From Large Language Models to Efficient System One Models via Adaptive Computation

Transferring an LLM into a System One decision model usually replaces the language modeling head with a decision head but keeps the full backbone, so every decision pays for all layers. The paper asks how much depth such decisions actually require and how to share computation across depths instead of only shrinking the backbone.

The authors find decision information is concentrated in intermediate layers: early layers build essential features, and the final six layers can be removed without loss. Independently trained smaller models largely mimic the full model rather than developing specialized decision capabilities.

It jointly trains a full model with prefix sub-models that share its final layers and decision head through self-distillation, so one model can operate at multiple layers.

A lightweight gate reads the hidden state at each exit and determines whether additional computation is needed, allowing a single model to run at multiple depths.

Across seven decision benchmarks, the approach matches or outperforms independently trained models at every depth while achieving better accuracy-efficiency trade-offs through adaptive computation.

The material provided does not state whether code or weights are released. It describes the method, the layer-depth findings, and the seven-benchmark evaluation, and is listed as an arXiv preprint.

It is a research method for depth efficiency in decision models, not a packaged runtime or hosted endpoint. Teams wanting a ready-made SDK, adapter, or service should look elsewhere, and the reported results come from seven decision benchmarks rather than production workloads.