Interpreting the Repeated Token Phenomenon in Large Language Models
By: Itay Yona , Ilia Shumailov , Jamie Hayes and more
Potential Business Impact:
Fixes AI that can't repeat words correctly.
Large Language Models (LLMs), despite their impressive capabilities, often fail to accurately repeat a single word when prompted to, and instead output unrelated text. This unexplained failure mode represents a vulnerability, allowing even end-users to diverge models away from their intended behavior. We aim to explain the causes for this phenomenon and link it to the concept of ``attention sinks'', an emergent LLM behavior crucial for fluency, in which the initial token receives disproportionately high attention scores. Our investigation identifies the neural circuit responsible for attention sinks and shows how long repetitions disrupt this circuit. We extend this finding to other non-repeating sequences that exhibit similar circuit disruptions. To address this, we propose a targeted patch that effectively resolves the issue without negatively impacting the model's overall performance. This study provides a mechanistic explanation for an LLM vulnerability, demonstrating how interpretability can diagnose and address issues, and offering insights that pave the way for more secure and reliable models.
Similar Papers
Why do LLMs attend to the first token?
Computation and Language
Helps computers remember important information better.
Memory Tokens: Large Language Models Can Generate Reversible Sentence Embeddings
Computation and Language
Lets computers remember and perfectly repeat any text.
Attention Sinks in Diffusion Language Models
Computation and Language
Helps computers learn language more like humans.