AnyJev
⚠️ Why not a model
Training-free: reads typed answers from any LLM’s next-token logits, averaged over option rotations (L0). No decision weights of its own.
Technical specs
- Base LLMs
- any causal LLM (Transformers, vLLM, SGLang, llama.cpp), Qwen3-8B (examples)
- Decision types
- choice, score, boolean
- Features
- training-free, logit-readout, option-order-debiasing, pypi
- License
- Apache-2.0
Links
AnyJev (GitHub, Apache-2.0, pip install anyjev, 0.3.0, technical report) wraps any open causal LLM and returns a probability for each option of a choice, yes_no or score question. Its L0 level reads the option logits once per rotation of the options and averages them, so reversing the option order no longer flips the answer. A rotation budget stops early when rotations agree; the repo reports 2.2–2.7× more decisions per second at unchanged accuracy (UnverifiedClaim). Backends: Transformers, vLLM, SGLang and llama.cpp; anyjev.serve exposes an HTTP endpoint.
Why it is here, not in the catalog. The Decider adds no trained weights. The project has since published Tacit (1.7B–9B), self-distilled models that are trained; they are on our watch list and would be scored separately.
