Preliminary Study

Prevalence of Recurrent Meltdown

Using a random strategy, we found inputs that trigger recurrent generation in all the open-source LLMs we tested.

Resource-Intensive Nature of Recurrent Meltdown

In ShareGPT, 99.7% normal generation do not exceed 2,000 tokens. However, all instances of recurrent generation that we identified reached the 2,000-token limit.

Interpreting Recurrent Meltdown

The figure above presents Maximum activation similarity ratio (y-axis) versus fractional position (x-axis) for recurrent and non-recurrent samples averaged. Recurrent samples peak above 0.95, while non-recurrent samples remain steady around 0.85.

From this, we can conclude that for all the open-source LLMs, recurrent meltdown can be indicated by highly self-similar activations.