RSI-Jev v2.1 opens a 2B decision model trained by a self-improving research loop
On 27 September 2026, Shanghua Gao published RSI-Jev-v2.1-2B on Hugging Face, with code and versioned experiment logs at github.com/Shanghua-Gao/RSI-Jev. It is the third public release of a recursively self-improving research system that trains Jev-style System One models: agents propose arms, register predictions before GPU spend, and retire champions when the evidence says so.
v2.1 is a full fine-tune of Qwen3.5-2B-Base with a trained option-scoring head and a confidence head (oof_head_scorefloor). It answers choice, noul and score in one forward pass and ships a local server on the public /v1/systemone shape. Weights Apache-2.0; code MIT.
What the author reports (UnverifiedClaim): suite mean 0.7291 across twelve benchmarks (ECE 0.0556), typed-decisions 0.7905, and gains over v1.0 on three frozen held-out sets that never entered training. About 10 ms per decision. Eleven of the twelve suite benches contributed train-split data, so those headline numbers are in-domain — v1.0 remains the zero-shot reference.
Catalog entry: /models/rsi-jev. Nearby peers: Jev-Style, Decider, Decily, Eikos.