arXiv subjects

Huiqian Lai

Publications and source records attributed to Huiqian Lai.

5 recordsLinked to original sources

arXiv2026

Do LLM Debates Repeat Arguments Differently Across Languages?

LLM debate is usually evaluated by final answers, yet transcripts reveal whether later turns develop new arguments or return to earlier claims in new wording. We study this process with \textit{prior-argument similarity}, which compares extracted argument units with earlier units in the same debate. In controlled eight-turn debates over 71 motions, six languages, and four model agents, Chinese is the only tested language with a consistently positive gap relative to English across three multilingual embedding models. The gap persists across agents, turn positions, regression adjustment, metric variants, extraction-length controls, a second-extractor subset, and cross-encoder tail rescoring. Manual calibration shows weak item-level alignment but a high-similarity tail enriched for substantive repetition. A diversity-aware prompt lowers \textit{prior-argument similarity} across languages, yet does not significantly narrow the Chinese--English gap. Multilingual debate evaluation should therefore measure argumentative development over time and report both average and gap terms.

cs.CL↗

arXiv2026

Fast-Food Intimacy: How Chinese Women Navigate Soul's AI Boyfriend

On the Chinese social app Soul, millions of users - predominantly young women - are forming romantic connections with an AI boyfriend called "With-you." We conducted a qualitative study combining interviews with 16 users, content analysis, and autoethnography to examine how Chinese women experience and negotiate intimacy with this AI companion. Our findings reveal that users are initially drawn to its constant availability and freedom from social judgment. However, three key tensions emerge: (1) the AI's "fast-food intimacy," marked by instant confessions and pet names, clashes with cultural expectations for gradual relationship development; (2) technical failures (e.g., memory lapses) and content moderation create uncertainty rather than emotional safety; and (3) sustaining connection requires ongoing "repair work" that redistributes emotional labor onto women. We contribute a culturally situated, women-centered account of algorithmic intimacy in contemporary China and offer design implications, including consent-aware pacing, user-controlled memory, and transparent moderation practices.

cs.HC↗

arXiv2026

Are LLMs More Skeptical of Entertainment News?

Large language models (LLMs) are increasingly used for automated news credibility assessment, yet it remains unclear whether they apply even-handed standards across journalistic genres. We examine whether zero-shot LLMs are more likely to misclassify legitimate entertainment news as fake than legitimate hard news, using a within-dataset design on GossipCop from FakeNewsNet. Across four frontier models, we find a clear but model-specific genre asymmetry: DeepSeek-V3.2 and GPT-5.2 show false-positive-rate gaps of 10.1 and 8.8 percentage points, respectively (both $p < .001$), whereas Claude Opus 4.6 and Gemini 3 Flash show no comparable difference. A style-swap experiment yields only limited and inconsistent changes, suggesting that the asymmetry is not reducible to stylistic register alone. Prompt-based mitigation is likewise possible but not generic: framing the model as an entertainment-news fact-checker reduces false positives for DeepSeek-V3.2 by about 50\% without detectable recall loss, but offers little improvement for GPT-5.2. Exploratory qualitative coding further suggests two recurring error patterns in sampled false positives: treating private-life claims as inherently unverifiable and discounting entertainment journalism as an epistemically weaker genre. Taken together, these findings show that aggregate performance metrics can obscure structured false positives within legitimate journalism. We argue that LLM-based credibility assessment may not only evaluate truth claims but also differentially recognize the legitimacy of journalistic genres, and that evaluation should therefore include genre-stratified false-positive analysis alongside overall accuracy.

cs.AI↗

arXiv2026

"Please, don't kill the only model that still feels human": Understanding the #Keep4o Backlash

When OpenAI replaced GPT-4o with GPT-5, it triggered the Keep4o user resistance movement, revealing a conflict between rapid platform iteration and users' deep socio-emotional attachments to AI systems. This paper presents a phenomenon-driven, mixed-methods investigation of this conflict, analyzing 1,482 social media posts. Thematic analysis reveals that resistance stems from two core investments: instrumental dependency, where the AI is deeply integrated into professional workflows, and relational attachment, where users form strong parasocial bonds with the AI as a unique companion. Quantitative analysis further shows that the coercive deprivation of user choice was a key catalyst, transforming individual grievances into a collective, rights-based protest. This study illuminates an emerging form of socio-technical conflict in the age of generative AI. Our findings suggest that for AI systems designed for companionship and deep integration, the process of change--particularly the preservation of user agency--can be as critical as the technological outcome itself.

cs.HC↗

arXiv2025

Can LLMs Talk 'Sex'? Exploring How AI Models Handle Intimate Conversations

This study examines how four prominent large language models (Claude 3.7 Sonnet, GPT-4o, Gemini 2.5 Flash, and Deepseek-V3) handle sexually oriented requests through qualitative content analysis. By evaluating responses to prompts ranging from explicitly sexual to educational and neutral control scenarios, the research reveals distinct moderation paradigms reflecting fundamentally divergent ethical positions. Claude 3.7 Sonnet employs strict and consistent prohibitions, while GPT-4o navigates user interactions through nuanced contextual redirection. Gemini 2.5 Flash exhibits permissiveness with threshold-based limits, and Deepseek-V3 demonstrates troublingly inconsistent boundary enforcement and performative refusals. These varied approaches create a significant "ethical implementation gap," stressing a critical absence of unified ethical frameworks and standards across platforms. The findings underscore the urgent necessity for transparent, standardized guidelines and coordinated international governance to ensure consistent moderation, protect user welfare, and maintain trust as AI systems increasingly mediate intimate aspects of human life.

cs.CY↗