Jeva-4B prints 0 flips on 600 Choice permutations, and BANKING77 at K=16 is 91.62 against Jev 82.50

Hugging Face DailyPapers posted Jeva on October 14, 2026, at 20:10 UTC. The paper prints 0 flips on 600 Choice items and BANKING77 at 16 candidates 91.62 against Jev-1.13.0 at 82.50. Apache-2.0 on Qwen3.5-4B. This desk did not run the weights.

DailyPapers posted Jeva on October 14, 2026, at 20:10 UTC. The text names a decision-native architecture that is null-aware, permutation-invariant, option-scalable, and cacheable, 2,048 candidates, zero decision flips under tested permutations, and 20x latency speedups via reuse. The share image on this page is that post’s still. The follow-up in the same minute links the paper, LIQIIIII/Jeva-4B, and the project page.

Jürgen R Plasser quoted that thread on October 25, 12:54 UTC, and wrote that Jeva works on his local machine.

arXiv 2610.18773 is Qi Li and Xinchao Wang, National University of Singapore. Submitted 13 Oct 2026, 17:31:09 UTC. v2 18 Oct 2026, 16:36:21 UTC. License CC BY 4.0. The abstract says Jeva unifies Choice, Noul, and Score.

What the manuscript prints

The HTML manuscript, opened October 26, names Jeva-4B on Qwen3.5-4B with one decision block. Rank-64 LoRA plus both projections and the decision head is 13.37M trainable parameters.

Table 1 first-argmax accuracy, including null where applicable: Jeva-4B Typed Test 78.30, BANKING77 K=16 91.62, MASSIVE K=16 90.22, Mind2Web cross-task 79.99. Jev-1.13.0 on that table is 74.10, 82.50, 81.61, and 66.26. InternDecision-4B Typed Test is 78.70.

600 Choice items in original, reversed, and up to three seeded random orders: Jeva 0 flips, mean TV 3.17e-8, original and all-order accuracy 76.67%. Jev-1.13.0 flip rate 17.00%, all-order accuracy 66.67%. InternDecision-4B flip rate 22.67%.

BANKING77 K=16 null-detection AUROC is 98.66% on 3,080 matched requests, 588 answer-absent. It detects 93.88% of absent answers at 5.02% FPR.

At K=2048 on a ToolRet diagnostic, Jeva prints 69.88% accuracy and 12.040 s p50 with length-aware batching. On 50 matched MLDR queries at full option reuse, Hybrid cache p50 falls from 2,102.0 ms to 75.1 ms, 27.98 times. The paper says no decisions change. This desk did not run the weights.

Weights

The Hugging Face card, opened October 26, lists Apache 2.0 for the decision head, adapter, and bundled Qwen3.5-4B backbone. GitHub LiQiiiii/Jeva is Apache-2.0. The card lists Choice, Noul, and Score. Load is Jeva.from_pretrained("LIQIIIII/Jeva-4B"). No hosted POST on the pages this desk opened.

91.62 is BANKING77 at K=16 on Table 1. 78.30 is Typed Test. 0 flips is the 600-item permutation test. 75.1 ms is MLDR Hybrid at full reuse. Those four stay on their own rows.

This site's reading

Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.

Verify

DailyPapers, @HuggingPapers, October 14, 2026, 20:10 UTC, status 2110463377029939703. The text names Jeva as a decision-native architecture, null-aware, permutation-invariant, option-scalable, and cacheable, 2,048 candidates, zero decision flips under tested permutations, and 20x latency speedups via reuse. The share image on this page is that post's still. The follow-up in the same minute links paperswithcode.co/paper/2610.18773, huggingface.co/LIQIIIII/Jeva-4B, and liqiiiii.github.io/Jeva/. Jürgen R Plasser, @__thetaphipsi, October 25, 12:54 UTC, status 2114339923813872012, quoted that thread and wrote that Jeva works on his local machine.

arXiv 2610.18773, Qi Li and Xinchao Wang, National University of Singapore. Submitted 13 Oct 2026, 17:31:09 UTC. v2 18 Oct 2026, 16:36:21 UTC. License CC BY 4.0. The abstract says Jeva unifies Choice, Noul, and Score. Across evaluations against Jev and open source models, it prints near-perfect null detection and no decision flips under any tested permutations.

The HTML manuscript, opened October 26. Jeva-4B uses Qwen3.5-4B with pooled width 2560, decision width 768, L=1, H=8, and f=3072. Rank-64 LoRA plus both projections and the decision head, 13.37M trainable parameters. Table 1 first-argmax accuracy, including null where applicable: Jeva-4B Typed Test 78.30, BANKING77 K=16 91.62, MASSIVE K=16 90.22, Mind2Web cross-task 79.99. Jev-1.13.0 on that table is 74.10, 82.50, 81.61, and 66.26. InternDecision-4B Typed Test is 78.70. 600 Choice items in original, reversed, and up to three seeded random orders: Jeva 0 flips, mean TV 3.17e-8, original and all-order accuracy 76.67%. Jev-1.13.0 flip rate 17.00%, all-order accuracy 66.67%. BANKING77 K=16 null-detection AUROC 98.66% on 3,080 matched requests, 588 answer-absent. At K=2048 on a ToolRet diagnostic, Jeva 69.88% accuracy and 12.040 s p50 with length-aware batching. On 50 matched MLDR queries at full option reuse, Hybrid cache p50 falls from 2,102.0 ms to 75.1 ms, 27.98 times. No decisions change. This desk did not run the weights.

The Hugging Face card LIQIIIII/Jeva-4B, opened October 26. License Apache 2.0 for the decision head, adapter, and bundled Qwen3.5-4B backbone. GitHub LiQiiiii/Jeva is Apache-2.0. The card lists Choice, Noul, and Score. from_pretrained("LIQIIIII/Jeva-4B"). No hosted POST on the pages this desk opened.

Compare

91.62 is BANKING77 at K=16 on Table 1, against Jev-1.13.0 at 82.50 on that same table. 78.30 is Typed Test, 2,000 requests, against InternDecision-4B at 78.70. Do not merge 91.62 with 78.30.

0 flips is 600 Choice items under tested permutations. Jev-1.13.0 on that test is 17.00%. It is not a JevBench composite and not Decision Index Full score.

98.66% is null-detection AUROC on BANKING77 with 16 candidates. 69.88% is ToolRet at 2,048 options, and it is also the all-size accuracy on 83 nested queries. 75.1 ms is MLDR Hybrid at full reuse. 2,102.0 ms is Cache-Off on that same stream. Do not merge 75.1 ms with 12.040 s.

GLiDE v1.2's live Decision Index 0.4.1 Full 71.5 is a different board. mammoth-1's card 71.24 is edition 0.4. This desk did not run Jeva on either index.

Terms

Jeva: Qi Li and Xinchao Wang's System One architecture and the Jeva-4B checkpoint, posted October 14, 2026. Apache-2.0, from Qwen3.5-4B. The paper prints 0 flips on 600 Choice items and BANKING77 at K=16 91.62 against Jev-1.13.0 at 82.50.
0 flips: Jeva's printed decision-flip rate on 600 Choice items under original, reversed, and seeded random candidate orders. Jev-1.13.0 on that test is 17.00%.