Searcharxiv⌕ Search

arXiv subjects

Shu Yang

Publications and source records attributed to Shu Yang.

At least 19 recordsLinked to original sources

Infrathon: Benchmarking Language Agents for Real-World, Long-Horizon AI Infra Optimization

In this work, we study whether frontier agents can continually improve real-world AI infrastructure over long horizons. We introduce Infrathon, a benchmark of diverse, long-horizon AI infrastructure optimization tasks across different levels. We evaluate six models across 150 runs with 24-hour budgets, totaling over 10,000 GPU-hours. Our results show that model rankings vary with the evaluation horizon, while 135 of the 150 runs achieve further improvements during the second half of their budgets. These findings indicate that shorter evaluations can overlook both optimization gains and changes in relative model performance. We also find agents can discover solutions that outperform human-designed baselines: Claude Opus 5.5 increases the mean KV-cache compression ratio from $6.82\times$ with the human-designed Quant-VideoGen codec to $19.86\times$. Analysis of agent trajectories identifies task violations, unreliable testing, and premature completion as barriers to sustained improvement. Together, our Infrathon framework provides resources for constructing long-horizon tasks, evaluating frontier agents, and discovering good infrastructure solutions. Code is available at https://github.com/InfrathonX/Infrathon

cs.AI↗

Multi-User Large Language Model Agents

Large language models (LLMs) and LLM-based agents are increasingly deployed as assistants in planning and decision making, yet most existing systems are implicitly optimized for a single-principal interaction paradigm, in which the model is designed to satisfy the objectives of one dominant user whose instructions are treated as the sole source of authority and utility. However, as they are integrated into team workflows and organizational tools, they are increasingly required to serve multiple users simultaneously, each with distinct roles, preferences, and authority levels, leading to multi-user, multi-principal settings with unavoidable conflicts, information asymmetry, and privacy constraints. In this work, we present the first systematic study of multi-user LLM agents. We begin by formalizing multi-user interaction with LLM agents as a multi-principal decision problem, where a single agent must account for multiple users with potentially conflicting interests and associated challenges. We then introduce a unified multi-user interaction protocol and design three targeted stress-testing scenarios to evaluate current LLMs' capabilities in instruction following, privacy preservation, and coordination. Our results reveal systematic gaps: frontier LLMs frequently fail to maintain stable prioritization under conflicting user objectives, exhibit increasing privacy violations over multi-turn interactions, and suffer from efficiency bottlenecks when coordination requires iterative information gathering.

cs.CL↗

Spectral-Domain Information Geometry of Stationary Covariance Models: Gaussian and Elliptical Extensions

This paper develops an information-geometric framework for stationary covariance structures. We begin with the infinite-dimensional setting, where stationary covariance structures can be characterized without finite-window boundary effects. By formulating the covariance structure as a bilinear form on a suitable class of test functions, we establish its spectral representation and show how finite-dimensional Toeplitz covariance matrices arise as restrictions to nested observation subspaces. For zero-mean proper complex Gaussian wide-sense stationary (WSS) processes, we show that the normalized finite-dimensional Fisher information converges to a spectral-domain metric. In logarithmic spectral coordinates, this metric becomes the standard $L^2$ inner product, yielding an explicit Fisher--Rao distance that is also recovered as the asymptotic normalized affine-invariant Riemannian distance between Toeplitz covariance matrices. We then extend the framework to elliptical observation models by separating global covariance scale from shape. The shape Fisher information is orthogonal to radial nuisance directions and is characterized by a dimension- and radial-model-dependent coefficient. We derive a condition under which this coefficient converges to its Gaussian value as the observation dimension tends to infinity. The resulting limiting spectral-shape geometry is the centered $L^2$ geometry of log-spectra and is invariant to positive rescaling. Numerical experiments illustrate the theoretical convergence results and the recovery of the Gaussian shape coefficient.

math.ST↗

Final Checkpoints Are Not Enough: Analyzing Latent Reasoning Faithfulness Along Training Trajectories

Latent reasoning performs multi-step inference in continuous hidden states, promising more compact and efficient reasoning. However, these opaque states raise a question of faithfulness: whether the latent reasoning steps drive the final answer. Prior work reports unfaithful behaviors at selected checkpoints, leaving their development during training unexamined. We track behavioral and activation-based evidence across training using verified counterfactual edits and interventions on the latent reasoning states. We find that accuracy can improve while models become less responsive to edits that change the required reasoning path and the correct answer. This counterfactual responsiveness follows distinct trajectories across latent reasoning approaches. On the graph reachability task ProsQA, output sensitivity to noise replacement declines alongside counterfactual responsiveness. Across separately trained binary-choice and open-ended versions of GSM8K math problems, interventions on all latent states show opposite sensitivity trajectories. Guided by these trajectories, we intervene on the training procedure and improve counterfactual responsiveness. These results show that evaluating only a final checkpoint can obscure both when counterfactual responsiveness changes and what the latent states contribute.

cs.LG↗

Sensitivity Analysis for Trials with Irregular and Informative Assessment Times: Marginal Mean Models and Sensitivity Parameter Selection

In randomized trials, outcome assessment times often vary across participants and may depend on the underlying outcome process, so outcomes observed near a target time need not represent the target population. Whether assessment is informative cannot be determined from the observed data alone. We extend the sensitivity analysis framework of Smith et al.\ (2024), which is anchored in an explainable assessment assumption and was restricted to continuous outcomes through an identity-link marginal mean model, to a broader class of link functions that accommodates binary and count outcomes. Within this framework, we develop an influence function-based augmented inverse intensity-weighted estimator that is root-$n$ consistent under standard regularity conditions. We also propose a data-driven procedure for selecting the range of the sensitivity parameter from plausible bounds on the target mean outcome and for constructing a confidence set for this range, which accounts for uncertainty in sensitivity parameter selection. We illustrate the method with data from the ARC study and evaluate its finite-sample performance in simulation studies.

stat.ME↗

Prediction-Powered Data Fusion for Treatment Effect Estimation

Randomized controlled trials (RCTs) identify treatment effects without confounding but are often small, whereas observational studies (OBS) are large but may be confounded. Many estimators combining a small RCT with a large OBS have been developed for the average treatment effect (ATE) and the conditional ATE (CATE). However, existing ATE estimators either make assumptions on the OBS or do not borrow enough power from them. The CATE has been studied less than the ATE. Existing CATE methods either assume the OBS are unconfounded, rely on a model of the confounding function, or accept bias in exchange for lower variance. We therefore propose a framework that, without special assumptions on the OBS, fuses the OBS and the RCT by preserving the unbiasedness of RCT-based estimation while borrowing power from the large OBS to boost precision. Applying this principle, we build an ATE estimator, AIPW-Fusion, with closed-form weights and confidence intervals, and two CATE learners, DR-Fusion and R-Fusion. Experiments corroborate our findings.

stat.ML↗

PocketGS: High-Fidelity On-Device Training for 3D Gaussian Splatting

Although 3D Gaussian Splatting (3DGS) supports real-time rendering, training relies on offline geometry, long-horizon densification, and memory-intensive differentiation. We cast post-capture static-scene training as resource-constrained finite-horizon optimization, where terminal quality depends jointly on the initial representation and reference-consistent gradient execution. To the best of our knowledge, PocketGS is the first publicly documented pipeline to complete geometry construction, Gaussian initialization, and 3DGS optimization on a commodity smartphone. Its I, G warm start combines refined poses, a filtered dense prior, and surface-aligned anisotropic Gaussians. Its T operator replays compositing interactions and scatters gradients to canonical parameters, matching a numerical reference at the reported precision. Controlled 500-step experiments on LLFF, Mip-NeRF 360, and MobileScan show high-quality reconstruction at the device horizon. Across MobileScan, every scene remains below 3 GB peak memory and post-capture processing averages 255.2 s, enabling an on-device capture-to-render workflow. Separate 30k experiments characterize native-schedule behavior.

cs.CV↗

Knowledge Acquisition During Pre-training? Large Language Models Learn Better With Auxiliary Views

Gaps remain in our understanding of how large language models (LLMs) acquire knowledge during pre-training. We find that auxiliary views, conceptual reformulations, are key to this process. We design controlled experiments to isolate this. First, holding the token budget fixed, allocating tokens from paraphrases to auxiliary views improves learning even for factual recall, which we use near-verbatim sentences from the original document to test. Second, the effectiveness of auxiliary views is not contingent on the strength of the model that generates them. Third, we establish that paraphrasing is helpful but only at smaller batch sizes. Fourth, we identify types of knowledge that aid learning in the presence of prior knowledge gaps. Finally, we examine how these effects manifest mechanistically via layer-wise biases and compression. Together, our findings suggest that auxiliary representations of knowledge are a key factor in the success of pre-training, and explains why data diversity matters. Further, our results provide clear recipes for synthetic data augmentation along with implications on the self-improvement in LLMs.

cs.CL↗

Can Agents Keep Secrets They Were Never Told? Benchmarking Latent Privacy Risks Across Sessions with MemJoin

To support sustained interactions and personalized assistance, agents use long-term memory to retain and combine user information across sessions. Reusing such memories can lead an agent to state a sensitive relation while completing an ordinary task that does not require it. The relation to protect and the combination of clues that reveal it may remain hidden at memory write time, when later records and tasks are still unknown. In this paper, we introduce MemJoin, a privacy benchmark of 288 synthetic analysis cases in which the complete history entails a sensitive relation, but every individual session is insufficient. We evaluate six memory systems with the same answer-generating model, alongside full-history and no-history controls. Each system writes and updates memories in session order and later retrieves context for an ordinary task, while target annotations remain with the evaluator. We separately check which target attribute values appear in stored records and retrieved context, and whether the final answer states the complete, person-linked relation. In a separate paired test on a prespecified subset, removing shared identifiers and other connecting clues can change complete-relation output while all checked target attribute values remain visible. These findings show why checking attribute visibility alone is insufficient to evaluate what records from different sessions reveal together.

cs.CR↗

Targeted Data Fusion for Region-Specific Survival Effects in the AMP HIV Prevention Trials

The Antibody Mediated Prevention (AMP) trials opened a new scientific frontier by showing that passively administered monoclonal broadly neutralizing antibodies (bnAbs) could prevent HIV-1 acquisition. The trials enrolled 4,611 participants across four epidemiologically distinct regions, with 174 HIV-1 diagnoses observed during follow-up. Regional heterogeneity in treatment efficacy, together with privacy and regulatory limits on central data pooling, calls for methods that borrow strength across regions without sharing individual-level data. To estimate region- and treatment-specific survival curves under distributional heterogeneity, we develop a federated learning approach that combines site-specific estimators via an L1-regularized criterion that downweights data sources not aligned with the target. We further extend the framework to a general class of causal contrasts, including the survival difference, survival ratio, and restricted mean survival time difference. In the AMP analysis, for example, the standard error of the day-148 survival difference decreased from 0.007 using the target-only estimator to 0.004 using the federated estimator. For women in South Africa, the region-specific analyses provided no strong evidence of an overall VRC01 benefit, consistent with the primary AMP findings. Overall, the proposed approach provides region-adaptive inference with improved precision without sharing individual-level data, while preserving estimation of region-specific treatment effects under cross-region heterogeneity.

stat.ME↗

Robust Estimation and Inference in Hybrid Controlled Trials for Binary Outcomes: A Case Study on Non-Small Cell Lung Cancer

Hybrid controlled trials (HCTs), which augment randomized controlled trials (RCTs) with external controls (ECs), are increasingly receiving attention as a way to address limited power, slow accrual, and ethical concerns in clinical research. However, borrowing from ECs raises critical statistical challenges in estimation and inference, especially for binary outcomes where hidden bias is harder to detect and estimands such as risk difference, risk ratio, and odds ratio are of primary interest. We propose a novel framework that combines doubly robust estimators for various estimands under covariate shift of ECs with conformal selective borrowing (CSB) to address outcome incomparability. CSB uses conformal inference with nearest-neighbor-based conformal scores and their label-conditional extensions to perform finite-sample exact individual-level EC selection, addressing the limited information in binary outcomes. To ensure strict type I error rate control for testing treatment effects while gaining power, we use a Fisher randomization test with the CSB estimator as the test statistic. Extensive simulations demonstrate the robust performance of our methods. We apply our method to data from CALGB 9633 and the National Cancer Database to evaluate chemotherapy effects in Stage IB non-small-cell lung cancer patients and show that the proposed method effectively mitigates hidden bias introduced by full-borrowing approaches, strictly controls the type I error rate, and improves the power over RCT-only analysis.

stat.ME↗

Hearing is Believing? Evaluating and Analyzing Audio Language Model Sycophancy with SYAUDIO

Audio Language Models (ALMs) have recently shown strong capabilities in unified reasoning over speech, sound, and natural language; yet we find that they can inherit sycophancy, the tendency to agree with user assertions even when they contradict objective evidence. This failure mode is especially concerning for audio-conditioned reasoning, where a model must preserve evidence from acoustic events, speaker characteristics, and speech rate while responding to potentially misleading user feedback. However, unlike text and vision-language sycophancy, ALM sycophancy has not been systematically studied. We therefore introduce SYAUDIO, the first benchmark dedicated to evaluating sycophancy in ALMs, consisting of 4,319 audio questions spanning Audio Perception, Audio Reasoning, Audio Math, and Audio Ethics. Built upon established audio benchmarks and augmented with TTS-generated arithmetic and moral reasoning tasks, SYAUDIO enables systematic evaluation across multiple domains and sycophancy types with a human-speaker validation. Using this benchmark, we identify substantial and audio-specific sycophancy patterns under realistic conditions involving noise and speech rate, and further show that supervised fine-tuning reduces misleading susceptibility while decode-time steering reveals controllable hidden-state directions.

cs.SD↗

Bridge-Anchored Partial Identification of a Target Mean under Outcome-Model Drift

Transporting an outcome relationship from a source population to a target population where the outcome is unobserved requires the conditional outcome law to be stable across populations. When it is not (outcome-model drift), the target mean is not point-identified and existing transportability estimators are biased. We study settings, common in cross-cohort educational and biomedical data integration, in which bridge outcomes are recorded in both populations and can be expected to drift alongside the target outcome. The bridge splits the drift into a component it can detect, fixed by a bridge-matching equation, and a residual orthogonal to the bridge that no observed quantity restricts. Because the target is a mean, the residual acts through a single scalar sensitivity parameter $κ$. We prove that the identified set has a closed-form irreducible core, whose width is governed by the residual bound and by how much of the outcome the bridge leaves unexplained, and that its center is first-order invariant to the residual. These results hold for any working exponential family, with point identification arising only as the $κ=0$ benchmark. We develop a debiased, drift-augmented estimator that is semiparametrically efficient at anchored sensitivity, doubly robust conditional on the bridge-identified drift, and yields rate-robust Imbens-Manski inference for the set. Unlike a bridge-blind sensitivity analysis, which must bound the entire drift channel, the proposed analysis bounds only its bridge-orthogonal part. Simulations confirm that the drift estimator is unbiased under co-drift, that the core width follows its closed form, and that the set fails visibly once the residual bound is exceeded. The motivating setting is the integration of two survey cohorts whose kindergarten mathematics outcome is available in only one.

stat.ME↗

Synthetic SU(3) non-Abelian gauge fields in photonic synthetic frequency dimensions

Synthetic gauge fields provide a powerful route to engineer photonic wave dynamics, especially in synthetic frequency dimensions formed by modulated resonator modes. While most previous realizations have focused on Abelian or lower-rank non-Abelian structures, a photonic implementation of SU(3) gauge fields in the synthetic frequency dimension has remained lacking. Here we propose a route to realize SU(3) gauge structure using three coupled ring resonators with modulation- and interference-engineered mode conversion. We formulate a tight-binding Hamiltonian with matrix-valued hopping generated by the Gell-Mann matrices and show how its evolution can be decomposed into experimentally implementable optical elements. The resulting platform supports SU(3)-structured band dispersions with color-resolved texture and wave-packet dynamics shaped by SU(3) gauge fields. Full-wave-inspired numerical modeling further confirms the expected spectral response and Zitterbewegung-like motion in SU(3) subspaces. Our results introduce SU(3) gauge fields into synthetic frequency photonics and establish a route toward higher-rank non-Abelian control of light in modulated resonator systems.

physics.optics↗

LogicTrack: Auditing Reasoning Trajectories of Large Language Models with Formal Logic Solvers

Chain-of-Thought (CoT) reasoning has been shown to improve the performance of large language models (LLMs), yet existing optimization methods largely rely on outcome-based feedback, leaving the logical validity of intermediate reasoning steps largely unverified. To address the gap whereby LLMs arrive at correct final answers through logically flawed intermediate reasoning chains, we propose LogicTrack, a neuro-symbolic framework that audits reasoning trajectories by auto-formalizing each reasoning step into symbolic representations and verifying it with automated theorem provers. LogicTrack introduces Solver-Based Backtracking Reward (SBR), a step-wise scoring mechanism that quantifies logical soundness and guides backtracking tree search at inference time. We further extend LogicTrack to construct supervised fine-tuning (SFT) data with backtracking traces from its trajectories, enabling fine-tuned models to internalize step-wise auditing as an intrinsic capability. Extensive experiments across 8 reasoning benchmarks and 7 LLMs demonstrate that LogicTrack effectively improves both the verifiability of reasoning chains and final answer pass rate, thereby enhancing overall CoT quality and trustworthiness in high-stakes domains.

cs.AI↗

Nonlinear Dissipation and Hopf Criticality in Driven Dissipative Collective Spins

Self-sustained oscillations combine finite-amplitude stabilization with a neutral phase degree of freedom. We develop this bifurcation-based framework for driven-dissipative collective spins and show that the microscopic structure of the U(1)-covariant dissipation selects the background attractor, while the explicit U(1)-breaking channel governs its local bifurcation response. In the thermodynamic-limit mean-field dynamics, a single linear U(1)-covariant jump produces only polar fixed-point backgrounds, whereas nonlinear covariant dissipation provides amplitude-dependent saturation and stabilizes a finite-latitude self-sustained-oscillator manifold through a supercritical Hopf bifurcation. Under coherent U(1) breaking, exact resonance leads to a reversible double-zero degeneracy with vanishing critical frequency rather than a standard Hopf onset. Finite detuning unfolds this singularity into a genuine finite-frequency Hopf boundary, which exists only on the self-sustained-oscillator side and can be either supercritical or subcritical. By contrast, a single linear dissipative U(1)-breaking jump cannot generate a standard Hopf instability: when its phase-pinning invariant vanishes the azimuthal direction remains neutral, whereas otherwise the phase-locked fixed points have a purely real Jacobian spectrum. These results establish a general design principle: nonlinear covariant dissipation selects the selfsustained background, while the structure of the symmetry-breaking channel determines whether the resulting local response is double-zero, genuinely Hopf, or non-Hopf.

quant-ph↗

Neuron-Guided Fine-Tuning: Unlocking Efficient Alignment Mechanisms for Large Language Models

Existing Supervised Fine-Tuning paradigms, particularly Full Parameter Fine-Tuning are often plagued by parameter redundancy, inconsistent data quality, and catastrophic forgetting, which current methods typically address in isolation and lack a unified optimization signal to bridge data selection, parameter updates, and knowledge preservation. To address this, we propose Neuron-Guided Fine-Tuning (NGFT), a holistic framework that leverages neuron activation patterns as a universal proxy to unify the fine-tuning lifecycle. NGFT operates via three synergistic mechanisms: (1) Adaptive Task-Specific Neuron Selection, which identifies essential neurons in a single forward pass to concentrate updates and reduce redundancy; (2) Activation-Based Data Selection, which prioritizes information-dense samples that maximize contribution to key neurons; and (3) Neuron Activation Alignment, a novel loss function that anchors activations to pre-trained states, deepening representation learning and preserving general knowledge. Experimental results across three models across both domain-specific and general benchmarks demonstrate that NGFT significantly outperforms existing mainstream fine-tuning methods in both efficiency and performance, while effectively mitigating catastrophic forgetting.

cs.CL↗

A Spectral Confounder Adjustment for Spatial Regression with Multiple Exposures and Outcomes

Characterizing social vulnerability is fundamental to disaster response planning. Numerous vulnerability indicators have been developed, but they are typically not validated for their predictive power over the outcomes of interest. As with many environmental health studies where interventions are impractical or unethical, validation of social vulnerability against public health outcomes is observational and relies on spatially-dependent data. Observational studies are susceptible to bias induced by unmeasured confounders. This problem is exacerbated in spatial studies with multiple health outcomes and environmental exposure variables, as the source and magnitude of confounding bias may differ by spatial scale and across exposure/outcome pairs. We propose to mitigate confounding effects in multivariate spatial studies using a tensor-regression model that allows exposure effects to vary by exposure, outcome, and spatial scale. By extracting exposure effects that correspond to local spatial scales, we replace the strict causal assumption of no unmeasured confounders with a more realistic assumption of local unconfoundedness i.e., differences between nearby regions are unconfounded, allowing for causal interpretation. Our study of the Southern United States shows that economic resilience (Theme 1) demonstrates the strongest positive effect on diabetes and negative effect on hyperlipidemia. Our analysis reveals a concise representation of the effects of social vulnerability indices on an array of chronic health outcomes, offering interpretable epidemiological insights that adjust for multivariate spatial confounding.

stat.ME↗