Awesome Latent Refinement
A curated reading list of research on models and agents that improve reasoning and planning by iteratively refining an internal latent state rather than producing explicit intermediate text, spanning supervised reasoners and reinforcement-learned planning.
This list organizes research on iterative refinement of latent state, where performance improves not from explicit intermediate outputs but from repeatedly updating internal representations through shared or recurrent computation. It connects supervised reasoning work with reinforcement-learned planning work, and tracks systems where extra inference-time compute improves results.
The supervised section covers dedicated reasoner architectures built around iterative latent-state updates, including hierarchical and recursive reasoning models, looped language models that reuse internal computation across passes, and vision-language-action models. The reinforcement-learning section covers model-free planning, recurrent reasoning agents, and how extra compute is spent inside reinforcement learning. The list states its inclusion criteria and links to related curated lists.
CC0-1.0 licensed.