JevCheese看懂決策模型的取捨。

公開題準確率: 86.1% · 封存題準確率: 37.0% · 差距: 49.1 pp

授權: Apache-2.0 (author adapter and server metadata)

公開題準確率: 89.6% · 封存題準確率: 38.0% · 差距: 51.6 pp

授權: Apache-2.0 (weights and code; NOTICE credits JevK5, Qwen and SemIf); base Qwen3.5-4B Apache-2.0

公開題準確率: 83.5% · 封存題準確率: 34.7% · 差距: 48.8 pp

授權: Apache-2.0 (package and weights)

公開題準確率: 86.6% · 封存題準確率: 36.7% · 差距: 49.9 pp

底層模型: closed

授權: proprietary API

公開題準確率: 85.3% · 封存題準確率: 33.1% · 差距: 52.2 pp

授權: Apache-2.0 (code/adapter); Apache-2.0 (Qwen3.5-4B base)

公開題準確率: 87.9% · 封存題準確率: 33.8% · 差距: 54.1 pp

授權: shim MIT; weights Apache-2.0 with Google's Gemma Prohibited Use Policy

公開題準確率: 82.3% · 封存題準確率: 34.1% · 差距: 48.2 pp

底層模型: Qwen3.5-4B plus HopitAI/hopper LoRA

授權: Component-specific terms recorded in RESULT.md; submitted adapter release and Qwen base retain their respective terms

公開題準確率: 85.7% · 封存題準確率: 33.1% · 差距: 52.6 pp

底層模型: google/gemma-4-12B-it LoRA fine-tune, merged and exported as Q8_0 GGUF

授權: Apache-2.0, including the applicable Gemma 4 base/derivative licence terms

公開題準確率: 79.2% · 封存題準確率: 28.2% · 差距: 51.0 pp

底層模型: Qwen/Qwen3.5-4B + kshetrajna12/reflex-qwen3.5-4b-lora

授權: MIT (code, adapter); Apache-2.0 (base)

公開題準確率: 84.0% · 封存題準確率: 29.9% · 差距: 54.1 pp

底層模型: inference method on google/diffusiongemma-26B-A4B-it (one structured denoising read), not a separately trained model

授權: Apache-2.0 code; Google DiffusionGemma Apache-2.0 weights; no djev-specific weights

公開題準確率: 88.7% · 封存題準確率: 32.1% · 差距: 56.6 pp

底層模型: google/gemma-4-12B-it fine-tuned and merged, with a trained 256-way decision head

授權: Apache-2.0, following Gemma 4; dataset rights stated separately by the author

公開題準確率: 79.7% · 封存題準確率: 27.6% · 差距: 52.1 pp

底層模型: Qwen3.5-4B merged r16 LoRA, candidate-logit readout

授權: Apache-2.0

公開題準確率: 81.0% · 封存題準確率: 26.3% · 差距: 54.7 pp

底層模型: Qwen/Qwen3.5-4B (frozen, BF16)

授權: MIT (code); Qwen3.5 weights Apache-2.0

公開題準確率: 81.0% · 封存題準確率: 25.6% · 差距: 55.3 pp

底層模型: Qwen/Qwen3.5-4B frozen BF16 backbone

授權: MIT code; Apache-2.0 weights

公開題準確率: 80.5% · 封存題準確率: 26.0% · 差距: 54.5 pp

底層模型: Qwen/Qwen3.5-4B text model, zero-shot next-token letter probabilities

授權: MIT code; Apache-2.0 Qwen weights

公開題準確率: 73.2% · 封存題準確率: 27.6% · 差距: 45.6 pp

底層模型: google/gemma-4-E2B-it + LoRA

授權: MIT (repository LICENSE; Gemma weights keep Google’s terms)

公開題準確率: 79.2% · 封存題準確率: 26.6% · 差距: 52.6 pp

底層模型: Qwen/Qwen3.5-4B with a LoRA, read at the option-letter logits with a fitted temperature (no extra head)

授權: Apache-2.0 (code and weights); base Qwen/Qwen3.5-4B Apache-2.0

公開題準確率: 74.9% · 封存題準確率: 23.4% · 差距: 51.5 pp

授權: CC-BY-NC-4.0, research use only (XNLI and RACE in the training mix)

公開題準確率: 80.1% · 封存題準確率: 28.2% · 差距: 51.8 pp

底層模型: Qwen/Qwen3-32B, bf16, read as a direct-logit classifier (no fine-tuning)

授權: Apache-2.0 (Qwen3-32B weights); serving code public

公開題準確率: 68.0% · 封存題準確率: 29.9% · 差距: 38.1 pp

底層模型: 4B instruction-aware generative reranker

授權: Apache-2.0

公開題準確率: 83.1% · 封存題準確率: 31.5% · 差距: 51.6 pp

底層模型: Qwen3.5-35B-A3B-Base with a trained decision readout, 34.7B parameters / 3B active

授權: Apache-2.0

公開題準確率: 69.7% · 封存題準確率: 27.3% · 差距: 42.4 pp

底層模型: Qwen3-4B-Instruct-2507 BF16

授權: Apache-2.0

公開題準確率: 78.4% · 封存題準確率: 26.3% · 差距: 52.1 pp

授權: MIT (repository code); Apache-2.0 (Qwen/Qwen3.5-4B base and unsloth/Qwen3.5-4B-GGUF Q4_K_M conversion)

公開題準確率: 70.1% · 封存題準確率: 28.6% · 差距: 41.6 pp

底層模型: Qwen3-Reranker-derived 4B cross-encoder

授權: Apache-2.0

公開題準確率: 67.5% · 封存題準確率: 25.6% · 差距: 41.9 pp

底層模型: decision-machine-1 (weights not published)

授權: proprietary API, closed weights

公開題準確率: 69.7% · 封存題準確率: 24.4% · 差距: 45.3 pp

授權: CC-BY-NC-4.0, research use only (XNLI and RACE in the training mix)

公開題準確率: 65.8% · 封存題準確率: 29.2% · 差距: 36.6 pp

底層模型: Phi-4-mini-instruct BF16

授權: MIT

公開題準確率: 75.3% · 封存題準確率: 29.5% · 差距: 45.8 pp

底層模型: Qwen3.5-9B Base NVFP4 with compact BF16 decision head

授權: Apache-2.0 entrant and upstream checkpoint

公開題準確率: 81.8% · 封存題準確率: 28.6% · 差距: 53.2 pp

底層模型: nvidia/diffusiongemma-26B-A4B-it-NVFP4

授權: Apache-2.0 (repo and weights)

公開題準確率: 66.2% · 封存題準確率: 22.4% · 差距: 43.8 pp

底層模型: Qwen3-4B-Base + LoRA + learned pointer head; jaredpalmer/kev-4b

授權: Apache-2.0

公開題準確率: 75.3% · 封存題準確率: 26.0% · 差距: 49.4 pp

底層模型: openbmb/MiniCPM5-2B with a trained LoRA adapter and pointer head (26.2M trainable parameters)

授權: Apache-2.0 notices on the included code and the pinned base; the weights are an evaluation preview under EVALUATION-PERMISSION.md, not a cleared commercial release

公開題準確率: 78.4% · 封存題準確率: 29.2% · 差距: 49.1 pp

授權: Apache-2.0 (adapter code and weights; Qwen3.5-9B base is Apache-2.0)

公開題準確率: 99.1% · 封存題準確率: 92.9% · 差距: 6.3 pp

授權: proprietary API

公開題準確率: 86.6% · 封存題準確率: 35.7% · 差距: 50.9 pp

底層模型: Qwen3.8-27B through SimpleJev's direct-logit classifier

授權: Apache-2.0 (Qwen weights); repository licence not stated

公開題準確率: 89.6% · 封存題準確率: 34.1% · 差距: 55.5 pp

底層模型: Qwen/Qwen3.8-Flash-Next, mixed NVFP4/FP8/INT4 NInfer artifact

授權: Qwen Community License 1.0 (model); Apache-2.0 (engine)

公開題準確率: 88.7% · 封存題準確率: 38.6% · 差距: 50.1 pp

授權: patches Apache-2.0, shim MIT; weights under the Qwen licence (LICENSE-NOTICE.md)

公開題準確率: 99.6% · 封存題準確率: 95.5% · 差距: 4.1 pp

授權: proprietary API

公開題準確率: 74.0% · 封存題準確率: 24.4% · 差距: 49.7 pp

底層模型: Qwen/Qwen3.5-4B (frozen, BF16)

授權: Apache-2.0 (code and weights)

公開題準確率: 74.9% · 封存題準確率: 29.5% · 差距: 45.3 pp

底層模型: Qwen/Qwen3.5-9B, frozen, per-option mean log-probability

授權: no licence stated in the repository (public code); Apache-2.0 base weights

公開題準確率: 63.2% · 封存題準確率: 25.6% · 差距: 37.6 pp

底層模型: Qwen/Qwen3-0.6B-Base with a trained LoRA adapter and pointer head (10.6M trainable parameters)

授權: Apache-2.0 notices on the included code and the pinned base; the weights are an evaluation preview under EVALUATION-PERMISSION.md, not a cleared commercial release

公開題準確率: 71.0% · 封存題準確率: 24.7% · 差距: 46.3 pp

底層模型: Qwen3.5-2B-Base with a trained decision readout, 1.9B

授權: Apache-2.0

公開題準確率: 62.8% · 封存題準確率: 33.1% · 差距: 29.7 pp

底層模型: GLiFormer large (knowledgator/gliformer-large-v1, ~400M) behind a TypeSafe-compatible /v1/systemone server

授權: MIT (code); GLiFormer weights per their model card

公開題準確率: 58.4% · 封存題準確率: 30.8% · 差距: 27.6 pp

底層模型: ModernBERT-large encoder + option-marker decision head, 421M, RLCD-trained

授權: Apache-2.0

公開題準確率: 92.8% · 封存題準確率: 45.8% · 差距: 47.1 pp

授權: Gemma terms of use (Google) for the base weights; hosted Autoloops service

公開題準確率: 76.6% · 封存題準確率: 26.6% · 差距: 50.0 pp

授權: Apache-2.0 (jev-adapter server, LoRA and merged weights; base Ministral 3 8B Apache-2.0)

公開題準確率: 58.4% · 封存題準確率: 25.0% · 差距: 33.4 pp

底層模型: LiquidAI/LFM2.5-350M with a LoRA and a 6.5M-parameter pointer head (a kev clone)

授權: Apache-2.0 (code); weights under LiquidAI's LFM1.0 licence, following the LFM2.5-350M base

公開題準確率: 85.3% · 封存題準確率: 33.1% · 差距: 52.2 pp

底層模型: Qwen3.6-35B-A3B

授權: no licence file in the repository as of 2026-09-19; Qwen3.6 weights keep their own terms

公開題準確率: 57.1% · 封存題準確率: 27.9% · 差距: 29.2 pp

授權: Apache-2.0 code; Apache-2.0 weights (answerdotai/ModernBERT-large base)

公開題準確率: 83.1% · 封存題準確率: 33.1% · 差距: 50.0 pp

底層模型: Qwen/Qwen3.8-27B, NInfer NVFP4/FP8 artifact

授權: Apache-2.0 (engine and submitted 27B artifact/base)

公開題準確率: 83.1% · 封存題準確率: 33.1% · 差距: 50.0 pp

底層模型: Qwen/Qwen3.8-27B, NInfer NVFP4/FP8 artifact

授權: Apache-2.0 (engine and submitted 27B artifact/base)

公開題準確率: 71.4% · 封存題準確率: 21.8% · 差距: 49.7 pp

底層模型: Qwen3-8B-Base + LoRA + learned pointer head; jaredpalmer/kev-8b

授權: Apache-2.0

公開題準確率: 89.6% · 封存題準確率: 33.8% · 差距: 55.8 pp

底層模型: Qwen3.6-35B-A3B BF16 with JevOne bidirectional option-logit mapping

授權: Component-specific JevOne/Qwen/SGLang terms recorded in RESULT.md

公開題準確率: 77.9% · 封存題準確率: 29.9% · 差距: 48.1 pp

授權: Apache-2.0 (package and weights)

公開題準確率: 81.4% · 封存題準確率: 28.2% · 差距: 53.1 pp

底層模型: Qwen3.6-35B-A3B through SimpleJev's direct-logit classifier

授權: Apache-2.0 (Qwen weights); repository licence not stated

公開題準確率: 66.7% · 封存題準確率: 24.0% · 差距: 42.6 pp

底層模型: Qwen3-0.6B-Base + LoRA + learned pointer head; jaredpalmer/kev-0.6b

授權: Apache-2.0

公開題準確率: 68.4% · 封存題準確率: 26.3% · 差距: 42.1 pp

底層模型: Qwen3-8B BF16

授權: Apache-2.0

公開題準確率: 71.9% · 封存題準確率: 24.4% · 差距: 47.5 pp

底層模型: Qwen/Qwen3-8B (frozen, BF16), the model of the author's demos

授權: no licence file in the repository as of 19 Sep; Qwen3 weights Apache-2.0

公開題準確率: 53.2% · 封存題準確率: 25.6% · 差距: 27.6 pp

底層模型: MoritzLaurer/ModernBERT-large-zeroshot-v2.0 (~400M) used as a zero-shot NLI decision engine

授權: Apache-2.0

公開題準確率: 86.1% · 封存題準確率: 30.8% · 差距: 55.3 pp

底層模型: Qwen/Qwen3.8-27B, frozen, read at the output head

授權: Apache-2.0 (code); Apache-2.0 base weights

公開題準確率: 57.6% · 封存題準確率: 27.9% · 差距: 29.7 pp

底層模型: GLiClass ModernBERT-base fine-tuned decision model, 151M; v1.4 fixed inference engine

授權: Apache-2.0

公開題準確率: 49.4% · 封存題準確率: 27.3% · 差距: 22.1 pp

底層模型: Qwen2.5-0.5B + LoRA + learned pointer head; jaredpalmer/kev-0.5b

授權: Apache-2.0

公開題準確率: 79.7% · 封存題準確率: 28.9% · 差距: 50.8 pp

底層模型: bespokelabs/Bespoke-Nimble-9B (LoRA, adapter unchanged since 93ec5d6), merged into Qwen/Qwen3.5-9B@c202236 with the author's PEFT safe-merge

授權: Apache-2.0 (weights); repository without a licence file as of 19 Sep

公開題準確率: 97.4% · 封存題準確率: 89.0% · 差距: 8.4 pp

底層模型: closed

授權: proprietary API

公開題準確率: 55.4% · 封存題準確率: 24.7% · 差距: 30.7 pp

底層模型: GLiClass ModernBERT-base (knowledgator/gliclass-modern-base-v2.0) fine-tuned, 151M

授權: Apache-2.0

公開題準確率: 54.1% · 封存題準確率: 26.0% · 差距: 28.1 pp

底層模型: Qwen3-1.7B BF16

授權: Apache-2.0

公開題準確率: 87.0% · 封存題準確率: 29.5% · 差距: 57.5 pp

底層模型: Qwen3.8-27B, frozen, direct-logit readout averaged across two option orders

授權: MIT code; Apache-2.0 Qwen weights

公開題準確率: 61.9% · 封存題準確率: 23.4% · 差距: 38.5 pp

底層模型: the author's own checkpoint model/jev1-2b-v2/checkpoint-45000-67.53 over a Qwen3.5-2B base, served by the author on his own machine; one request per decision returns a native probability for every offered option

授權: not stated (the author published only an endpoint and the evaluation client)

公開題準確率: 87.4% · 封存題準確率: 60.1% · 差距: 27.4 pp

底層模型: google/diffusiongemma-26b-a4b-it, BF16; full generation with thinking enabled

授權: Apache-2.0

公開題準確率: 56.7% · 封存題準確率: 28.6% · 差距: 28.1 pp

底層模型: GLiNER2 large schema-conditioned extractor (DeBERTa-v3-large encoder)

授權: Apache-2.0

公開題準確率: 88.7% · 封存題準確率: 42.2% · 差距: 46.5 pp

底層模型: google/diffusiongemma-26b-a4b-it, BF16; OpenJev think=512

授權: Apache-2.0

公開題準確率: 59.3% · 封存題準確率: 34.7% · 差距: 24.6 pp

底層模型: Qwen3.5-0.8B-Base fine-tuned as a JevLite decision model with per-question calibration

授權: MIT code and training data; Apache-2.0 model weights

公開題準確率: 87.0% · 封存題準確率: 38.6% · 差距: 48.4 pp

底層模型: closed

授權: proprietary API

公開題準確率: 52.4% · 封存題準確率: 29.5% · 差距: 22.8 pp

底層模型: microsoft/deberta-v3-large

授權: Apache-2.0 (model card); DeBERTa-v3 keeps its own terms

公開題準確率: 60.6% · 封存題準確率: 26.9% · 差距: 33.7 pp

底層模型: MiniCPM5-2B-Base, 2.5B dense, with LoRA and native semantic decision heads

授權: Apache-2.0

公開題準確率: 58.0% · 封存題準確率: 29.2% · 差距: 28.8 pp

底層模型: DeBERTa-v3-base schema-conditioned extractor (GLiNER2.5 boundary architecture), 194M

授權: Apache-2.0

公開題準確率: 86.6% · 封存題準確率: 35.1% · 差距: 51.5 pp

底層模型: Qwen3.8-27B (Apache-2.0 base weights) with no additional fine-tuning; their own serving stack reads candidate-token scores in one forward pass per question and normalizes them into a decision probability, with no autoregressive decoding. The serving implementation is not public.

授權: proprietary service over Apache-2.0 base weights (Qwen3.8-27B)

公開題準確率: 77.5% · 封存題準確率: 29.9% · 差距: 47.6 pp

底層模型: Qwen3.5-9B plus rank-8 LoRA and trained scalar decision head

授權: MIT (loader); Apache-2.0 (adapter and pinned Qwen base); CC0-1.0 public training projection

公開題準確率: 64.5% · 封存題準確率: 26.3% · 差距: 38.2 pp

底層模型: Qwen3.5-2B plus rank-8 LoRA and trained scalar decision head

授權: MIT (loader); Apache-2.0 (adapter and pinned Qwen base); CC0-1.0 public training projection

公開題準確率: 48.9% · 封存題準確率: 32.8% · 差距: 16.1 pp

底層模型: GLiNER2.5 multilingual schema-conditioned extractor, 287M

授權: Apache-2.0

公開題準確率: 40.7% · 封存題準確率: 24.0% · 差距: 16.7 pp

授權: Apache-2.0 (code and CLM-v0.1-8B head); Qwen3-8B encoder Apache-2.0

公開題準確率: 54.5% · 封存題準確率: 34.7% · 差距: 19.8 pp

底層模型: Qwen3.5-0.8B through SimpleJev native assistant-prefill option-logit scorer

授權: Apache-2.0 server; Apache-2.0 Qwen3.5-0.8B checkpoint

公開題準確率: 45.9% · 封存題準確率: 28.6% · 差距: 17.3 pp

底層模型: GLiNER2.5 small schema-conditioned extractor, 74M

授權: Apache-2.0

公開題準確率: 48.1% · 封存題準確率: 25.6% · 差距: 22.4 pp

底層模型: Qwen3-0.6B BF16

授權: Apache-2.0

公開題準確率: 53.7% · 封存題準確率: 28.9% · 差距: 24.8 pp

底層模型: intfloat/multilingual-e5-small (118M multilingual bi-encoder) fine-tuned on a typed-decision mix; each option is scored against the rendered state by cosine similarity, with temperature scaling and a conformal abstain set on top. One encoder pass per option, no tokens generated.

授權: Apache-2.0 (verdictml code and the Manav2op/verdict-small checkpoint over intfloat/multilingual-e5-small)

公開題準確率: 97.8% · 封存題準確率: 94.8% · 差距: 3.0 pp

底層模型: DeepSeek-V4.1-Flash

授權: open weights, proprietary API route

公開題準確率: 41.1% · 封存題準確率: 9.1% · 差距: 32.0 pp

底層模型: Mirror DeBERTa-v3-large 436M classification-span scorer

授權: Apache-2.0 wrapper and Mirror release; upstream DeBERTa/model assets retain their own terms

公開題準確率: 37.2% · 封存題準確率: 34.4% · 差距: 2.8 pp

底層模型: 494M cross-encoder

授權: Apache-2.0

公開題準確率: 39.4% · 封存題準確率: 27.9% · 差距: 11.5 pp

底層模型: 568M multilingual cross-encoder

授權: Apache-2.0

公開題準確率: 33.8% · 封存題準確率: 33.4% · 差距: 0.3 pp

底層模型: 149M ModernBERT cross-encoder

授權: Apache-2.0

公開題準確率: 31.6% · 封存題準確率: 29.5% · 差距: 2.1 pp

底層模型: ModernBERT-large with a per-option query/scoring head, ~400M parameters

授權: MIT

公開題準確率: 84.4% · 封存題準確率: 31.2% · 差距: 53.2 pp

底層模型: google/diffusiongemma-26B-A4B-it BF16, one-step JSON canvas

授權: MIT code; Apache-2.0 DiffusionGemma weights

公開題準確率: 85.3% · 封存題準確率: 34.4% · 差距: 50.9 pp

未排名: runs on Jev (TypeSafe) — listed, not ranked

底層模型: Jev (TypeSafe) behind classifier.dev's zero-shot classification API

授權: MIT (code); hosted service

公開題準確率: 71.9% · 封存題準確率: 21.8% · 差距: 50.1 pp

底層模型: Qwen3.8-27B

授權: open weights

公開題準確率: 22.1% · 封存題準確率: —% · 差距: — pp

未排名: No completed sealed v1.4 run; retained as an unranked historical row

底層模型: Needle 3 (121M parameters, 2-bit)

授權: Apache-2.0 (model and package)

公開題準確率: 22.5% · 封存題準確率: —% · 差距: — pp

未排名: No completed sealed v1.4 run; retained as an unranked historical row

底層模型: Needle 3 (121M parameters, 2-bit)

授權: Apache-2.0 (model and package)