Fastino's looped GLiDE tops decision leaderboard
Fastino says that running some of its GLiDE decision model's layers more than once, with no new weights, took it to first place on Hugging Face's Decision Index, evidence that latent-space reasoning can improve accuracy.
Fastino reports that Recurrent Looped GLiDE scores 73.37 on the Hugging Face Decision Index. That puts it first of 274 models, at 112 ms per decision. The company says the result is 1.5 points ahead of the best model of similar size and 5.4 ahead of Perplexity's 27B decision model. It also beats Fastino's previous leader, GLiDE V1.2, which scored 71.52. The leaderboard's maintainers score every model themselves on test questions that developers never see.
Decision models take a context, a question and a fixed set of options. They return a probability for each option without writing any reasoning text. Typical jobs include routing requests and choosing which tool to call. Fastino says this makes them a cleaner setting than chat models for isolating gains from latent reasoning. The looped model runs two middle bands of layers more than once with shared weights. The extra compute is fixed and known in advance, memory use stays the same, and the answer still comes from one forward pass.
Fastino reports three main findings:
- Scale: The benefit of extra depth grows with model size across five sizes. It is absent below 1B parameters and clearest at 27B.
- Loop design: How the loop is built matters more than whether there is one.
- Training: Training a loop directly didn't work. Four trainable variants failed or produced gains that weren't significant, and one diverged within a few hundred steps.
The company says a single-pass version also ranks first on the Artificial Analysis Decision Index. The post comes after The Information reported in September that OpenAI's GPT-6 Astra is a looped transformer. The excerpt available here doesn't explain the loop setup that finally worked, and the results are Fastino's own.