CHI 2026 Best Paper｜社会模拟迈入可控、可量化时代：为AI Agent加上「认知滑条」

Core Insights - The article discusses the transition of AI social simulation from anecdotal narratives to a controlled and reproducible experimental science paradigm through the CoBRA framework [3][21]. Group 1: CoBRA Framework - CoBRA provides a quantifiable, verifiable, and reproducible agent control framework that transforms classic social science experiments into reusable calibration environments [2][21]. - The framework allows for the measurement, feedback, and convergence of agent behavior, establishing a variable control mechanism akin to experimental science [2][21]. Group 2: Cognitive Bias Measurement - CoBRA utilizes a cognitive bias index to quantify the degree of bias in agents, covering four representative types of cognitive biases: authority effect, bandwagon effect, confirmation bias, and framing effect [11]. - Each type of bias corresponds to two classic experimental paradigms for cross-calibration and validation, with agent performance quantified on a continuous scale from 0 to 4 [11]. Group 3: Behavior Regulation Engine - The behavior regulation engine of CoBRA operates across three intervention spaces: input space, activation space, and parameter space [12]. - The input space uses numerical instructions to replace vague qualitative descriptions, while the activation space employs comparative samples to extract clean bias direction vectors [13]. - The parameter space involves training "biased" and "unbiased" LoRA models to fine-tune control signals, allowing precise adjustments [13]. Group 4: Cross-Model Consistency and Robustness - CoBRA has been systematically evaluated across various open-source and closed-source systems, demonstrating cross-model, cross-inference mode, and cross-scenario stability [16]. - The framework significantly reduces behavioral variance and maintains statistical equivalence across different temperature ranges, showcasing high consistency in control curves [16]. Group 5: Practical Application - CoBRA's practical value is illustrated through a simulation of emotional contagion, where agents generate content based on varying proportions of negative posts, revealing clear dose-response relationships [18][19]. - The results indicate that higher cognitive bias indices correlate with stronger emotional contagion, providing distinct and stable differentiation between levels [18][19]. Group 6: Conclusion - The significance of this work lies in its shift from "looking plausible" to "controllable and reproducible scientific research," marking a critical advancement in AI social simulation [21][22]. - CoBRA enables agents to have clear calibration and adjustment mechanisms, facilitating a transition to reproducible engineering in social simulations [22].