Jeva Posts 27.86x Reuse Speedup, 92.05% on BANKING77 K=8

Jeva, a decision-native architecture from Qi Li and Xinchao Wang, posts 92.05% accuracy on BANKING77 with eight candidates and a 27.86x median paired latency speedup under full option reuse, according to a preprint posted to arXiv on October 13, 2026.

The pitch starts from a complaint about how system-one classifiers are usually built. The paper argues that "recent practice of predicting an answer from serialized text does not, by itself, respect the structure of a decision problem," and names four requirements it says are underemphasized: null-aware, permutation-invariant, option-scalable, and cacheable. Jeva's shared encoder represents the request and each option independently, while "a compact decision stack refines only the request state against immutable option memory."

The numbers back the structural claims one at a time. On BANKING77 with 16 candidates and 588 answer-absent examples, null detection AUROC hits 98.66%, with a 93.88% detection rate at roughly 5% false positive. Choice evaluation reports 0.00% prediction flips under reordering across 600 examples, with a mean total variation distance of 3.17e-08 between orderings. On ToolRet the model holds 78.40% aggregate accuracy across 500 requests and 69.88% at K=2048.

The caching story is where the headline figure lives. A hybrid cache under full option reuse produces the 27.86x median paired speedup, with maximum probability difference across cache conditions of 0.00736; MLDR priming drops from 2,097.3 ms to 73.9 ms. At partial reuse the speedup is 1.95x, so the full-reuse case is the ceiling, not the typical case.

The abstract page lists no institutional affiliation and the paper publishes no comparison against an LLM-based classifier at matched latency. Jeva lands amid a run of agent-architecture work our feed has been tracking; the agents section has logged 475 stories in the last 90 days.