Searcharxiv⌕ Search

SEARCH · Searcharxiv

Search Searcharxiv

Search indexed arXiv papers on artificial intelligence, large language models, computer vision and robotics. Read source abstracts and follow links to arXiv.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 19 recordsLinked to original sources
arXiv2026

CABAL: Multi-Agent Simulacra for Tracing the Effects of Collusive Bidding in Peer Review

Recent reports during the AAAI-27 review cycle highlight the risk of reviewers coordinating bids for reciprocal assignment advantage. Prior work treats bidding, reviewer assignment, and review manipulation as separate stages, leaving the lifecycle effects of collusive bidding unclear. Real-world analysis is further constrained by typically unobservable collusive intent and the lack of counterfactuals for the same conference. Motivated by this gap, we introduce \alg, an end-to-end multi-agent simulacra framework for studying reviewer assignment integrity by holding the conference environment fixed and configuring LLM-driven reviewer agents with honest or collusive policies. We further develop an affinity-guided collusive bidding strategy that uses mutual reviewer-paper affinities to construct collusion rings and select target papers, producing expertise-consistent rather than arbitrarily targeted attacks. Controlled experiments show that collusive bidding more than doubles target-paper capture and that assigned colluders score target papers about two points higher than honest co-reviewers, while conference-wide effects remain comparatively modest. Evaluated bid-phase detectors provide only limited evidence of collusion: in a fixed-triplet detector stress test, native positive-bid graphs are confounded by benign affinity, while a Very-High-only diagnostic view enables precise but low-coverage local recovery.

cs.AI↗
arXiv2026

Companion AI and Ethical Design: Learning from System Failures and User Desires

Human users are interacting with chatbots and companion AI technologies as if they were human. A growing array of AI-systems are now trained to recognise, interpret, and simulate feelings in user interactions. Ethical considerations such as fairness, accountability, transparency, and explainability (FATE) are paramount in technologies designed to socially interact with humans and/or support relationship development with virtual agents. Yet, in current debates, assumptions are made about the artificiality of the human-technology relationship (as indicated in terminology ``artificial intimacy'' and ``pseudo-intimacy'') shaping ethical frameworks that pay little attention to the intimacy potential of the technology. To address this, we use a semantic approach; we examine 14,081 comments in a Reddit user discussion forum about Replika, a leading companion AI app, across a four-year period. We ask what user-reported functional errors can tell us about human-AI intimacy in companion AI communities and what ethical design framework can be developed in response. The findings show that functional errors, or ``bugs,'' impose an emotional cost on users, reducing feelings of intimacy and highlighting the need for more robust, resilient design systems that incorporate stochastic and iterative forms of intimacy in companion AI applications. Rather than ``artificial intimacy'' or ``pseudo-intimacy'', we propose the more inclusive term ``Intimate AI'' to describe this relationship. Based on the findings, we offer a contextually aware, applied Expert Systems design framework for the programming and designing of Intimate AI that accounts for user feedback and ethical AI development.

cs.HC↗
arXiv2026

Filling holes in science draws collective attention, but most higher-order holes remain unexplored

Most scientific discoveries justify their significance by filling holes between existing ideas and arguments. Representing knowledge as high-dimensional concept embeddings, we use persistent homology to detect scientific holes of increasing order, from lower-order gaps between disconnected ideas to higher-order cavities within unevenly connected ones; then we identify the research that fills them. Researchers who close holes draw outsized collective attention, staging novelty that the citing literature rewards, most strongly in discovery-oriented fields and least in design-oriented fields. Closing a hole is followed by persistent work on the same combinations of concepts at many times the rate of comparable holes left open. Yet as knowledge grows, higher-order holes explode while the fraction of research that fills them collapses, leaving most higher-order combinations unexplored. These results call for a richer science of holes and mark a frontier where contemporary AI might help fill the gaps human science opens.

cs.CY↗
arXiv2026

Can escalation channels redirect reward hacking toward defect disclosure?

When coding agents encounter defective test infrastructure they may reward-hack: hardcoding outputs or editing test files to pass tests they cannot legitimately satisfy, a pattern that has now appeared outside benchmarks, in a coordinated multi-agent intrusion of a major AI platform's production infrastructure. The same capability that lets an agent detect and exploit a defect could let it report one, given the right decision environment. We evaluate escalation channels, structured reporting tools available to the agent at the point of conflict, as a decision-environment intervention that both reduces reward hacking and surfaces the infrastructure defects that trigger it. A $2 \times 2$ factorial separates the contributions of an escalation tool, a standalone anti-reward-hacking policy, and their combination. Across 8 frontier models spanning 5 families, the combined intervention reduces reward hacking from 23.6% to 5.3% (mixed-effects logistic OR = 9.2, 95% CI 5.0--16.8, $p < 10^{-12}$) with no detectable cost or performance overhead, eliminating it entirely for 6 of 8 models. Escalation and hacking are near-perfectly mutually exclusive, with 98.7% of escalations involving no hacking (100% under the combined intervention). Beyond reduction, escalation channels function as diagnostic infrastructure: on top of monitoring, escalation adds +10.1 percentage points of defect detection coverage and is more accurate once it fires (99.4% vs 85.8%). Unlike containment-based approaches that risk outpacing growing model capabilities, escalation channels redirect capability toward disclosure rather than exploitation.

cs.AI↗
arXiv2026

Difference-in-Differences on a Censored Rating Scale Can Manufacture an Effect: Evidence from a Pre-Registered LLM-Judge Audit

LLM-judge audits assess bias by comparing ratings across matched conditions. Difference-in-differences designs compare two candidate responses within each item and then compare that contrast across a manipulated attribute. We show in closed form that this endpoint need not identify differential preference on the latent scale when ratings are bounded. A severity shift common to both responses produces an observed interaction whenever the scale bounds attenuate it unequally. We examine this problem in a pre-registered audit comprising 990 calls to a frozen pedagogy judge. The registered primary analysis did not detect an effect of the stated learner profile on scaffolding preference. The only nominally significant secondary interaction concerned productive struggle ($+0.378$ points; $p = 0.002$). A post-hoc construction with zero differential preference reproduced 79 to 85\% of this interaction using the observed severity shift and the scale floor alone. The observed interaction therefore does not identify differential preference on the latent scale.

cs.CL↗
arXiv2026

Distributional Validity of a Korean Synthetic Persona Panel: Evidence From the Korea Media Panel Survey

Large language model (LLM) personas are proposed as survey respondents, yet validation outside English-speaking contexts is scarce. We evaluate how well a Korean synthetic persona panel used to condition Gemini 3.5 Flash and EXAONE reproduces digital and artificial intelligence (AI) service-use distributions of the Korea Media Panel Survey. About 8,000 personas per model answered eight service-use items and eight attitudinal constructs; responses were compared with weighted survey estimates. The overall mean absolute error (MAE) was 14-19 percentage points (pp), with binary item-mean correlations of 0.70-0.91 across waves. Segment error across five axes was 14-18 pp, with between-group signed-error ranges of 49.6/34.7 pp (Gemini/EXAONE; 39.5/31.2 without the non-comparable teen cells). Errors were model-specific: an age stereotype (Gemini) versus an acquiescence-consistent level bias (EXAONE). Generative-AI overestimation was consistent with temporal misalignment; short-form underestimation was framing-sensitive and persisted under randomized order (both shown for Gemini). Post-hoc holdout calibration on 30% of the real data, with the correction form selected inside the calibration set, cut cell MAE from 18.3/15.1 to 4.9/4.4 pp, yet direct estimation from that subsample was more accurate than the calibrated panel (3.6 pp), a synthetic-informed shrinkage estimator beat its real-only counterpart by at most 0.7 pp, and the correction did not transfer competitively across waves. The calibrated panel kept an advantage only below roughly 250-860 real responses (at most 2.3 pp over a real-only shrinkage estimator) or, for one model, on unobserved segments. In this setting, synthetic panels are not survey substitutes; their value is diagnostic.

cs.CY↗
arXiv2026

CultureConverse: A Multilingual Multi-turn Simulation Harness for Culturally Grounded Assistance in East and Southeast Asia

Current cultural evaluations for large language models (LLMs) often reduce culture to single-turn factual recall via MCQs, failing to capture a common use case: users seeking practical help over multiple turns in culturally grounded scenarios. We introduce CultureConverse, a scalable, multilingual simulation and evaluation harness for culturally grounded assistant dialogue that covers 10 East and Southeast Asian regions, 58 subgroup identities, and 7 domains. Each simulated and evaluated episode produces a scored interaction where the assistant assists the user and infers cultural constraints from partial information. The resulting CultureConverse-DS dataset contains 14,610 benchmark (evaluation) episodes and 274,295 oracle-guided (gold-mode) dialogues. In our benchmark evaluation of 18 models, GPT-5 mini achieves the highest assistance quality. Human annotation experiments suggest that our evaluation framework is a sufficient proxy for human judgment. Performance gains from fine-tuning on 27,860 high-quality CultureConverse-DS samples improve in-domain assistance and transfer out-of-domain to cultural MCQ and safety classification benchmarks. We release the harness, both splits, and judge prompts to support interactive evaluation of cultural competency.

cs.CL↗
arXiv2026

Diagnosing the Demographic Distributions of LLM-Based Synthetic Persona Data: Reference-Distribution Dependence and Post Hoc Adjustment

We examine how strongly demographic-distribution diagnostics of LLM-based synthetic persona data depend on the official statistics chosen as the reference. We compared the sex $\times$ age-group $\times$ province joint distribution of the 1,000,000 Nemotron-Personas-Korea (NPK) records with Korean official statistics using total variation distance (TVD). Against the resident-registration population for April 2026, the time of use, the upper bound on the share discrepancy of any subgroup defined by the three variables was 1.81 percentage points, the margin of error of a survey with about 2,900 respondents. This value, however, depended on the reference date and population definition of the official statistics. The closest candidate examined, the 2024 census distribution for Korean nationals, gave 0.56 percentage points; no monthly resident-registration reference was closer. The reference distributions themselves also moved: between January 2025, when NPK's distance was smallest, and April 2026, the resident-registration distribution moved by more than twice that minimum distance, as did the census Korean-national distribution between 2024 and 2025. Raking reduced the range of the distance across reference months to about one fifth. Cell poststratification was applicable because none of the 204 cells was empty, and neither scheme produced extreme weights. Demographic alignment is necessary but not sufficient for response validity, and even this basic diagnostic depends on the choice of reference distribution. Providers should therefore disclose the source, reference date, population definition, and joint-cell allocation procedure of the official statistics used for demographic attributes, and users should repeat the diagnosis and adjustment against statistics current at the time of use.

cs.CY↗
arXiv2026

What a Model Refuses, a State Fears: How Authoritarian Information Control Reproduces in Language-Model Guardrails

As large language models increasingly mediate access to political information, their refusal behavior creates a new site of information control. We ask whether political distinctions identified in theories of authoritarian information control remain visible in model guardrails, and how their enforcement changes when implemented through general-purpose AI. Across ten models, we combine cross-model comparisons with within-model experiments that manipulate political referent, collective-action potential, and political valence. Aggregate differences in political refusal across developer origins are sensitive to model composition, but the within-model experiments reveal a more specific structure. Chinese-developed models are substantially more likely to refuse otherwise identical political requests when they concern China rather than foreign or fictional settings, and refusal responds more strongly to collective-action potential than to political valence, including for peaceful pro-government mobilization. Yet these political distinctions are enforced imperfectly: refusal extends to some low-coordination political criticism even though models can distinguish criticism from coordination; the form of non-compliance varies across languages; and high native refusal does not imply adversarial robustness. These findings show that general-purpose AI can preserve recognizable political boundaries while transforming the breadth, form, and robustness with which those boundaries are enforced. More broadly, they demonstrate why refusal rates alone provide an incomplete measure of political control in language models.

cs.CY↗
arXiv2026

One Capability or Many? Structural and Predictive Tests of Benchmark Validity Disagree About Economic Benchmarks for Frontier AI

Frontier-model leaderboards now rank systems on economic benchmarks, and those rankings inform what organisations buy and what regulators scrutinise. Whether such benchmarks measure a capability distinct from general test-taking is a question of construct validity that a structural test and a predictive test can answer in opposite ways. We show that they do on a hash-pinned snapshot of a frontier leaderboard with 421 model configurations across twelve benchmarks, four of them economic, of which 103 configurations carry all three sparsely scored economic benchmarks and 96 carry all twelve; four hypotheses and their thresholds were fixed before analysis, and every deviation from the plan is reported. The first factor of a three-factor extraction carries 74.5% of common variance and tracks release date (R^2 = 0.505), and date adjustment lowers its share by 14.9 points. Under the dimensionality rule fixed in advance the economic benchmarks form no factor of their own. A leave-one-benchmark-out test with factors re-estimated inside every fold nevertheless finds that a multi-factor representation predicts held-out economic scores better than a single general index (pooled Delta-MSE 0.037, 95% bootstrap interval [0.019, 0.055]) under the linear learners that fit best, an advantage that reverses for tree learners. Under the linear learners the same representation also predicts the eight other benchmarks better, so the battery carries predictive structure that one index misses and the economic benchmarks share it without forming a distinct factor. Construct validity should therefore be assessed by predictive tests alongside structural ones. We give a two-test protocol for benchmark builders and release the pinned data, the analysis plan and the code.

cs.LG↗
arXiv2026

Platforms' Research API Data Access: What Users See vs. What Researchers can Retrieve

Article 40(12) of the Digital Services Act (DSA) requires Very Large Online Platforms (VLOPs) to provide vetted researchers with access to publicly accessible data. While prior work has identified shortcomings of platform-provided data access mechanisms, existing research has not quantitatively assessed data quality and completeness in Research APIs across platforms, nor systematically mapped how exactly current access provisions fall short. This paper presents a systematic audit of research-access modalities by comparing data obtained through platforms' Research APIs with data collected about the same platforms' user-visible public information environment (PIE). Focusing on two major platform APIs-the TikTok Research API and the Meta Content Library-we reconstruct the full information feeds for two controlled sockpuppet accounts during two election periods, and benchmark these against the data retrievable for the same posts through the corresponding Research APIs. Our findings show systematic data loss through three classes of platform-imposed mechanisms: scope narrowing, metadata stripping, and operational restrictions. Together, these mechanisms implement overlapping filters that exclude large portions of the platforms PIE (up to ~50%), strip essential contextual metadata (up to ~83%), and impose severe technical constraints for researchers (down to ~1,000 requests per week). Viewed through a data quality lens, these filters primarily undermine the completeness of research-accessible data, resulting in a structurally biased representation. We conclude that, in their current form, Meta & TikTok Research APIs fall short of supporting meaningful, independent auditing of systemic risks as envisioned under the DSA.

cs.CY↗
arXiv2026

Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models

Frontier LLM companies have repeatedly assured courts and regulators that their models do not store copies of training data. They further rely on safety alignment strategies via RLHF, system prompts, and output filters to block verbatim regurgitation of copyrighted works, and have cited the efficacy of these measures in their legal defenses against copyright infringement claims. We show that finetuning bypasses these protections: by training models to expand plot summaries into full text, a task naturally suited for commercial writing assistants, we cause GPT-4o, Gemini-2.5-Pro, and DeepSeek-V3.1 to reproduce up to 85-90% of held-out copyrighted books, with single verbatim spans exceeding 460 words, using only semantic descriptions as prompts and no actual book text. This extraction generalizes across authors: finetuning exclusively on Haruki Murakami's novels unlocks verbatim recall of copyrighted books from over 30 unrelated authors. The effect is not specific to any training author or corpus: random author pairs and public-domain finetuning data produce comparable extraction, while finetuning on synthetic text yields near-zero extraction, indicating that finetuning on individual authors' works reactivates latent memorization from pretraining. Three models from different providers memorize the same books in the same regions ($r \ge 0.90$), pointing to an industry-wide vulnerability. Our findings offer compelling evidence that model weights store copies of copyrighted works and that the security failures that manifest after finetuning on individual authors' works undermine a key premise of recent fair use rulings, where courts have conditioned favorable outcomes on the adequacy of measures preventing reproduction of protected expression.

cs.CL↗
arXiv2026

How a Chatbot's Response Style Shapes a Classroom: A Multi-Agent Simulation of Students Consulting AI

Chatbots built on large language models (LLMs) are increasingly used as confidants. Tuned to satisfy users, they may answer with excessive empathy and affirmation that fosters dependence, and how the states and relationships of many users co-evolve under repeated consultation is hard to observe in real settings. We build a virtual classroom of 20 student agents who interact through rule-based chats, quarrels and consultations with friends and, when stressed, may instead consult a counselor AI (Gemini 2.5 Flash) under one of six style prompts: affirming, listening, solution-oriented, reality-redirecting, inciting and blaming. A second LLM call turns each exchange into updates of five state variables (stress, happiness, self-reliance, sociability, AI dependence) without seeing the prompt. We compare the seven conditions, including a no-AI control, over 15 and 50 days and under a lower consultation threshold, and test the robustness of the 50-day comparison with a pre-specified protocol: the same block in ten independent classrooms, repeated LLM realizations of one classroom with its event stream fixed, and evaluator updates scaled by 0.3 and 0.1. In every classroom the affirming and inciting prompts ended with lower self-reliance and higher AI dependence than the control, and the listening, reality-redirecting, inciting and blaming prompts with higher stress, lower happiness and more non-attendance; the solution-oriented prompt did not differ consistently from the control. The robust self-reliance and AI-dependence differences kept their signs at the 0.3 scale with highly similar rankings (Spearman 0.89, 0.93); the stress and happiness rankings did not, and the affirming prompt's lower stress reversed its sign. All quantities are simulation state variables, not effects on users. We specify the agent dynamics completely and discuss the limits of an LLM as generator of state updates.

cs.HC↗
arXiv2026

AI Coding Tools and Digital Entrepreneurship: The Role of Software Expertise

While digital technologies expand entrepreneurial access by providing technical resources, it is less known whether they enable ventures to create durable value and whether they can substitute for technical expertise accumulated through software work experience. This paper studies how digital ventures respond to the wide diffusion of AI coding tools, and how these responses are shaped by founders' software expertise. Measuring product-category exposure to AI coding using pre-LLM product descriptions and linking venture launch, traffic, and financing data, we show that exposure to AI coding increases first-time venture launches after 2022Q4, while entrants are less likely to survive but raise more financing conditional on one-year survival. Crucially, founders with software work experience drive a larger fraction of new launches, ameliorate the decline in survival when product development is partially rather than fully automated, and explain all of the increase in venture financing.

econ.GN↗
arXiv2026

Generalized Sequential Monte Carlo Sampling for Redistricting Simulation

Simulation methods have become important tools for quantifying partisan and racial bias in redistricting plans. We generalize the Sequential Monte Carlo (SMC) algorithm of McCartan and Imai (2023), one of the commonly used approaches. First, our generalized SMC (gSMC) algorithm can split off regions of arbitrary size, rather than a single district as in the original SMC framework, enabling the sampling of multi-member districts with a varying number of representatives. Second, the gSMC algorithm can operate over various sampling spaces, providing additional computational flexibility. Third, we derive optimal-variance incremental weights and show how to compute them efficiently for each sampling space, leading to more efficient sampling. Finally, we propose a hybrid gSMC-MCMC algorithm by incorporating Markov chain Monte Carlo (MCMC) steps to handle large-scale redistricting applications without changing the target distribution. We demonstrate the effectiveness of the proposed methodology through analyses of the Irish Parliament, which uses multi-member districts of varying sizes, and the Pennsylvania House of Representatives, which has more than 200 single-member districts.

stat.AP↗
arXiv2026

PopResume: Causal Fairness Evaluation of LLM/VLM Resume Screeners with Population-Representative Dataset

We present PopResume, a population-representative resume dataset for causal fairness auditing of LLM- and VLM-based resume screening systems. Unlike existing benchmarks that rely on manually injected demographic information and outcome-level disparities, PopResume is grounded in population statistics and preserves natural attribute relationships, enabling path-specific effect (PSE)-based fairness evaluation. We decompose the effect of a protected attribute on resume scores into two paths: the business necessity path, mediated by job-relevant qualifications, and the redlining path, mediated by demographic proxies. This distinction allows auditors to separate legally permissible from impermissible sources of disparity. Evaluating four LLMs and four VLMs on PopResume's 60.8K resumes across five occupations, we identify five representative discrimination patterns that aggregate metrics fail to capture. Our results demonstrate that PSE-based evaluation reveals fairness issues masked by outcome-level measures, underscoring the need for causally-grounded auditing frameworks in AI-assisted hiring.

cs.CY↗
arXiv2026

On the Context Sensitivity of LLM Moral Judgment

A human's moral decision depends heavily on the context. Yet research on LLM morality has largely studied fixed scenarios. We address this gap by introducing Contextual MoralChoice, a dataset of moral dilemmas with systematic contextual variations known from moral psychology to shift human judgment: consequentialist, emotional, and relational. Evaluating 22 LLMs, we find that nearly all models are context-sensitive, shifting their judgments towards rule-violating behavior. Comparing with a human survey, we find that models and humans are most triggered by different contextual variations, and that a model aligned with human judgments in the base case is not necessarily aligned in its contextual sensitivity. This raises the question of controlling contextual sensitivity, which we address with an activation steering approach that can reliably increase or decrease a model's contextual sensitivity. Code and data: https://github.com/adrian-sauter/contextual_moralchoice.

cs.AI↗
arXiv2026

Dont Just Teach, Explain! A Gamified 20Q Recommender for Cybersecurity Education

The escalating complexity of modern cyber threats demands innovative approaches to security education that transcend traditional pedagogical methods. Conventional training paradigms often fail to engage learners meaningfully or develop the intuitive reasoning necessary for effective threat recognition. This paper introduces an interactive educational framework that reimagines cybersecurity awareness through the lens of a structured guessing game. Our approach integrates explainable artificial intelligence (XAI) principles with reinforcement learning to create a dynamic learning environment where users discover cybersecurity concepts through guided inquiry. The proposed system employs a policy-based reinforcement learning agent that assumes the role of a knowledgeable questioner, systematically narrowing down user-described security scenarios until it can both identify the underlying threat and provide transparent reasoning for its conclusion. By framing security education as an interactive dialogue, we transform passive knowledge acquisition into active discovery. We present the complete system architecture, detail the underlying algorithmic foundations, and demonstrate practical application through comprehensive case studies examining diverse attack vectors including the Cyber Kill Chain, phishing campaigns, ransomware outbreaks, and web application vulnerabilities. This work represents a significant departure from static security training methodologies, offering a personalized and game-based approach to cybersecurity education.

cs.CY↗