arXiv subjects

Hao Cui

Publications and source records attributed to Hao Cui.

At least 19 recordsLinked to original sources

arXiv2026

Idea Search: Guiding Tree Search with Ideas to Explore Diverse Scientific Methods

Tree Search-based test-time scaling of LLMs is a powerful tool for automated scientific coding. However, pure Tree Search sometimes struggles with systematic exploration, becoming trapped in local optima, or unproductive loops, especially in the vast search space of scientific methods. To address this limitation, we propose Idea Search, a framework that systematically integrates a dynamic "Idea Bank" into Tree Search. Idea Search involves three steps: (1) decomposing existing methods into atomic ideas, (2) sampling from this bank of ideas to guide branches of code mutations, and (3) dynamically updating the bank with new ideas discovered through execution. On single-cell RNA-sequencing (scRNA-seq) batch integration, Idea Search reliably breaks the plateau of a strong pure Tree Search baseline, improving the mean score from 0.678 to 0.697 and reaching a best score of 0.728. We then characterize which design choices drive these gains: bank augmentation helps bandit sampling but not random sampling, "Exploratory" prompting that prioritizes new ideas surfaces the rare best-performing solutions, while increasing sampling-level exploration is counterproductive.

cs.LG↗

arXiv2026

CARE: A Molecular-Guided Foundation Model with Adaptive Region Modeling for Whole Slide Image Analysis

Foundation models have recently achieved impressive success in computational pathology, demonstrating strong generalization across diverse histopathology tasks. However, existing models overlook the heterogeneous and non-uniform organization of pathological regions of interest (ROIs) because they rely on natural image backbones not tailored for tissue morphology. Consequently, they often fail to capture the coherent tissue architecture beyond isolated patches, limiting interpretability and clinical relevance. To address these challenges, we present Cross-modal Adaptive Region Encoder (CARE), a foundation model for pathology that automatically partitions WSIs into several morphologically relevant regions. Specifically, CARE employs a two-stage pretraining strategy: (1) a self-supervised unimodal pretraining stage that learns morphological representations from 34,277 whole-slide images (WSIs) without segmentation annotations, and (2) a cross-modal alignment stage that leverages RNA and protein profiles to refine the construction and representation of adaptive regions. This molecular guidance enables CARE to identify biologically relevant patterns and generate irregular yet coherent tissue regions, selecting the most representative area as ROI. CARE supports a broad range of pathology-related tasks, using either the ROI feature or the slide-level feature obtained by aggregating adaptive regions. Based on only one-tenth of the pretraining data typically used by mainstream foundation models, CARE achieves superior average performance across 33 downstream benchmarks, including morphological classification, molecular prediction, and survival analysis, and outperforms other foundation model baselines overall.

cs.CV↗

arXiv2026

Visual Anthropomorphism Shifts Evaluations of Gendered AI Managers

This research examines whether competence cues can reduce gender bias in evaluations of AI managers and whether these effects depend on how the AI is represented. Across two preregistered experiments (N = 2,505), each employing a 2 x 2 x 3 design manipulating AI gender, competence, and decision outcome, we compared text-based descriptions of AI managers with visually generated AI faces created using a reverse-correlation paradigm. In the text condition, evaluations were driven by competence rather than gender. When participants received unfavourable decisions, high-competence AI managers were judged as fairer, more competent, and better leaders than low-competence managers, regardless of AI gender. In contrast, when the AI manager was visually represented, competence cues had attenuated influence once facial information was present. Instead, participants showed systematic gender-differentiated responses to AI faces, with feminine-appearing managers evaluated as more competent and more trustworthy than masculine-appearing managers, particularly when delivering favourable outcomes. These gender effects were largely absent when outcomes were unfavourable, suggesting that negative feedback attenuates the influence of both competence information and facial cues. Taken together, these findings show that competence information can mitigate negative reactions to AI managers in text-based interactions, whereas facial anthropomorphism elicits gendered perceptual biases not observed in text-only settings. The results highlight that representational modality plays a critical role in determining when gender stereotypes are activated in evaluations of AI systems and underscore that design choices are consequential for AI governance in evaluative contexts.

cs.CY↗

arXiv2026

QuEPT: Quantized Elastic Precision Transformers with One-Shot Calibration for Multi-Bit Switching

Elastic precision quantization enables multi-bit deployment via a single optimization pass, fitting diverse quantization scenarios.Yet, the high storage and optimization costs associated with the Transformer architecture, research on elastic quantization remains limited, particularly for large language models.This paper proposes QuEPT, an efficient post-training scheme that reconstructs block-wise multi-bit errors with one-shot calibration on a small data slice. It can dynamically adapt to various predefined bit-widths by cascading different low-rank adapters, and supports real-time switching between uniform quantization and mixed precision quantization without repeated optimization. To enhance accuracy and robustness, we introduce Multi-Bit Token Merging (MB-ToMe) to dynamically fuse token features across different bit-widths, improving robustness during bit-width switching. Additionally, we propose Multi-Bit Cascaded Low-Rank adapters (MB-CLoRA) to strengthen correlations between bit-width groups, further improve the overall performance of QuEPT. Extensive experiments demonstrate that QuEPT achieves comparable or better performance to existing state-of-the-art post-training quantization methods.Our code is available at https://github.com/xuke225/QuEPT

cs.CV↗

arXiv2025

Boolean function monotonicity testing requires (almost) $n^{1/2}$ queries

We show that for any constant $c>0$, any (two-sided error) adaptive algorithm for testing monotonicity of Boolean functions must have query complexity $\Omega(n^{1/2-c})$. This improves the $\tilde\Omega(n^{1/3})$ lower bound of [CWX17] and almost matches the $\tilde{O}(\sqrt{n})$ upper bound of [KMS18].

cs.CC↗

arXiv2025

From Birdwatch to Community Notes, from Twitter to X: four years of community-based content moderation

Community Notes (formerly known as Birdwatch) is the first large-scale crowdsourced content moderation initiative launched by X (formerly Twitter) in January 2021. As the Community Notes model gains momentum across other social media platforms, there is a growing need to assess its underlying dynamics and effectiveness. This paper provides a descriptive investigation of Community Notes during its first four years, examining its linguistic diversity, sourcing practices, Contributor activity, rating behaviour, and interaction networks. In addition, we release a curated dataset and accompanying source code to support future research, along with a review of prior research on Community Notes. We parsed Notes and ratings data from the first four years of the program and conducted language detection across all Notes. For English-language Notes, we extracted embedded URLs and identified discussion topics in each Note. Additionally, we constructed monthly interaction networks among the Contributors. Together, the descriptive analysis, dataset, code, and literature review provide a foundation for advancing research on Community Notes and community-based content moderation more broadly.

cs.SI↗

arXiv2025

An AI system to help scientists write expert-level empirical software

The cycle of scientific discovery is frequently bottlenecked by the slow, manual creation of software to support computational experiments\cite{hannay2009how}. To address this, we present Empirical Research Assistance (ERA), an AI system that creates expert-level scientific software whose goal is to maximize a quality metric. The system uses a Large Language Model (LLM) and Tree Search (TS)\cite{silver2016mastering} to systematically improve the quality metric and intelligently navigate the large space of possible solutions. ERA achieves expert-level results when it explores and integrates complex research ideas from external sources. The effectiveness of tree search is demonstrated across a diverse range of tasks. In bioinformatics, ERA discovered 40 novel methods for single-cell data analysis that outperformed the top human-developed methods on a public leaderboard. In epidemiology, ERA generated 14 models that outperformed the CDC ensemble and all other individual models for forecasting COVID-19 hospitalizations. ERA also produced expert-level software for geospatial analysis, neural activity prediction in zebrafish, and numerical solution of integrals, and a novel rule-based construction for time series forecasting. By devising and implementing novel solutions to diverse tasks, ERA represents a significant step towards accelerating scientific progress.

cs.AI↗

arXiv2025

FEABench: Evaluating Language Models on Multiphysics Reasoning Ability

Building precise simulations of the real world and invoking numerical solvers to answer quantitative problems is an essential requirement in engineering and science. We present FEABench, a benchmark to evaluate the ability of large language models (LLMs) and LLM agents to simulate and solve physics, mathematics and engineering problems using finite element analysis (FEA). We introduce a comprehensive evaluation scheme to investigate the ability of LLMs to solve these problems end-to-end by reasoning over natural language problem descriptions and operating COMSOL Multiphysics$^\circledR$, an FEA software, to compute the answers. We additionally design a language model agent equipped with the ability to interact with the software through its Application Programming Interface (API), examine its outputs and use tools to improve its solutions over multiple iterations. Our best performing strategy generates executable API calls 88% of the time. LLMs that can successfully interact with and operate FEA software to solve problems such as those in our benchmark would push the frontiers of automation in engineering. Acquiring this capability would augment LLMs' reasoning skills with the precision of numerical solvers and advance the development of autonomous systems that can tackle complex problems in the real world. The code is available at https://github.com/google/feabench

cs.AI↗

arXiv2025

CURIE: Evaluating LLMs On Multitask Scientific Long Context Understanding and Reasoning

Scientific problem-solving involves synthesizing information while applying expert knowledge. We introduce CURIE, a scientific long-Context Understanding,Reasoning and Information Extraction benchmark to measure the potential of Large Language Models (LLMs) in scientific problem-solving and assisting scientists in realistic workflows. This benchmark introduces ten challenging tasks with a total of 580 problems and solution pairs curated by experts in six disciplines - materials science, condensed matter physics, quantum computing, geospatial analysis, biodiversity, and proteins - covering both experimental and theoretical work-flows in science. We evaluate a range of closed and open LLMs on tasks in CURIE which requires domain expertise, comprehension of long in-context information,and multi-step reasoning. While Gemini Flash 2.0 and Claude-3 show consistent high comprehension across domains, the popular GPT-4o and command-R+ fail dramatically on protein sequencing tasks. With the best performance at 32% there is much room for improvement for all models. We hope that insights gained from CURIE can guide the future development of LLMs in sciences. Evaluation code and data are in https://github.com/google/curie

cs.CL↗

arXiv2025

Gender Bias in Perception of Human Managers Extends to AI Managers

As AI becomes more embedded in workplaces, it is shifting from a tool for efficiency to an active force in organizational decision-making. Whether due to anthropomorphism or intentional design choices, people often assign human-like qualities, including gender, to AI systems. However, how AI managers are perceived in comparison to human managers and how gender influences these perceptions remains uncertain. To investigate this, we conducted randomized controlled trials (RCTs) where teams of three participants worked together under a randomly assigned manager. The manager was either a human or an AI and was presented as male, female, or gender-unspecified. The manager's role was to select the best-performing team member for an additional award. Our findings reveal that while participants initially showed no strong preference based on manager type or gender, their perceptions changed notably after experiencing the award process. As expected, those who received awards rated their managers as more trustworthy, competent, and fair, and they were more willing to work with similar managers in the future. In contrast, those who were not selected viewed them less favorably. However, male managers, whether human or AI, were more positively received by awarded participants, whereas female managers, especially female AI managers, faced greater skepticism and negative judgments when they did not give awards. These results suggest that gender bias in leadership extends beyond human managers to include AI-driven decision-makers as well. As AI assumes more managerial responsibilities, understanding and addressing these biases will be crucial for designing fair and effective AI management systems.

cs.CY↗

arXiv2024

Understanding Privacy Norms through Web Forms

Web forms are one of the primary ways to collect personal information online, yet they are relatively under-studied. Unlike web tracking, data collection through web forms is explicit and contextualized. Users (i) are asked to input specific personal information types, and (ii) know the specific context (i.e., on which website and for what purpose). For web forms to be trusted by users, they must meet the common sense standards of appropriate data collection practices within a particular context (i.e., privacy norms). In this paper, we extract the privacy norms embedded within web forms through a measurement study. First, we build a specialized crawler to discover web forms on websites. We run it on 11,500 popular websites, and we create a dataset of 293K web forms. Second, to process data of this scale, we develop a cost-efficient way to annotate web forms with form types and personal information types, using text classifiers trained with assistance of large language models (LLMs). Third, by analyzing the annotated dataset, we reveal common patterns of data collection practices. We find that (i) these patterns are explained by functional necessities and legal obligations, thus reflecting privacy norms, and that (ii) deviations from the observed norms often signal unnecessary data collection. In addition, we analyze the privacy policies that accompany web forms. We show that, despite their wide adoption and use, there is a disconnect between privacy policy disclosures and the observed privacy norms.

cs.CR↗

arXiv2024

Understanding the Impact of Referent Design on Scale Perception in Immersive Data Visualization

Referents are often used to enhance scale perception in immersive visualizations. Common referent designs include the considerations of referent layout (side-by-side vs. in-situ) and referent size (small vs. medium vs. large). This paper introduces a controlled user study to assess how different referent designs affect the efficiency and accuracy of scale perception across different data scales, on the performance of the size-matching task in the virtual environment. Our results reveal that in-situ layouts significantly enhance accuracy and confidence across various data scales, particularly with large referents. Linear regression analyses further confirm that in-situ layouts exhibit greater resilience to changes in data scale. For tasks requiring efficiency, medium-sized referents emerge as the preferred choice. Based on these findings, we offer design guidelines for selecting referent layouts and sizes in immersive visualizations.

cs.HC↗

arXiv2024

AI-enhanced Collective Intelligence

Current societal challenges exceed the capacity of humans operating either alone or collectively. As AI evolves, its role within human collectives will vary from an assistive tool to a participatory member. Humans and AI possess complementary capabilities that, together, can surpass the collective intelligence of either humans or AI in isolation. However, the interactions in human-AI systems are inherently complex, involving intricate processes and interdependencies. This review incorporates perspectives from complex network science to conceptualize a multilayer representation of human-AI collective intelligence, comprising cognition, physical, and information layers. Within this multilayer network, humans and AI agents exhibit varying characteristics; humans differ in diversity from surface-level to deep-level attributes, while AI agents range in degrees of functionality and anthropomorphism. We explore how agents' diversity and interactions influence the system's collective intelligence and analyze real-world instances of AI-enhanced collective intelligence. We conclude by considering potential challenges and future developments in this field.

cs.CY↗

arXiv2024

C2Ideas: Supporting Creative Interior Color Design Ideation with Large Language Model

Interior color design is a creative process that endeavors to allocate colors to furniture and other elements within an interior space. While much research focuses on generating realistic interior designs, these automated approaches often misalign with user intention and disregard design rationales. Informed by a need-finding preliminary study, we develop C2Ideas, an innovative system for designers to creatively ideate color schemes enabled by an intent-aligned and domain-oriented large language model. C2Ideas integrates a three-stage process: Idea Prompting stage distills user intentions into color linguistic prompts; Word-Color Association stage transforms the prompts into semantically and stylistically coherent color schemes; and Interior Coloring stage assigns colors to interior elements complying with design principles. We also develop an interactive interface that enables flexible user refinement and interpretable reasoning. C2Ideas has undergone a series of indoor cases and user studies, demonstrating its effectiveness and high recognition of interactive functionality by designers.

cs.HC↗

arXiv2023

BehaVR: User Identification Based on VR Sensor Data

Virtual reality (VR) platforms enable a wide range of applications, however, pose unique privacy risks. In particular, VR devices are equipped with a rich set of sensors that collect personal and sensitive information (e.g., body motion, eye gaze, hand joints, and facial expression). The data from these newly available sensors can be used to uniquely identify a user, even in the absence of explicit identifiers. In this paper, we seek to understand the extent to which a user can be identified based solely on VR sensor data, within and across real-world apps from diverse genres. We consider adversaries with capabilities that range from observing APIs available within a single app (app adversary) to observing all or selected sensor measurements across multiple apps on the VR device (device adversary). To that end, we introduce BehaVR, a framework for collecting and analyzing data from all sensor groups collected by multiple apps running on a VR device. We use BehaVR to collect data from real users that interact with 20 popular real-world apps. We use that data to build machine learning models for user identification within and across apps, with features extracted from available sensor data. We show that these models can identify users with an accuracy of up to 100%, and we reveal the most important features and sensor groups, depending on the functionality of the app and the adversary. To the best of our knowledge, BehaVR is the first to analyze user identification in VR comprehensively, i.e., considering all sensor measurements available on consumer VR devices, collected by multiple real-world, as opposed to custom-made, apps.

cs.HC↗

arXiv2023

A CI-based Auditing Framework for Data Collection Practices

Apps and devices (mobile devices, web browsers, IoT, VR, voice assistants, etc.) routinely collect user data, and send them to first- and third-party servers through the network. Recently, there is a lot of interest in (1) auditing the actual data collection practices of those systems; and also in (2) checking the consistency of those practices against the statements made in the corresponding privacy policies. In this paper, we argue that the contextual integrity (CI) tuple can be the basic building block for defining and implementing such an auditing framework. We elaborate on the special case where the tuple is partially extracted from the network traffic generated by the end-device of interest, and partially from the corresponding privacy policies using natural language processing (NLP) techniques. Along the way, we discuss related bodies of work and representative examples that fit into that framework. More generally, we believe that CI can be the building block not only for auditing at the edge, but also for specifying privacy policies and system APIs. We also discuss limitations and directions for future work.

cs.CR↗

arXiv2022

Sums of Reciprocals of Recurrence Relations

There is a growing literature on sums of reciprocals of polynomial functions of recurrence relations with constant coefficients and fixed depth, such as Fibonacci and Tribonacci numbers, products of such numbers, and balancing numbers (numbers $n$ such that the sum of the integers less than $n$ equals the sum of the $r$ integers immediately after, for some $r$ which is called the balancer of $n$; If $n$ is included in the summation, we have the cobalancing numbers, and $r$ is called the cobalancer of $n$). We generalize previous work to reciprocal sums of depth two recurrence sequences with arbitrary coefficients and the Tribonacci numbers, and show our method provides an alternative proof of some existing results. We define $(a,b)$ balancing and cobalancing numbers, where $a$ and $b$ are constants that multiply the left-hand side and right-hand side respectively, and derive recurrence relations describing these sequences. We show that for balancing numbers, the coefficients $(3,1)$ is unique such that every integer is a $(3,1)$ balancing number, and proved there does not exist an analogous set of coefficients for cobalancing numbers. We also found patterns for certain coefficients that have no balancing or cobalancing numbers.

math.NT↗

arXiv2022

PoliGraph: Automated Privacy Policy Analysis using Knowledge Graphs (Journal Version)

Privacy policies disclose how an organization collects and handles personal information. Recent work has made progress in leveraging natural language processing (NLP) to automate privacy policy analysis and extract data collection statements from different sentences, considered in isolation from each other. In this paper, we view and analyze, for the first time, the entire text of a privacy policy in an integrated way. In terms of methodology: (1) we define PoliGraph, a type of knowledge graph that captures statements in a policy as relations between different parts of the text; and (2) we revisit the notion of ontologies, previously defined in heuristic ways, to capture subsumption relations between terms. We make a clear distinction between local and global ontologies to capture the context of individual policies, application domains, and privacy laws. We develop PoliGrapher, an NLP tool to automatically extract PoliGraph from the text using linguistic analysis. Using a public dataset for evaluation, we show that PoliGrapher identifies 40% more collection statements than prior state-of-the-art, with 97% precision. In terms of applications, PoliGraph enables automated analysis of a corpus of policies and allows us to: (1) reveal common patterns in the texts across different policies, and (2) assess the correctness of the terms as defined within a policy. We also apply PoliGraph to: (3) detect contradictions in a policy, where we show false alarms by prior work, and (4) analyze the consistency of policies and network traffic, where we identify significantly more clear disclosures than prior work. Finally, leveraging the capabilities of the emerging large language models (LLMs), we also present PoliGrapher-LM, a tool that uses LLM prompting instead of NLP linguistic analysis, to extract PoliGraph from the policy text, and we show that it further improves coverage.

cs.CR↗