Skip to main content
Back to timeline
arXivSource publication:

Jeva splits decisions out of text generation: reordering candidates no longer flips the answer, and 2,048 options are scored in one pass

Synopsis

For requests that need a direct answer, such as tool selection or document identification, Jeva encodes the request and each candidate separately, then uses a compact decision module to refine only the request state, so candidate representations can be reused across requests. On 600 Choice examples in original, reversed and up to three seeded random orders, Jeva records zero decision flips, while Jev-1.13.0, InternDecision-4B and JevK5 flip at 17.00%, 22.67% and 26.17%. On 3,080 BANKING77 requests, its null-detection AUROC is 98.66%, above KEV-BF16's 93.61%.

AI-generated editorial illustration: Jeva: System One Models from First Principles

Interpretation

System one models, which emit a choice, judgment or score directly instead of generating reasoning text, can now keep the decision fixed when candidates are reordered and avoid redefining the output space as the candidate set grows. Earlier open-source system one models concatenate request and candidates into one text and read the decision from answer-token logits or a task-specific prediction head, so candidate content sits inside request-specific computation and candidate position can influence predictions. On 600 Choice examples evaluated in original order, its reversal and up to three seeded random orders, aligned by candidate identity, Jeva records zero flips and identical original and all-order accuracy of 76.67%; Jev-1.13.0, InternDecision-4B and JevK5 flip at 17.00%, 22.67% and 26.17%.

Jeva uses a shared language encoder to map the request and each option into vectors independently, then a compact decision module refines only the request state while option vectors stay fixed within a forward pass. Interactions among candidates no longer have to be expressed by independent scoring: a single decision block can reverse the preference between two retained options when the surrounding set changes, which an independent scorer cannot do. On 500 ToolRet requests, Jeva scores the complete candidate set in one decision-head pass and reaches 78.40% aggregate accuracy, versus 64.00% for recursive JevK5; on 83 queries whose candidate sets expand from 32 to 2,048 options, Jeva's all-size accuracy is 69.88%, 12.05 points above JevK5.

Answer absence and uncertainty among candidates are modeled separately: a coverage gate allocates probability between null and the candidate set, and competition inside the set is resolved on its own. Normalizing scores over the supplied candidates only says which candidate ranks higher; it cannot establish whether any acceptable candidate exists. On 3,080 matched BANKING77 requests with 16 candidates, 588 of them answer-absent, Jeva reaches 98.66% null-detection AUROC versus 93.61% for KEV-BF16, 84.27% for InternDecision-4B and 79.44% for JevK5; it detects 93.88% of absent answers at 5.02% FPR.

Option-side semantic vectors and layer projections can be cached across requests, including newly assembled candidate sets, while request encoding and decision refinement still run per request. When candidates sit inside request-specific computation, changing the question or recombining candidates requires processing their joint context again. On 50 matched MLDR queries, hybrid caching at full option reuse cuts p50 latency from 2,102.0 ms to 75.1 ms; on a shared candidate bank, Typed Decisions p50 falls from 93.1 ms to 46.6 ms (exact-set cache) and 46.7 ms (hybrid cache), avoiding 99.2% of candidate-row encodings, with no decision changes and a maximum absolute probability difference of 0.00736.

Perspective

The evaluation covers Typed Decisions, BANKING77, MASSIVE, three Mind2Web slices, ToolRet and MLDR, using a Qwen3.5-4B backbone with one decision block and synchronized timing on a single H200; the conclusions apply to these candidate-selection, null-judgment and scoring settings.

The permutation-invariance and cache-equivalence proofs assume exact arithmetic, and floating-point reductions or batching changes can perturb nearly tied scores; when maxima tie, an implementation returning the first maximizing index need not select the same tied candidate after reordering. Cache equivalence also requires unchanged encoder and projection weights, tokenizer, templates, length policy and numeric configuration, and memories must be invalidated after parameter updates.