Revising Context, Shifting Simulated Stance: Auditing LLM-Based Stance Simulation in Online Discussions
Summary
This paper studies how LLM-based stance simulation in online discussions is sensitive to counterfactual revisions of conversational context, and proposes an auditing framework comparing text-only and multimodal strategies.
View Cached Full Text
Cached at: 06/05/26, 02:08 PM
Paper page - Revising Context, Shifting Simulated Stance: Auditing LLM-Based Stance Simulation in Online Discussions
Source: https://huggingface.co/papers/2606.06443
Abstract
LLM-based stance simulation exhibits context sensitivity when subjected to counterfactual revisions, with both text-only and multimodal approaches showing robust stance transitions across different polarization mechanisms.
Large language modelsare increasingly used to simulate social media users and infer how individuals may respond to online discussions. However, it remains unclear whether these simulations reflect precise user-specific beliefs or whether they are highly sensitive to semantically independent changes inconversational contexts. In this work, we studycounterfactual context revisionas a framework for auditing LLM-basedstance simulation. Given an original online conversation, we first infer a target user’s stance toward a specific topic. We then apply controlled revision strategies to theconversational contextand simulate the user’s stance again under the revised context. We compare text-only revision strategies with a multimodal one that incorporates meme-based context and evaluate two main effectiveness metrics, i.e., average directional stance shift andstance transition rate. The results reveal effective and robust stance transitions in both text-only and multimodal strategies across differentpolarization-preference mechanisms. Our study contributes an evaluation framework for understanding the context sensitivity of LLM-basedstance simulation. More broadly, it highlights both the promise and risk of using LLMs to simulate online opinion dynamics.
View arXiv pageView PDFAdd to collection
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2606.06443 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2606.06443 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2606.06443 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
LLM-Assisted Stance Detection in Scientific Discourse: A Test Case in Bayesian Cognitive Science
This paper presents a method using LLMs for stance detection in scientific discourse, specifically identifying realism vs. instrumentalism in Bayesian cognitive science articles. The approach combines theory-driven coding, expert annotations, and prompt optimization to achieve high reliability.
When Should LLMs Trust Their Own Revisions? A Risk-Aware Study of Intrinsic Self-Correction
This arXiv paper studies intrinsic self-correction across 29 open-weight LLMs on BoolQ, GSM8K, and Corr2Cause, finding that revision can flip previously correct answers into errors and that self-correction should be treated as a selective revision policy rather than a uniformly beneficial second pass.
Semantic Variability of Replies Across LLMs: Implications for Designing Conversation-Based Assessment
The study examines the semantic consistency of LLM-generated replies across different models and conversational contexts, highlighting the need for infrastructure and design strategies to maintain stable responses for conversation-based assessments.
Robust Critics: Defending LLMs Against Multi-Turn Attacks
This paper proposes Dialogue Critic Guided Sampling (DCGS), a framework that defends LLMs against multi-turn adversarial attacks by inferring user intent from conversation history and using value/regret-based critics to score responses, achieving improved robustness without fine-tuning.
Belief Engine: Configurable and Inspectable Stance Dynamics in Multi-Agent LLM Deliberation
The paper introduces the Belief Engine, an auditable belief-update layer for LLM agents that makes stance changes in multi-agent deliberation configurable and inspectable by treating belief as an evidential state with explicit update rules.