Penelope: Localized Latent Recurrence for Efficient Structured Reasoning
Penelope moves extra reasoning work into a localized recurrent path inside a decoder-only Transformer.
The paper says the model evaluates the lower decoder once, builds a problem-conditioned boundary memory, then refines selected internal states before generating an answer. Its curriculum transfers chain-of-thought behavior into that latent route, avoiding long visible reasoning traces. On open-source structured-reasoning benchmarks, the authors report competitive accuracy against established latent-reasoning models at validation-selected latent budgets, with lower measured inference latency. ArXiv · AI/CL/LG's note
The paper says the model evaluates the lower decoder once, builds a problem-conditioned boundary memory, then refines selected internal states before generating an answer. Its curriculum transfers chain-of-thought behavior into that latent route, avoiding long visible reasoning traces. On open-source structured-reasoning benchmarks, the authors report competitive accuracy against established latent-reasoning models at validation-selected latent budgets, with lower measured inference latency. ArXiv · AI/CL/LG's note
score 5