👥 Social Computing¶

🤖 AAAI2026 · 11 paper notes

Argumentative Debates for Transparent Bias Detection: This paper proposes ABIDE (Argumentative BIas Detection by DEbate), which constructs Quantitative Bipolar Argumentation Frameworks (QBAFs) via neighborhood-based argument schemes, models the bias detection process as a structured debate, enables transparent bias reasoning from individual neighborhoods to the global level, and formally proves the correspondence between QBAF semantics and the expected behavior of bias detection.
Beyond Detection: Exploring Evidence-based Multi-Agent Debate for Misinformation Intervention and Persuasion: This paper proposes ED2D, a framework that integrates an evidence retrieval module into a multi-agent debate (MAD) system to enhance misinformation detection accuracy. Through controlled human experiments, it provides the first comparative evaluation of AI-generated debate transcripts versus expert human fact-checks in terms of persuasiveness and belief correction, revealing a double-edged-sword effect: the AI debate system achieves expert-level persuasiveness when correct, but may amplify misinformation when wrong.
Bias Association Discovery Framework for Open-Ended LLM Generations: This paper proposes the Bias Association Discovery Framework (BADF), which systematically extracts both known and unknown bias associations between demographic identities and descriptive concepts from LLM open-ended story generation, overcoming the limitation of prior methods that rely on predefined bias concepts.
Cross-modal Prompting for Balanced Incomplete Multi-modal Emotion Recognition: This paper proposes Cross-modal Prompting (ComP), which addresses the modality imbalance problem in incomplete multi-modal emotion recognition (IMER) via progressive prompt generation, cross-modal knowledge propagation, and a dynamic scheduler, achieving state-of-the-art performance across 4 datasets and 7 missing rates.
Fact2Fiction: Targeted Poisoning Attack to Agentic Fact-checking System: This paper proposes Fact2Fiction, the first poisoning attack framework targeting agentic fact-checking systems (e.g., DEFAME, InFact). It employs a Planner Agent to simulate claim decomposition and generate sub-questions, reverse-engineers key reasoning points from system justifications to craft targeted malicious evidence, and allocates the poisoning budget according to sub-claim importance. At a poisoning rate of only 1%, Fact2Fiction achieves 8.9%–21.2% higher attack success rate (ASR) than the state-of-the-art PoisonedRAG.
FactGuard: Event-Centric and Commonsense-Guided Fake News Detection: This paper proposes FactGuard, a framework that leverages LLMs to extract event-centric content (with style removed) and generate commonsense rationales. A Rationale Usability Evaluator dynamically assesses the reliability of LLM suggestions. Knowledge distillation yields a lightweight variant, FactGuard-D, that operates without LLM inference, achieving both robustness and efficiency in fake news detection.
From Imitation to Discrimination: Toward A Generalized Curriculum Advantage Mechanism Enhancing Cross-Domain Reasoning Tasks: This paper proposes CAPO (Curriculum Advantage Policy Optimization), an adaptive curriculum mechanism based on advantage signals. Through a two-phase strategy — first imitation (using only positive-advantage samples) then discrimination (introducing negative signals) — CAPO stably and significantly improves LLM performance on mathematical reasoning and multimodal GUI reasoning tasks.
Multi-modal Dynamic Proxy Learning for Personalized Multiple Clustering: This paper proposes the Multi-DProxy framework, which leverages learnable textual proxies for personalized multiple clustering through three key innovations: gated cross-modal fusion, dual-constraint proxy optimization, and dynamic candidate management, achieving state-of-the-art performance on all public benchmarks.
Reasoning About the Unsaid: Misinformation Detection with Omission-Aware Graph Inference: This paper proposes OmiGraph, the first omission-aware misinformation detection framework. By constructing omission-aware graphs, leveraging LLMs to reason about omission intent, and employing omission-guided message passing and aggregation mechanisms, OmiGraph extracts deception patterns from "what is unsaid," achieving average gains of +5.4% F1 and +5.3% ACC on bilingual datasets.
SceneJailEval: A Scenario-Adaptive Multi-Dimensional Framework for Jailbreak Evaluation: This paper proposes SceneJailEval, a scenario-adaptive multi-dimensional jailbreak evaluation framework that defines 14 jailbreak scenarios and 10 evaluation dimensions. Through a pipeline of scenario classification → dynamic dimension selection → multi-dimensional detection → weighted harm scoring, it achieves F1 of 0.917 on a self-constructed dataset (surpassing SOTA by 6%) and 0.995 on JBB (surpassing SOTA by 3%), while supporting harm severity quantification beyond binary classification.
T2Agent: A Tool-augmented Multimodal Misinformation Detection Agent with Monte Carlo Tree Search: This paper proposes T2Agent, a misinformation detection agent integrating an extensible toolset with Monte Carlo Tree Search (MCTS). By decomposing detection into sub-tasks targeting distinct forgery sources via a multi-source verification mechanism, T2Agent achieves a new state of the art on MMfakebench, improving the accuracy of the baseline MMDAgent by 28.7% using GPT-4o as the backbone.