RSI-Jev v3.0: first release where listwise RL earns its place
On 28 September 2026, Shanghua Gao published RSI-Jev-v3.0-2B with the v3.0 version record. The author frames it as the first release where reinforcement learning works: a listwise ranking reward (Plackett-Luce / NDCG@5) on top of a wider supervised corpus and an MLP readout combine.
Author suite mean 0.756 (v2.1 was 0.736 on the fifteen-benchmark suite), typed-decisions 0.791, suite ECE 0.066, hippo R@1 0.192 → 0.308 — all UnverifiedClaim. Same local /v1/systemone server path as v2.1. GitHub stars roughly 13 → 20 between the 28 Sep and 29 Sep checks.
Catalog entry (refreshed): /models/rsi-jev. Prior news: v2.1.