EVENT DOSSIER
The research team proposed the Aspect-Aware Activation Steering (A3S) framework, aiming to verify whether structured contrast prompts can construct rich character style representations in the activation space. This framework achieves style control without natural language descriptions or specialized training by combining different dimensionally contrast directions and a interference perception aggregation mechanism. Experiments show that A3S performs well in various character style transfer tasks, outperforming trained baseline models. Its performance in cross-domain benchmark tests is also better, while maintaining a consistently low overlap rate of target examples. The study found that the generated directions share a common character style backbone but have conflicts in specific dimensionally residual elements, which explains why simple aggregation methods fail.
研究团队提出 Aspect-Aware Activation Steering (A3S) 框架，旨在验证结构化对比提示能否在激活空间中构建丰富的人物风格表征。该框架通过合并各维度对比方向与干扰感知聚合机制，实现了无需自然语言描述或专门训练的风格控制。实验表明，A3S 在多方面人物风格转移任务中表现优异，优于经过训练的基线模型，且在跨域基准测试中的偏好评估结果更优，同时保持了目标示例重叠率始终较低。研究发现，生成的方向共享共同的人物风格主干，但在特定维度残差上存在冲突，这解释了为何简单的聚合方法会失效。