nokia-applied-research/AnyJev
AnyJev turns open LLMs into calibrated typed-decision endpoints without generating text.
The repo describes levels from raw logits to L2, where a small closed-form head reads hidden states partway through a model. Its examples say L2 can run through vLLM as an embed server plus a lightweight head, with no parsing or patched engine. On the cited BANKING77 and typed-decisions tests, the authors report better calibration and higher auto-decidable coverage than raw logits. The limitations are explicit: L2 is per question and per model, only Qwen3 heads ship, and the agent-loop evaluation is still on the roadmap. GitHub · LLM repos' note
The repo describes levels from raw logits to L2, where a small closed-form head reads hidden states partway through a model. Its examples say L2 can run through vLLM as an embed server plus a lightweight head, with no parsing or patched engine. On the cited BANKING77 and typed-decisions tests, the authors report better calibration and higher auto-decidable coverage than raw logits. The limitations are explicit: L2 is per question and per model, only Qwen3 heads ship, and the agent-loop evaluation is still on the roadmap. GitHub · LLM repos' note
score 4