New 'prolepsis' phenomenon identified in small transformer models

Researchers have identified a phenomenon called 'prolepsis' in small transformer models, where the model commits to a decision early in its processing and cannot correct it. This commitment is sustained by task-specific attention heads and is not easily detectable by standard residual-stream methods, though CLT-based steering shows some success. The study found that this prolepsis motif appears across different tasks in decoder-only models like Gemma 2-2B and Llama 3.2 1B, suggesting a shared underlying mechanism. AI

IMPACT Identifies a new limitation in small transformer models, potentially impacting their reliability and interpretability.

RANK_REASON The cluster contains an academic paper detailing a new phenomenon observed in transformer models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Anthropic
• arXiv
• central limit theorem
• Eric Jacopin
• Gemma 2-2B
• Hugging Face
• Lindsey et al.
• Llama 3.2 1B
• Prolepsis
• transformers

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →