Build your own single-pass decision model: masking vocabulary with Qwen3-1.7B to mimic Jev
bibryam · x · 2026-10-11
Nish Tahir breaks down how "System one" decision models like Jev work and how to replicate them. Even with Structured Output, an LLM needs one forward pass per token (11 in his example) to emit valid JSON. Decision models instead assume a fixed option set (A–E), mask the vocabulary to those tokens, and pick the highest-probability answer in a single pass. The author demos this with Qwen/Qwen3-1.7B and transformers, including full code and a visualization. He also cautions that constrained output doesn't guarantee correctness, and that raw token probabilities aren't calibrated confidence scores without extra training.
More from Research
- Lab automation only handles cookie-cutter assays — the long tail is what stifles innovation — anshulkundaje · 2026-10-11
- ArXiv caps submissions at two per month as AI paper flood overwhelms server — The Decoder · 2026-10-11
- Warning: engram/n-gram methods may markedly amplify overfitting under data repeat — SonglinYang4 · 2026-10-11
- MIT's FleXray segments 60 anatomical structures in any X-ray, runs in-browser — Dr_Alex_Crimi · 2026-10-11
- OpenAI's math formalization push called 'greatest gift to mathematics' — basedjensen · 2026-10-11
- Numerical-memory method validated on DeepSeek-V2-Lite: 2.58 MiB, 69/72 correct — Nearby_Indication474 · 2026-10-11