Penelope: Localized Latent Recurrence for Efficient Structured Reasoning

За 30 секунд

Что произошло

Penelope is a framework that performs structured reasoning inside transformer decoder layers using localized recurrent computation instead of visible chain-of-thought tokens.

Почему это важно

Matters for engineers optimizing inference latency and cost on reasoning tasks where chain-of-thought expansion significantly increases output length.

На что обратить внимание

Results shown on open-source benchmarks only. Unclear how well latent reasoning generalizes to domains beyond structured reasoning or to different model scales.