arXiv subjects

Qingqing Mao

Publications and source records attributed to Qingqing Mao.

At least 19 recordsLinked to original sources

arXiv2026

Train What You Deploy: Closing the MLP Reachability Gap in Low-Rank Clone Distillation

A compressed student has two shapes that need not agree: the weight it deploys at inference and the weight family its training can reach. We show that a state-of-the-art weight-inheritance distiller, Low-Rank Clone (LRC), deploys a full-width student MLP but ties training to a teacher-induced slice, leaving 62.5-81.4% of each deployed matrix's independent linear degrees of freedom unreachable-paid for at inference, never trainable. Our principle is one line: train what you deploy. From the identical LRC warm start, we make the training object the entire deployed matrix, with no change in deployed shape, deployed parameter count, or inference FLOPs, via two mergeable realizations (Dense-LRC and CORE-LRC) that both collapse to one deployed weight. This recovers stranded capacity: taking the stronger realization per teacher, +2.36/+2.71/+10.45 Avg9 over matched-budget plain-LRC baselines across three teachers (Llama3.2-3B, Llama3.1-8B, Qwen2.5-3B), with the largest gain on the widest teacher (Qwen), where it reaches the original recipe's approx. 20B-token accuracy at 10B tokens (2x token efficiency); there the strictly same-lineage arm still recovers +6.39, the fully controlled figure. Controls strongly support attributing the gain to the enlarged reachable set, rather than to added parameters or the recipe. From approx. 10B distillation tokens plus a short SFT, a half-parameter 1.5B student matches its approx. 9T-token teacher's 9-task macro-average, within evaluation noise and with a residual MMLU deficit, and a 2.7B student beats Meta's own official compression of Llama3.1-8B at ~900x fewer compression tokens (a token count under unmatched recipes, not a compute claim). All results are from single-seed runs on the LRC backbone.

cs.LG↗

arXiv2026

Reconstruction: A Blind Benchmark for Recovering Research Ideas from Pre-Publication Bibliographies

Can a language model recover the true research idea of a published paper when given only that paper's pre-publication bibliography? We introduce Reconstruction, a blind idea-recovery benchmark that withholds the seed paper and all contemporaneous or future literature, and asks models to propose hypotheses that an independent large language model judge matches against the held-out ground-truth idea. A strict anti-leakage protocol-temporal citation cutoff, anonymous reference IDs, and frozen per-paper bibliographies, which prevents prompt-time leakage of the seed idea. Across six scientific domains and 643 evaluated papers, seven frontier models achieve only modest Match rates (approx. 3-15%). We then evaluate a reference-only multi-agent (top 4) pipeline that combines cross-model review with a Swiss tournament over aligned hypothesis slots, without external web search. Cross-model review plus tournament selection raises Match rates to approx. 23-42% across all six domains, which is an observed approx. 2.4x lift over the best single-model baseline. This draft reports the protocol, anti-leakage design, and current results as an arXiv timestamp.

cs.AI↗

arXiv2026

How Do Agents Fail on AutoResearch: End-to-End Diagnostic Evaluation on 100 Real-World Frontier Research Tasks

AI has long assisted scientific research, but the rapid advance of LLMs and agentic scaffolds is reshaping the landscape; a single system can now carry whole-stage research from an initial hypothesis all the way to final published paper, which is a paradigm now referred to as AutoResearch. Existing evaluations reveal little about how these agents operate or where they break down. Tasks are narrowly-scoped, evaluation measures performance but not process, and failure diagnoses lack systematic coverage or artifact-level visibility. To address this gap, we introduce AutoResearchEval, featuring 100 tasks grounded in published frontier science across 7 scientific domains and the full research lifecycle, including ideation, retrieval, execution, analysis, writing, and review. Evaluating 8 harness-model combinations yields 800 autoresearch agent trajectories, with process-level annotation. We organize these insights into AutoResearch Failure Taxonomy or ARFT, a framework of 45 empirically-grounded failure patterns. To enable scalable fine-grained attribution, we leverage a human-calibrated agent-as-a-judge pipeline to inspect complete trajectories and intermediate artifacts. Failure patterns converge on a single overarching limitation, namely that current agents lack a metacognitive loop, which entails the ability to check what they produced against what they found, revise when it does not hold up, and question whether the path they took was sound. The same patterns recur across all 8 harness-model combinations, including the strongest models tested, locating the deficit at the model level rather than in any particular scaffold; whether orchestration-level interventions can close it is an open question this work does not test. We publicly release AutoResearchEval and ARFT to facilitate continued research and development in autonomous scientific discovery.

cs.CL↗

arXiv2026

Reaction-Network-Level Discovery of Ammonia Synthesis Catalysts via Ten-Million-Scale Generative Exploration

Catalyst discovery for ammonia synthesis is inherently a reaction-network challenge because catalytic performance is governed not by a single adsorbed intermediate, but by a surface's orchestrated compatibility with multiple distinct intermediates across competing dissociative and associative pathways. However, navigating ultra-large chemical spaces under such multi-intermediate constraints remains a formidable bottleneck for conventional screening workflows. Here, we report a reaction-network-level catalyst discovery framework driven by ten-million-scale generative exploration. By coupling adsorbate-specific generative Transformers with high-throughput machine learning potentials, we systematically map the structure-property landscapes of four critical intermediates (N*, NH*, NNH*, and HNNH*). Scale-dependent overlap analysis shows that the full four-intermediate compatibility space remains strongly under-sampled at conventional 105-106 generative scales, emerging exclusively under ten-million-scale exploration. By generating approximately 15 million configurations per adsorbate, followed by structural compression and machine-learning-potential predictions, we identified 279 highly potential target materials. This sparse compatibility space successfully recovers traditional Fe- and Ru-based motifs while uncovering previously unexplored catalyst families. Representative DFT calculations validate pathway-dependent mechanisms: Fe-V emerges as a dissociative-pathway lead by significantly lowering the initial N2 dissociation barrier, whereas Al-Pd-Zr efficiently stabilizes associative intermediates as an associative-pathway lead. These findings establish multi-intermediate reaction-network compatibility as a robust criterion for discovering advanced catalysts from multi-million generative chemical spaces.

physics.chem-ph↗

arXiv2026

CLaaS: Continual learning as a service for sample efficient online learning

Deployed large language model agents must adapt to distribution shift in dynamic environments. Ideally, adaptation can be performed from accumulated agent experiences and retain prior capabilities while transferring to future tasks. However, agent actions and environmental transitions can only be sampled once per scenario, as real-world environments cannot be trivially reset. To this end, we investigate an experiential and online continual learning setting in which agents learn from a stream of scenarios. We propose continual learning as-a-service (CLaaS), a system which enables agents to improve during deployment, abstracted behind a chat API. To increase sample efficiency, CLaaS stores rollouts in an experience replay buffer for gradient reuse during asynchronous training. We evaluate CLaaS on an adversarial task, demonstrating that parametric updates lead to superior forward transfer and less forgetting than in-context learning, with replay being a critical choice for sample efficiency.

cs.LG↗

arXiv2026

AdaDPO: Self-Adaptive Direct Preference Optimization with Balanced Gradient Updates

DPO has become a widely adopted alternative to RLHF for aligning LLMs with human preferences, eliminating the need for a separate reward model or RL loop. Recent theoretical analysis uncovers an asymmetric gradient behavior in DPO: the loss suppresses dispreferred responses substantially faster than it promotes preferred ones, causing the model to learn to avoid bad answers rather than to generate good ones. We propose AdaDPO, a Self-Adaptive variant of the DPO algorithm that introduces per-preference-pair, stop-gradient-based coefficients derived directly from the policy model's generation probabilities, with the reference model's probabilities as an optional component. AdaDPO is constructed to enforce equality of gradient magnitudes between preferred and dispreferred probabilities; the practical implementation balances per-token gradients and applies a numerical clipping bound for stability, while retaining DPO's original hyperparameter structure. On Llama-3-8B-Instruct trained on UltraFeedback under a SimPO similar setup, AdaDPO consistently outperforms DPO on AlpacaEval 2: it achieves higher length-controlled win rates (LC) in 81% of hyperparameter combinations, attains the global best LC (48.3%) and raw win rate (46.1%), and enlarges the LC-over-WR margin in 88% of combinations, indicating effective mitigation of length bias. Additional analyses on KL divergence, reward margin, and reward accuracy confirm that AdaDPO rectifies the gradient imbalance and yields more efficient optimization. Because it operates purely at the loss level, AdaDPO can be dropped into existing preference-based alignment pipelines without changing data collection or model architectures. The method requires only a few lines of code, and the same self-adaptive principle generalizes to a broad family of pairwise contrastive preference losses including SimPO, R-DPO, IPO, CPO, and ORPO.

cs.CL↗

arXiv2026

Interpretability-Guided Layer Selection over Subspace Projection: SAEs as Stethoscopes, Not Scalpels, for Raw Task Vector Model Editing

LLMs increasingly require surgical model editing to enhance domain-specific capabilities without incurring the computational cost or catastrophic forgetting associated with full fine-tuning. Sparse Autoencoders (SAEs) have emerged as a promising tool in this setting, in principle allowing for feature-level identification of where to intervene. In this work, we rigorously evaluate an SAE-guided editing pipeline for mathematical reasoning on Gemma-3-4B-IT and uncover a fundamental failure mode: the intuitively appealing approach of projecting task vectors onto SAE feature subspaces acts as an information bottleneck that discards approximately 97% of the modification energy, yielding no statistically significant improvements across seven math subjects. We show that this failure stems from a geometric misalignment between activation-space SAE directions and weight-space task vectors. We then propose a shift in perspective: SAE as a Stethoscope, Not a Scalpel, where SAEs are used for layer-level diagnosis rather than intervention-level filtering. By injecting unfiltered raw task vectors only into layers identified by an SAE-derived specificity score, we improve Number Theory accuracy from 29.6% to 39.4% (z=+3.41, p=0.0007) on the Minerva Math benchmark; 5 of 7 math subjects significantly improved and none significantly degraded. Our method is fully deterministic, requires no additional inference cost, and provides a principled framework for interpretability-guided model editing.

cs.LG↗

arXiv2026

State-of-the-Art Arabic Language Modeling with Sparse MoE Fine-Tuning and Chain-of-Thought Distillation

This paper introduces Arabic-DeepSeek-R1, an application-driven open-source Arabic LLM that leverages a sparse MoE backbone to address the digital equity gap for under-represented languages, and establishes a new SOTA across the entire Open Arabic LLM Leaderboard (OALL). Our four-phase CoT distillation scheme integrates Arabic-specific linguistic verification and regional ethical norms into a 372M-token, contamination-controlled 80/20 Arabic-English training mixture. Arabic-DeepSeek-R1 achieves the highest average score across the seven-benchmark OALL suite while establishing SOTA or near-SOTA, including dominant results on grammar-focused MadinahQA (surpassing both GPT-5.1 and the OALL leader by substantial margins), safety-oriented AraTrust, multi-ability AlGhafa, and retrieval-augmented ALRAGE. Our results indicate that the combination of sparse MoE architecture, culturally-informed CoT distillation with explicit Arabic linguistic checks, and strategic bilingual data curation enables an open-source adapted model to systematically outperform the proprietary frontier system GPT-5.1 on the majority of benchmarks evaluating comprehensive language-specific tasks: the first such demonstration for Arabic LLMs. These findings indicate that much of Arabic's performance deficit in current LLM ecosystems stems from under-specialization rather than architectural limitations, and that parameter-efficient adaptation of open reasoning models can yield breakthrough SOTA performance without industrial-scale pretraining costs. Arabic-DeepSeek-R1 establishes a validated and replicable framework for sovereign and domain-specific language technologies, demonstrating that strategic, culturally-grounded adaptation of sparse MoE backbones offers a viable and cost-effective pathway to achieving record-breaking performance across standardized benchmarks for low-resource languages.

cs.CL↗

arXiv2026

Application-Driven Pedagogical Knowledge Optimization of Open-Source LLMs via Reinforcement Learning and Supervised Fine-Tuning

We present an innovative multi-stage optimization strategy combining reinforcement learning (RL) and supervised fine-tuning (SFT) to enhance the pedagogical knowledge of large language models (LLMs), as illustrated by EduQwen 32B-RL1, EduQwen 32B-SFT, and an optional third-stage model EduQwen 32B-SFT-RL2: (1) RL optimization that implements progressive difficulty training, focuses on challenging examples, and employs extended reasoning rollouts; (2) a subsequent SFT phase that leverages the RL-trained model to synthesize high-quality training data with difficulty-weighted sampling; and (3) an optional second round of RL optimization. EduQwen 32B-RL1, EduQwen 32B-SFT, and EduQwen 32B-SFT-RL2 are an application-driven family of open-source pedagogical LLMs built on a dense Qwen3-32B backbone. These models remarkably achieve high enough accuracy on the Cross-Domain Pedagogical Knowledge (CDPK) Benchmark to establish new state-of-the-art (SOTA) results across the interactive Pedagogy Benchmark Leaderboard and surpass significantly larger proprietary systems such as the previous benchmark leader Gemini-3 Pro. These dense 32-billion-parameter models demonstrate that domain-specialized optimization can transform mid-sized open-source LLMs into true pedagogical domain experts that outperform much larger general-purpose systems, while preserving the transparency, customizability, and cost-efficiency required for responsible educational AI deployment.

cs.CL↗

arXiv2024

OpenMedLM: Prompt engineering can out-perform fine-tuning in medical question-answering with open-source large language models

LLMs have become increasingly capable at accomplishing a range of specialized-tasks and can be utilized to expand equitable access to medical knowledge. Most medical LLMs have involved extensive fine-tuning, leveraging specialized medical data and significant, thus costly, amounts of computational power. Many of the top performing LLMs are proprietary and their access is limited to very few research groups. However, open-source (OS) models represent a key area of growth for medical LLMs due to significant improvements in performance and an inherent ability to provide the transparency and compliance required in healthcare. We present OpenMedLM, a prompting platform which delivers state-of-the-art (SOTA) performance for OS LLMs on medical benchmarks. We evaluated a range of OS foundation LLMs (7B-70B) on four medical benchmarks (MedQA, MedMCQA, PubMedQA, MMLU medical-subset). We employed a series of prompting strategies, including zero-shot, few-shot, chain-of-thought (random selection and kNN selection), and ensemble/self-consistency voting. We found that OpenMedLM delivers OS SOTA results on three common medical LLM benchmarks, surpassing the previous best performing OS models that leveraged computationally costly extensive fine-tuning. The model delivers a 72.6% accuracy on the MedQA benchmark, outperforming the previous SOTA by 2.4%, and achieves 81.7% accuracy on the MMLU medical-subset, establishing itself as the first OS LLM to surpass 80% accuracy on this benchmark. Our results highlight medical-specific emergent properties in OS LLMs which have not yet been documented to date elsewhere, and showcase the benefits of further leveraging prompt engineering to improve the performance of accessible LLMs for medical applications.

cs.CL↗

arXiv2021

Optimal discharge of patients from intensive care via a data-driven policy learning framework

Clinical decision support tools rooted in machine learning and optimization can provide significant value to healthcare providers, including through better management of intensive care units. In particular, it is important that the patient discharge task addresses the nuanced trade-off between decreasing a patient's length of stay (and associated hospitalization costs) and the risk of readmission or even death following the discharge decision. This work introduces an end-to-end general framework for capturing this trade-off to recommend optimal discharge timing decisions given a patient's electronic health records. A data-driven approach is used to derive a parsimonious, discrete state space representation that captures a patient's physiological condition. Based on this model and a given cost function, an infinite-horizon discounted Markov decision process is formulated and solved numerically to compute an optimal discharge policy, whose value is assessed using off-policy evaluation strategies. Extensive numerical experiments are performed to validate the proposed framework using real-life intensive care unit patient data.

cs.LG↗

arXiv2016

The lensing and temperature imprints of voids on the Cosmic Microwave Background

We have searched for the signature of cosmic voids in the CMB, in both the Planck temperature and lensing-convergence maps; voids should give decrements in both. We use zobov voids from the DR12 SDSS CMASS galaxy sample. We base our analysis on N-body simulations, to avoid a posteriori bias. For the first time, we detect the signature of voids in CMB lensing: the significance is $3.2σ$, close to $Λ$CDM in both amplitude and projected density-profile shape. A temperature dip is also seen, at modest significance ($2.3σ$), with amplitude about 6 times the prediction. This temperature signal is induced mostly by voids with radius between 100 and 150 Mpc/h, while the lensing signal is mostly contributed by smaller voids -- as expected; lensing relates directly to density, while ISW depends on gravitational potential. The void abundance in observations and simulations agree, as well. We also repeated the analysis excluding lower-significance voids: no lensing signal is detected, with an upper limit of about twice the $Λ$CDM prediction. But the mean temperature decrement now becomes non-zero at the $3.7σ$ level (similar to that found by Granett et al.), with amplitude about 20 times the prediction. However, the observed dependence of temperature on void size is in poor agreement with simulations, whereas the lensing results are consistent with $Λ$CDM theory. Thus, the overall tension between theory and observations does not favour non-standard theories of gravity, despite the hints of an enhanced amplitude for the ISW effect from voids.

astro-ph.CO↗

arXiv2016

The Astropy Problem

The Astropy Project (http://astropy.org) is, in its own words, "a community effort to develop a single core package for Astronomy in Python and foster interoperability between Python astronomy packages." For five years this project has been managed, written, and operated as a grassroots, self-organized, almost entirely volunteer effort while the software is used by the majority of the astronomical community. Despite this, the project has always been and remains to this day effectively unfunded. Further, contributors receive little or no formal recognition for creating and supporting what is now critical software. This paper explores the problem in detail, outlines possible solutions to correct this, and presents a few suggestions on how to address the sustainability of general purpose astronomical software.

astro-ph.IM↗

arXiv2016

Cosmic Voids in the SDSS DR12 BOSS Galaxy Sample: The Alcock-Paczynski Test

We apply the Alcock-Paczynski (AP) test to the stacked voids identified using the large-scale structure galaxy catalog from the Baryon Oscillation Spectroscopic Survey (BOSS). This galaxy catalog is part of the Sloan Digital Sky Survey (SDSS) Data Release 12 and is the final catalog of SDSS-III. We also use 1000 mock galaxy catalogs that match the geometry, density, and clustering properties of the BOSS sample in order to characterize the statistical uncertainties of our measurements and take into account systematic errors such as redshift space distortions. For both BOSS data and mock catalogs, we use the ZOBOV algorithm to identify voids, we stack together all voids with effective radii of 30-100Mpc/h in the redshift range 0.43-0.7, and we accurately measure the shape of the stacked voids. Our tests with the mock catalogs show that we measure the stacked void ellipticity with a statistical precision of 2.6%. We find that the stacked voids in redshift space are slightly squashed along the line of sight, which is consistent with previous studies. We repeat this measurement of stacked void shape in the BOSS data assuming several values of Omega_m within the flat LCDM model, and we compare to the mock catalogs in redshift space in order to perform the AP test. We obtain a constraint of $Ω_m = 0.38^{+0.18}_{-0.15}$ at the 68% confidence level from the AP test. We discuss the various sources of statistical and systematic noise that affect the constraining power of this method. In particular, we find that the measured ellipticity of stacked voids scales more weakly with cosmology than the standard AP prediction, leading to significantly weaker constraints. We discuss how AP constraints will improve in future surveys with larger volumes and densities.

astro-ph.CO↗

arXiv2016

A Cosmic Void Catalog of SDSS DR12 BOSS Galaxies

We present a cosmic void catalog using the large-scale structure galaxy catalog from the Baryon Oscillation Spectroscopic Survey (BOSS). This galaxy catalog is part of the Sloan Digital Sky Survey (SDSS) Data Release 12 and is the final catalog of SDSS-III. We take into account the survey boundaries, masks, and angular and radial selection functions, and apply the ZOBOV void finding algorithm to the galaxy catalog. We identify a total of 10,643 voids. After making quality cuts to ensure that the voids represent real underdense regions, we obtain 1,228 voids with effective radii spanning the range 20-100Mpc/h and with central densities that are, on average, 30% of the mean sample density. We release versions of the catalogs both with and without quality cuts. We discuss the basic statistics of voids, such as their size and redshift distributions, and measure the radial density profile of the voids via a stacking technique. In addition, we construct mock void catalogs from 1000 mock galaxy catalogs, and find that the properties of BOSS voids are in good agreement with those in the mock catalogs. We compare the stellar mass distribution of galaxies living inside and outside of the voids, and find no significant difference. These BOSS and mock void catalogs are useful for a number of cosmological and galaxy environment studies.

astro-ph.CO↗

arXiv2015

KELT-4Ab: An inflated Hot Jupiter transiting the bright (V~10) component of a hierarchical triple

We report the discovery of KELT-4Ab, an inflated, transiting Hot Jupiter orbiting the brightest component of a hierarchical triple stellar system. The host star is an F star with $T_{\rm eff}=6206\pm75$ K, $\log g=4.108\pm0.014$, $\left[{\rm Fe}/{\rm H}\right]=-0.116_{-0.069}^{+0.065}$, ${\rm M_*}=1.201_{-0.061}^{+0.067} \ {\rm M}_{\odot}$, and ${\rm R_*}=1.610_{-0.068}^{+0.078} \ {\rm R}_{\odot}$. The best-fit linear ephemeris is $\rm {BJD_{TDB}} = 2456193.29157 \pm 0.00021 + E\left(2.9895936 \pm 0.0000048\right)$. With a magnitude of $V\sim10$, a planetary radius of $1.699_{-0.045}^{+0.046} \ {\rm R_J}$, and a mass of $0.902_{-0.059}^{+0.060} \ {\rm M_J}$, it is the brightest host among the population of inflated Hot Jupiters ($R_P > 1.5R_J$), making it a valuable discovery for probing the nature of inflated planets. In addition, its existence within a hierarchical triple and its proximity to Earth ($210$ pc) provides a unique opportunity for dynamical studies with continued monitoring with high resolution imaging and precision radial velocities. In particular, the motion of the binary stars around each other and of both stars around the primary star relative to the measured epoch in this work should be detectable when it rises in October 2015.

astro-ph.EP↗

arXiv2015

Probing Galactic Structure with the Spatial Correlation Function of SEGUE G-dwarf Stars

We measure the two-point correlation function of G-dwarf stars within 1-3 kpc of the Sun in multiple lines-of-sight using the Schlesinger et al. G-dwarf sample from the SDSS SEGUE survey. The shapes of the correlation functions along individual SEGUE lines-of-sight depend sensitively on both the stellar-density gradients and the survey geometry. We fit smooth disk galaxy models to our SEGUE clustering measurements, and obtain strong constraints on the thin- and thick-disk components of the Milky Way. Specifically, we constrain the values of the thin- and thick-disk scale heights with 3% and 2% precision, respectively, and the values of the thin- and thick-disk scale lengths with 20% and 8% precision, respectively. Moreover, we find that a two-disk model is unable to fully explain our clustering measurements, which exhibit an excess of clustering at small scales (< 50 pc). This suggests the presence of small-scale substructure in the disk system of the Milky Way.

astro-ph.GA↗

arXiv2014

VIDE: The Void IDentification and Examination toolkit

We present VIDE, the Void IDentification and Examination toolkit, an open-source Python/C++ code for finding cosmic voids in galaxy redshift surveys and N-body simulations, characterizing their properties, and providing a platform for more detailed analysis. At its core, VIDE uses a substantially enhanced version of ZOBOV (Neyinck 2008) to calculate a Voronoi tessellation for estimating the density field and a performing a watershed transform to construct voids. Additionally, VIDE provides significant functionality for both pre- and post-processing: for example, vide can work with volume- or magnitude-limited galaxy samples with arbitrary survey geometries, or dark matter particles or halo catalogs in a variety of common formats. It can also randomly subsample inputs and includes a Halo Occupation Distribution model for constructing mock galaxy populations. VIDE uses the watershed levels to place voids in a hierarchical tree, outputs a summary of void properties in plain ASCII, and provides a Python API to perform many analysis tasks, such as loading and manipulating void catalogs and particle members, filtering, plotting, computing clustering statistics, stacking, comparing catalogs, and fitting density profiles. While centered around ZOBOV, the toolkit is designed to be as modular as possible and accommodate other void finders. VIDE has been in development for several years and has already been used to produce a wealth of results, which we summarize in this work to highlight the capabilities of the toolkit. VIDE is publicly available at http://bitbucket.org/cosmicvoids/vide public and http://www.cosmicvoids.net.

astro-ph.CO↗