Researchers have identified specific linear representations, termed modal difference vectors, within language models that effectively distinguish between different categories of event plausibility. These vectors demonstrate that language models possess a more robust ability to categorize sentences by modality than previously understood. The study also found that these modal difference vectors develop in a consistent manner as models increase in training steps, layers, and parameter count, and can even be used to model fine-grained human categorization behaviors. AI
IMPACT Provides new insights into how language models process and categorize event plausibility, potentially informing human cognitive models.
RANK_REASON Academic paper on language model representations and human judgments of event plausibility.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
