDM cosmologies, we build Gaussian-process emulators for the MCFs as functions of cosmological parameters and tracer bias. We then apply the emulators to mock halo catalogues from the independent \textsc{Jiutian} simulation and perform a joint likelihood analysis to quantify the resulting cosmological constraints. We consider two marker choices: a discrete morphology marker and a continuous morphology strength marker. The continuous marker improves the Figure of Merit (FoM) by a factor of $\sim 8.6$ relative to the standard 2PCF and reduces the $1σ$ uncertainty on $σ_8$ by a factor of $\sim 5$. The discrete marker gives a more modest FoM improvement of $\sim 17\%$. We further test the impact of tracer selection by varying the halo mass threshold by a factor of $\sim 4.5$. Even for the lowest mass threshold, the continuous marker remains unbiased and achieves a FoM about $\sim 3.4$ times higher than that of the 2PCF alone. These results show that morphology-based MCFs, combined with simulation-based emulation, provide a useful framework for extracting additional cosmological information from large-scale structure surveys.

astro-ph.CO↗
arXiv2026

Wasserstein Filtering: A Sample Selection Method for Robust Distribution Learning

Given a dataset where a portion of the samples are contaminated, our goal is to recover the underlying clean population distribution. To this end, we propose Wasserstein Filtering (WF), a novel sample selection framework that discards a fraction of suspicious samples and estimates the target distribution using the empirical measure of the remaining data. The core insight is to select a subset of samples whose empirical distribution maximizes its Wasserstein distance to the fully contaminated empirical distribution, thereby preferentially isolating and removing geometrically influential outliers. To render this optimization computationally tractable, we introduce three algorithms: a marginal screening scheme, SinkMarg, and two joint optimization algorithms, SinkWF and SlicedWF, leveraging entropic optimal transport and sliced Wasserstein approximations, respectively. On the theoretical front, we introduce the Far Exclusion and Local Projection (FELP) contamination model, which characterizes corruptions consisting of well-separated outliers and locally indistinguishable perturbations. Under this model, we prove that the WF estimator achieves minimax optimality over distribution families with bounded covariance. Extensive numerical experiments on synthetic datasets, benchmark anomaly detection suites, and robust generative learning with diffusion models demonstrate that WF serves as a highly practical, model-agnostic preprocessing tool. It delivers competitive outlier detection performance and provides substantial downstream benefits for generative modeling under heavy contamination.

stat.ML↗
arXiv2026

Extending CSST Emulator to post-DESI era

The recent DESI BAO measurements have revealed a potential deviation from a cosmological constant, suggesting a dynamic nature of dark energy. To rigorously test this result, complementary probes such as weak gravitational lensing are crucial, demanding highly accurate and efficient predictions of the nonlinear matter power spectrum within the $w_0w_a$CDM framework. However, most existing emulators fail to cover the full parameter posterior from DESI DR2+CMB constraints in the $w_0\mbox{-}w_a$ plane. In this work, we extend the spectral equivalence method outlined in Casarini et al. 2016 to use auxiliary $w_0w_a$CDM models for approximating the power spectrum of a target $w_0w_a$CDM cosmology, moving beyond the previous use of $w$CDM auxiliaries. Incorporating this enhanced module, the extended CSST Emulator achieves a prediction accuracy of $\leq1\%$ over the $1σ$ confidence region from DESI DR2+CMB constraints for $z\leq3$, with a mild degradation in accuracy outside this posterior region. This performance is rigorously validated by additional simulations of dynamic dark energy cosmologies. The emulator's applicable parameter space has been generalized to fully encompass the $2σ$ region, greatly enhancing its utility for cosmological analysis in the post-DESI era.

astro-ph.CO↗
arXiv2026

Evaluating Language Models in Realistic Conversational Contexts

As Large Language Models (LLMs) are increasingly deployed to serve open-ended, multi-turn interactions, evaluating conversational quality at human scale has become a central challenge. Existing evaluation frameworks built for summarization, translation, or short-form QA tasks fall short of adequately measuring the consistency of human-scale dialogue, especially when derivation and validation of these metrics themselves often rely on synthetic rather than human sources. We fill the gap by introducing UPHELD (UPwork Human-Scale Evaluated Long Dialogues), a large, reference-full benchmark for evaluating human-scale conversational ability beyond factual correctness. UPHELD consists of hundreds of complete human-to-human dialogues authored by professional script writers, with realistic turn densities and 36,000+ per-turn human annotations across 30,000+ expert-generated dialogue turns. Using UPHELD, we systematically evaluate classical automatic metrics and reference-free LLM-as-a-judge approaches, and find them unreliable when correlated with expert human judgment. Building off this analysis, we use UPHELD to develop a Mixture-of-Judges framework that combines multiple evaluative signals and improves correlation with human assessments by approximately 30%. Overall, UPHELD provides a robust, human-grounded foundation for evaluating human-scale conversational intelligence that fills a crucial gap in the pre-existing LLM dataset landscape.

cs.CL↗
arXiv2026

ELUCID-DESI I: A Parallel MPI Implementation of the Initial Condition Solver for Large-Scale Reconstruction Simulations

We present a highly scalable, MPI-parallelized framework for reconstructing the initial cosmic density field, designed to meet the computational demands of next-generation cosmological simulations, particularly the upcoming ELUCID-DESI simulation based on DESI BGS data. Building upon the Hamiltonian Monte Carlo approach and the FastPM solver, our code employs domain decomposition to efficiently distribute memory between nodes. Although communication overhead increases the per-step runtime of the MPI version by roughly a factor of eight relative to the shared-memory implementation, our scaling tests-spanning different particle numbers, core counts, and node layouts-show nearly linear scaling with respect to both the number of particles and the number of CPU cores. Furthermore, to significantly reduce computational costs during the initial burn-in phase, we introduce a novel ``guess'' module that rapidly generates a high-quality initial density field. The results of the simulation test confirm substantial efficiency gains: for $256^3$ particles, 53 steps ($\sim$ 54 core hours) are saved, accelerating convergence by a factor of $\sim$ 18; for $1024^3$, 106 steps ($\sim$7500 core hours), achieving a speedup factor of $\sim$ 3. The total core hour gain grows with the number of particles, rendering large-volume reconstructions computationally practical for upcoming surveys, including our planned ELUCID-DESI reconstruction simulation with $4096^3$ particles. We estimate that achieving convergence for this scale (targeting DESI-BGS data) requires about 800 HMCMC steps ($\sim$ 5 million core hours). Our initial guess module will save approximately 360 steps ($\sim$2.3 million core hours), reducing the total computational time by about 45\%.

astro-ph.GA↗
arXiv2026

Cosmological constraints from neighbor-density-weighted marked correlation functions

We investigate whether neighbor-density-weighted marked correlation functions (MCFs) can extract cosmological information beyond the standard redshift-space two-point correlation function (2PCF). Using the Kun suite of 129 $w_0w_a$CDM$+\sum m_ν$ simulations in $1~h^{-1}{\rm Gpc}$ boxes, we construct Gaussian-process emulators for the normalized scale statistic $\widehat{W}^α(s)$ and the angular statistic $\widehat{W}^α_{Δs}(μ)$. We perform joint analyses combining multiple mark parameters $α$ and quantify the information gain using the FoM in the $Ω_m$--$σ_8$ plane. Relative to the 2PCF case, three-mark combinations improve the FoM by factors of $1.7$--$2.5$, while five-mark combinations increase the gain to $1.9$--$2.4$, depending on the statistic and mark definition. We further compare density and normalized-gradient marks, finding that they are nearly redundant for isotropic statistics but complementary for angular statistics, where their combination improves the FoM by up to $43\%$. Tests of scale range and halo selection show that the marked statistics remain robust under changes in analysis choices, with the angular statistic retaining additional cosmological information that is less sensitive to tracer selection. Our results demonstrate that MCFs substantially enhance cosmological constraints beyond the standard 2PCF and provide a robust probe for next-generation galaxy surveys.

astro-ph.CO↗
arXiv2026

Cosmological Constraints from Bias-Robust Wavelet Scattering Statistics for Stage-IV Galaxy Surveys

A central challenge in precision cosmology with galaxy surveys is to extract non-Gaussian information from large-scale structure while controlling systematic uncertainties such as tracer bias. Conventional clustering statistics, such as the two-point correlation function (2PCF), capture limited nonlinear information and typically require explicit bias modeling, which can introduce systematic errors if the adopted bias prescription is inaccurate. To address this problem, we introduce $R^{\rm wst}$, a bias-robust statistic constructed from $m$-mode ratios of the wavelet scattering transform (WST). Using simulation-based inference, we train a Gaussian-process-regression emulator on the \texttt{Kun} simulation suite and use \texttt{JiuTian} simulations for covariance estimation and validation. The emulator achieves percent-level accuracy, sufficient for the expected observational uncertainties. We show that $R^{\rm wst}$ yields unbiased constraints on $Ω_m$, $σ_8$, $n_s$, and $w_0$, and improves the breaking of the $Ω_m$--$σ_8$ degeneracy by about a factor of two compared with 2PCF. Its constraining power remains stable across a broad range of tracer-bias scenarios, demonstrating that $R^{\rm wst}$ can mitigate bias-induced systematics without explicit bias modeling. These results establish $R^{\rm wst}$ as a powerful and robust statistic for precision cosmology with Stage-IV surveys.

astro-ph.CO↗
arXiv2026

The first AKRA mass map reconstruction from HSC Y1 data

Weak lensing mass-mapping from shear catalogs faces systematic challenges from survey masks and spatially varying noise. To overcome these issues and reconstruct unbiased convergence $κ$ maps, we have constructed the AKRA (Accurate Kappa Reconstruction Algorithm), a prior-free and maximum-likelihood based analytical method. It has been validated for mock shear catalogs with a variety of survey masks. In this work, we present the first real-data application of the AKRA on the Subaru Hyper Suprime-Cam Year 1 (HSC Y1) data. We first validate AKRA using mock shear catalogs from the \texttt{Kun} simulation suite, with masks corresponding to the six HSC Y1 regions (\texttt{GAMA09H}, \texttt{GAMA15H}, \texttt{HECTOMAP}, \texttt{VVDS}, \texttt{WIDE12H}, and \texttt{XMMLSS}). The investigated statistics, including the lensing power spectrum, $\langle κ^2\rangle$, $\langle κ^3\rangle$, and the one-point probability distribution function of $κ$, are all unbiased. We then apply AKRA to the HSC Y1 shear catalog and provide reconstructed $κ$ maps ready for subsequent scientific analyses.

astro-ph.CO↗