Simulating Diverse Human Behavior with Anacreon

Summary & Methodology Analysis

The paper addresses the issue of model collapse, where large language models fail to simulate heterogeneous individuals by regressing toward generic, agreeable responses. To combat this, the researchers developed Anacreon, built on a Gemma 4 12B backbone. The core methodology employs QLoRA, a method of parameter-efficient fine tuning (the process of adapting a pre-trained model to specific tasks by updating only a small subset of weights), to create a mixture of minds architecture. By clustering unstructured corpus data around seed individuals, the team trained dedicated LoRA (Low-Rank Adaptation) adapters for each cluster, allowing the model to switch behaviors based on specific individual profiles. The training process was further refined by augmenting records with a chain of emotion, which maps the evolution of emotional states throughout the interaction. To ensure robustness, the researchers employed a multi-objective checkpoint selection process, balancing multiple-choice accuracy, ordinal alignment, and chain of emotion perplexity. This architectural shift significantly improves the model's ability to maintain distinct viewpoints, achieving an ordinal alignment of 0.775. The training protocol also focused on reducing prompt brittleness by shuffling response orders and balancing questions to mitigate sycophancy, which is the tendency of a model to provide answers that align with user bias. Despite these improvements, the model still exhibits a residual positive bias and demonstrates inconsistent performance across different clusters. Additionally, the paper notes that the system's predictive accuracy is generally limited to the domain of its training data.