AnyJev
Nokia's training-free library that turns open LLMs into Jev-style deciders robust to option order, plus self-distilled Tacit models.
Overview
AnyJev, from Nokia Applied Research with collaborators, turns any open-weights LLM into a Jev-style decision model without training. It reads typed decisions and probabilities from the prefill distribution, averages option logits over permutations (cyclic shifts) and subtracts a label-free prior so answers barely change when options are reordered, reducing order-flip rate and calibration error. Results are labelled raw, L0 or L1 by the debiasing and calibration applied.
It also ships Tacit, self-distilled models from 1.7B to 9B on Hugging Face that answer in one forward pass; with adaptive=True, a capped share of low-confidence decisions is sent to the model's own reasoning. It runs on transformers, vLLM, SGLang or llama.cpp and has a server mode. The README reports the training-free Decider on unchanged Gemma-4-31B scoring 60.22 on the Decision Index 0.3, 5th of 149 entries on 2026-10-12. A paper is on arXiv.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which label fits this item (e.g. BANKING77 intent)? | — |
| Yes / no | Yes/no decision read from prefill probabilities | yes · no |
| Score | Ordered-rubric decision | — |
How it works
- Load an open model through a backend (HF, vLLM, SGLang, llama.cpp)
- Ask a Question; options are rotated and logits averaged
- Subtract the label-free prior to debias
- Return calibrated probabilities; optionally route low-confidence cases to reasoning
Strengths
- No fine-tuning required
- Reduced option-order sensitivity
- Runs locally; served via a System One-style server
- Public leaderboard result and paper
Caveats
- Multiple forward passes for rotations add cost
- Quality depends on the base model
Getting started
pip install "anyjev[hf]" then Tacit.from_pretrained("morriszjm/Tacit-9B")At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- library + server (transformers, vLLM, SGLang, llama.cpp)
- Has code
- yes
- License
- Apache-2.0
- Stars
- 1,255
Use cases and integrations
How curated lists describe it
Open research: turns open LLMs into Jev-style decision models that read typed decisions and calibrated probabilities from next-token prefill distributions with zero fine-tuning, reducing order-flip rate and calibration error.
yibie/awesome-jevNokia's training-free method for turning any LLM into a Jev-style decision model whose answers hold when option order changes, plus self-distilled Tacit models (1.7B to 9B) that can send a capped share of low-confidence decisions to the model's own reasoning. `pip install anyjev`.
AnotiaWang/awesome-decision-modelsTurn any LLM into a Jev-style decision model: typed decisions, real probabilities, no training. (continue updating, welcome any issue and PR request)
wh000wh000/awesome-jev-liveTurn any LLM into a Jev-style decision model: typed decisions, real probabilities, no training. (continue updating).
hellogumbo/awesome-jevTurns any open-weights LLM into a typed decider by averaging the option logits over permutations and subtracting a label-free prior, so the answer barely moves when you reorder the options.
MrJev/awesome-jev
Frequently asked questions
Does it need training?
No for the Decider; Tacit models are optional self-distilled checkpoints.
Why rotate options?
Averaging over permutations and removing a prior reduces sensitivity to option order.
Does it need a TypeSafe key?
No, it runs open models locally.
Related projects
Exact-target audit finding Jev's Choice probabilities are a step function on stated base rates while per-option Noul tracks them closely.
ProjectresearchPythonResearch replicating Jev-style typed decisions with open LLMs via logprobs; Qwen3.5-9B reached 94% with good calibration in ~1 s.
ProjectresearchPythonLocal reproduction of Jev-style inference that reads option-token logits from a small LLM in one pass, with an honest failure log.
ProjectresearchPythonOpen decision models that derive calibrated confidence from recurrent looped-layer dynamics instead of RL, with early exit and a Jev API.
ProjectresearchPythonLaptop-scale reimplementation of RL for Calibrated Decisions: parallel constrained decoding trained with GRPO and a Brier reward.
ProjectresearchPythonJev-style replica: frozen Qwen3-1.7B plus a ~0.8M-parameter trained head returns calibrated Choice, Boolean and Score distributions.
ServiceresearchPython