DIY-Jev: no fine-tuning needed — reading true/false logits hits 75.5% on 32k benchmark
Malfeitor1235 · reddit · 2026-09-21
Developer VakeDomen released DIY-Jev, a Jev-style inference setup that needs no NLI fine-tuning or classifier head and works with completely unmodified open-weight LLMs.
Method: each candidate answer becomes a boolean verification — after building a <state>/<question>/<options>/<candidate> prompt, instead of generating text the code reads the true/false logits, subtracts them per candidate, and softmaxes the differences. The expensive prefix is evaluated once; candidate branches (differing only in the last tokens) are batched through llama.cpp.
Results on a 32,235-example benchmark (laptop RTX 5090 24GB):
- Qwen3-4B: 65.0% at 27 req/s
- Qwen3-27B: 75.3% at 2.9 req/s
- Qwen3.6-35B-A3B: 75.5% at 5.3 req/s
Notably, the same model with zero training outperforms the OpenJev fine-tune. The author cautions it's not a strict Jev reproduction and the benchmark isn't perfectly apples-to-apples. Code: GitHub DIY-Jev.
More from Research
- minpts-check: one-run predictor tells if your HDBSCAN clusters survive min_samples changes — leland_mcinnes · 2026-11-12
- Architecture study sparks monitoring concerns over a multi-hop planning spike — ChowdhuryNeil · 2026-11-12
- SGLang Megakernel Pushes Muse Spark to 1,109 tok/s, 2.1x Over Baseline — hsu_byron · 2026-11-12
- ArcGrav: a cross-model Agent Control Language to cut token overhead — ArchitecturalGravity · 2026-11-12
- Open-source dual-arm robot solves tangram puzzles with 90 demos, ~60% success — DanGrover · 2026-11-12
- FutureHouse's deslopper trains Kosmos to write indistinguishably from human experts — proteinrosh · 2026-11-12