Penelope Localizes Latent Reasoning to a Narrow Decoder Interval Instead of Emitting Chain-of-Thought Tokens

arXiv 2607.25915·low signalXBlueskyLinkedInCopy link

Penelope (arXiv 2607.25915, 2026-07-28) offers a third option for buying extra reasoning computation beyond scaling parameters or serializing steps as chain-of-thought tokens, both of which tie cost to either model size or output length. It evaluates the lower decoder prefix once to build a problem-conditioned boundary memory, then iteratively refines it through time-modulated GRU dynamics and recurrent readout states before generating the answer, with a progressive CoT-to-latent curriculum transferring visible reasoning into the internal recurrent path. Because refinement is confined to a selected decoder interval, extra computation does not require repeatedly executing the full decoder. At validation-selected latent budgets it matches established latent-reasoning models on open-source structured-reasoning benchmarks while reducing measured inference latency.