AnyJev
A training-free layer that turns an open LLM into a decision model. It reads typed choice, yes/no and score answers from one prefill, removes option-order bias with no labels and, from a few hundred labels, calibrates or fits a closed-form head. Weights stay untouched.
Every decision carries the level that produced it. L0 needs no labels and averages out option order (on Qwen3-8B, BANKING77 20-way, answer flips under reversed options fall from 0.230 to 0.073); L1 adds a temperature from 100 to 500 labels; L2 solves a closed-form head per question on a hidden state partway down, from 100 to 300 labels. This entry is AnyJev on Qwen3-8B: on LocalLLaMA/typed-decisions L2 scores 77.1% with a calibration error of 0.034 at block 24 of 36 (Jev's published figure: 72.7%). Heads for five Qwen3 models, 1.7B to 32B (73.0% to 79.9%), ship in the GitHub repository rather than on Hugging Face, and the Qwen weights come from Qwen, so there are no files to pull here. A head serves one question on one model, and the letter readout takes at most 26 options. Code and heads are Apache-2.0. By Jiamu Zhang, Tianze Yang and Liang Wu (Nokia) with Yucheng Shi (Tencent Hunyuan).
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one calibrated probability
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 8B |
| Base model | qwen/qwen3-8b |
| Maker | Nokia Applied Research |
| Released | 2026-09-21 |
| License | apache-2.0 |
| Reported accuracy | 77.1% |
| Reported latency | 33.9 ms per decision alone, 7.5 ms batched, on an H100 NVL (110-token state, block 24) |