Study purpose: This research investigates how social bots participate in and shape public opinion through behavioral patterns and emotional contagion, using the “Xinjiang Cotton” Twitter discourse as a case study. It aims to advance empirical understanding of bot–human interaction dynamics in geographically sensitive online discourse. Methodology: Combining social network analysis, LDA topic modeling, sentiment analysis and non-parametric statistical testing, we examined 6,781 Chinese-language tweets from March 24, 2021 to March 31, 2021. We used Botometer to identify bot accounts, while Gephi was employed to visualize interaction networks. Emotional expressions were examined with the Dalian University of Technology sentiment dictionary, and differences between user groups were further tested using the Mann–Whitney U test. Main findings: Bots comprised 49.1 % of active users, exhibiting high retweet volumes (68 % of posts) and content homogeneity. They formed clustered networks and exhibited lower levels of positive emotional expression compared with human users, while no significant difference was found in negative emotion. Emotional convergence was observed within bot clusters as well as across human–bot interactions. Multi-level emotional contagion occurred both within bot/human groups and across bots and humans. Social implications: Bot participation in emotionally charged discourse may influence public perception and contribute to polarization in transnational communication, particularly in geopolitically sensitive contexts. Their ability to amplify divisive narratives risks eroding public trust in digital spaces. Practical implications: Platforms should prioritize multi-dimensional bot detection (e.g., activity patterns, network structures, and emotional signals) over simple metrics. Policymakers must establish cross-border regulations, while educators should develop media literacy programs addressing emotional manipulation. Originality/value: This study integrates network visualization with sentiment-labeled edges and statistical testing to examine emotional contagion pathways. It provides the first systematic analysis of China-related bot behavior in Western platforms, offering methodological and contextual advancements for computational propaganda research.
30 August 2026 in Physica A: Statistical Mechanics and its Applications
Large Language Models (LLMs) have achieved remarkable progress in complex scientific reasoning, yet they face a fundamental trade-off between deductive rigor and computational efficiency. In modern long-chain reasoning models, this trade-off manifests as a persistent tendency toward ``overthinking'' and redundant token overhead during test-time compute. Existing optimization methodologies typically treat response quality and generation efficiency independently, leading to suboptimal trade-offs and prohibitive operational serving costs in practical deployments. To address this technological bottleneck, we propose SynergiSci, a novel dual-stage reinforcement learning framework that explicitly decouples quality learning from efficiency optimization for LLM-based reasoning systems. The first stage employs an information-gain-guided quality learning mechanism to capture the mutual information between derivation trajectories and answer correctness, successfully mitigating the trivial reward distribution problem in policy optimization. The second stage implements a dynamic length control strategy to explicitly penalize low-information-gain tokens, thereby learning highly streamlined deduction paths without compromising empirical accuracy. Extensive experiments across six complex scientific reasoning benchmarks demonstrate that SynergiSci achieves superior Pareto-optimal quality-efficiency synergy. Crucially, it reduces trajectory length by an average of 28.6% while simultaneously yielding a statistically significant accuracy improvement of 3.5% over state-of-the-art LLM baselines. Consequently, SynergiSci provides a principled, computationally economical paradigm for deploying highly accurate and viable reasoning policies for complex scientific domains.