Researchers find interpretable circuit for entity tracking in LLMs

Researchers have identified a specific circuit within large language models that handles dynamic entity tracking. This mechanism, termed a retrieval conditioned rebinding circuit, is responsible for binding entities to their attributes and updating this information as the model processes changing states. The study found this circuit present in models like Gemma and Llama, though its implementation varies, with Gemma expressing binding information in query/key subspaces and Llama primarily in key vectors. AI

IMPACT Reveals an interpretable mechanism for state tracking, potentially aiding in understanding and improving LLM reasoning capabilities.

RANK_REASON The cluster contains an academic paper detailing a new finding about the internal mechanisms of large language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Gemma
• Large language models
• Llama

• paper
• model release

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