qwen-rlcd
Prototype Jev-style decision model on Qwen3.5-0.8B using prefix-fork execution to answer typed questions in one pass without generation.
Overview
qwen-rlcd is a prototype System One decision model inspired by Jev. It takes a state and typed questions and returns probability distributions over fixed answers in a single forward pass. The README separates public facts from the authors' own hypotheses, noting TypeSafe has not published Jev's architecture. It targets Qwen/Qwen3.5-0.8B-Base and is currently inference-only; training is on hold.
Because Qwen3.5 mixes Gated DeltaNet linear-attention and full-attention layers, a tree attention mask cannot isolate branches, so the repo uses prefix-fork execution: prefill the state once, copy the cache per branch, and run every question-plus-answer branch as one batch, making results independent of question and option order. Built pieces include the fork engine (8-19x faster than one forward per branch), a request schema with typed Choice/Score/Noul answers, several scorers (letter, likelihood, PMI, a read head), invariance tests, metrics (accuracy, NLL, Brier, ECE, MAE) and zero-shot baselines on ARC-Challenge, BoolQ and SST-5 at n=100 on CPU.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Multiple-choice science questions (ARC-Challenge) | — |
| Yes / no | Yes/no reading-comprehension questions (BoolQ) | — |
| Score | Five-level sentiment (SST-5) | — |
How it works
- Prefill the state once and keep its cache
- Expand the cache along the batch dimension, one row per branch
- Run all question+answer branches in one batch and read hidden states
- Turn scores into typed answers via softmax/sigmoid with entropy-based confidence
Strengths
- Order and fan-out invariant by construction
- 8-19x faster than separate forwards per branch
- Clear separation of public facts and hypotheses
Caveats
- Inference only; training and calibration fitting not started
- Zero-shot numbers are small-sample CPU runs
- No licence stated
At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- PyTorch
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Jev-style calibrated decision model (Choice/Score/Noul) on Qwen3.5-0.8B.
wh000wh000/awesome-jev-liveJev-style calibrated decision model (Choice/Score/Noul) on Qwen3.5-0.8B.
hellogumbo/awesome-jev
Frequently asked questions
Is this Jev?
No; it is an independent prototype based on the authors' hypothesis of how a Jev-like model could work.
Why prefix-fork instead of a tree mask?
Qwen3.5's recurrent DeltaNet layers cannot isolate sibling branches with a mask, so the cache is copied per branch instead.
Is it trained?
Not yet; current results are zero-shot and training is on hold.
Related projects
- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++MLX library that turns an existing local model into a decision engine by reading candidate logits, with no fine-tuning, on Apple Silicon.
ProjectprototypePythonPython library that turns an open LLM into a Jev-style classifier via constrained single-token choice indexes in one forward pass.
ProjectprototypePythonOpen Jev-style decision model built on Gemma 3 270M with a scoring head, scoring all options in one forward pass on CPU.
ProjectresearchPythonOfficial code for the FastJEV paper: training-free reduction of redundant candidate computation in multimodal JEV-style models.
ProjectresearchPython