PaperlayerView exact revision on arXiv

Source-linked AI summary

Jeva: System One Models from First Principles

Qi Li, Xinchao Wang

arXiv:2610.18773v1cs.AI

TL;DR

System-one models need to make compact decisions without treating alternatives as an ordered, fixed output space or conflating candidate competition with answer absence. Jeva independently encodes requests and options, then refines the request state against reusable option memory. Across evaluations, it combines strong typed-decision performance with near-perfect null detection, stable tested permutations, large-set accuracy, and cross-request latency savings.

  • Problem

    Many system-one designs jointly encode requests and candidates, which can entangle reusable candidate content with request computation, expose decisions to candidate order, limit candidate counts, and leave answer availability unresolved.

    Source: §1 · p. 2

  • Method

    Jeva independently encodes the request and each option, refines only the request state against immutable option memory, and uses symmetric scoring with a separate coverage gate.

    Source: §1 · p. 2

  • Results

    Jeva combines strong typed-decision performance with nearly perfect null detection and stable choices under all tested permutations, and improves accuracy and latency at 2,048 options over recursive selection.

    Source: §1 · p. 3, §5 · p. 9

  • Takeaways & Limitations

    Jeva connects explicit null probability, stable predictions under reordering, variable candidate counts, and reusable option computation in a decision-native architecture.

    Source: §5 · p. 9

    Each text has its own encoder length budget, and emitting a dense distribution retains an Ω(K) cost in the candidate count.

    Source: §3.4 · p. 6

Abstract

from arXiv · show

System one models enable direct, low-latency decision-making. Yet recent practice of predicting an answer from serialized text does not, by itself, respect the structure of a decision problem. We argue that current designs underemphasize four important structural requirements. (i) null-aware: Models should distinguish answer absence from uncertainty among candidates; (ii) permutation-invariant: Reordering alternatives should not change the decision; (iii) option-scalable: Growing candidate sets should not require redefining the output space; (iv) cacheable: Request-independent computation over recurring alternatives should be reusable across requests. We introduce Jeva, a decision-native architecture that addresses these requirements jointly. A shared encoder independently represents the request and each option, while a compact decision stack refines only the request state against immutable option memory. Symmetric aggregation yields permutation-equivariant probabilities; shared scoring accommodates variable-size candidate sets. A separate coverage estimate distinguishes answer absence from candidate competition, enabling explicit null outcomes. Request-independent option representations and layer projections support reuse across newly assembled sets. These mechanisms unify Choice, Noul, and Score outputs. Across evaluations, Jeva combines strong typed-decision performance with near-perfect null detection and no decision flips under any tested permutations. It scales to thousands of candidates and substantially reduces latency through cross-request reuse without changing decisions.

1 Introduction

Jeva addresses structural limits in direct system-one decisions by separating reusable option representations from set-dependent judgment. Its evaluation reports strong typed-decision performance alongside null detection, permutation stability, large-set accuracy, and latency gains.

Motivation

  • Direct system-one decisions must handle reordered and variable-size candidate sets, recognize when no option applies, and reuse recurring alternatives to reduce latency.The introduction describes tool choice, document identification, condition testing, and scoring as examples of compact decisions.Source: §1 · p. 2

Problem

  • Jointly encoding requests and candidates can require recomputing their context, expose predictions to candidate order, restrict inventory size, and leave answer availability unresolved.The proposed architecture separates semantic representation, candidate interaction, and answer availability while retaining context-sensitive judgment.Source: §1 · p. 2

Architecture

  • A shared encoder represents requests and options independently, while a compact decision module refines only the request state against immutable option memory.This supports request-independent caching while allowing set-dependent judgment; symmetric aggregation and shared scoring provide permutation-equivariant probabilities and variable-size candidate sets.Source: §1 · p. 2

Evaluation

  • Jeva combines strong typed-decision performance with nearly perfect null detection and stable choices under all tested candidate permutations.At 2,048 options, native full-set scoring has higher accuracy and over 2× median-latency speedup over recursive selection; reuse gives over 20× speedup for newly assembled long-document sets without changing decisions.Source: §1 · p. 3

2 Related Work

Prior system-one approaches use several readouts over request and candidate representations, while related work supplies foundations for independent encoding, set modeling, and abstention. Jeva combines reusable candidate representations with set-conditioned decisions and a separate coverage estimate.

Direct and typed decision models

  • Jeva separates independent text encoding from compact set-conditioned state updates, reusing candidate computation while retaining interactions among alternatives.Its request state is refined against the full candidate set before scoring and coverage estimation.Source: §2 · p. 3
  • SemIf and JevK5 read option probabilities from answer-token logits, Intern-Decision uses decision markers, and KEV and Laya use dedicated decision readouts.ReFlex aggregates predictions across option orders to reduce candidate-position sensitivity.Source: §2 · p. 3

Independent encoding and late interaction

  • Dense Passage Retrieval and ColBERT illustrate independent encoding and late interaction, while Jeva applies reusable representations to typed Choice, Noul, and Score decisions.Jeva refines the request against the full candidate set before scoring and coverage estimation.Source: §2 · p. 3

Set functions and latent attention

  • Deep Sets, Set Transformer, and Perceiver provide related foundations for symmetric set modeling, attention-based interactions, and compact latent processing.Jeva uses symmetric aggregation and shared scoring for permutation-equivariant probabilities and variable-size sets; refinement and readout scale linearly with candidate count over prepared memory.Source: §2 · p. 3

Null awareness and calibration

  • Jeva’s coverage gate distinguishes absence of a valid candidate from uncertainty among alternatives, complementing prior work on answerability and abstention.Post-training temperature scaling is described as complementary probability calibration.Source: §2 · p. 3, Figure 4 · p. 4

3 Jeva: A Native Decision Architecture

Jeva separates reusable option encoding from set-dependent judgment: it refines a request state against immutable option memory, then scores candidates and estimates coverage.

3 Jeva: A Native Decision Architecture

  • Jeva independently encodes requests and options, then refines only the request state against immutable option memory for set-dependent scoring.The shared encoder and compact decision module support typed Choice, Noul, and Score outputs.Source: §3 · p. 4, §3.1 · p. 4, §3.3 · p. 6

3.1 Independent Encoding, Collective Decisions

  • With independently encoded options, Jeva conditions scoring on the full set, allowing retained-option preferences to reverse; option-bank hits rise from 0/3 to 2/3 to 3/3 across new sets.Every whole-set lookup misses, while state encoding and decisions remain fresh.Source: §3.1 · p. 4, Figure 5 · p. 4

3.2 Candidate-Conditioned State Refinement

  • A single decision block can uniquely reverse the preferred retained option across candidate sets, which independent scoring cannot do; the mixer avoids a K × K interaction matrix.The construction preserves a unique winner over the added candidate.Source: §3.2 · p. 5

3.3 Competition, Coverage, and Typed Outputs

  • Jeva separates conditional candidate probabilities from coverage, combining them into open-set probabilities over candidates and null while retaining candidate-only probabilities for closed sets.At fixed final scores, adding options below the existing maximum leaves coverage unchanged; set refinement can nevertheless alter scores and the joint argmax.Source: §3.3 · pp. 5–6

3.4 Symmetry, Reuse, and Scaling

  • With fixed weights and deterministic exact arithmetic, candidate permutations reorder candidate probabilities but leave null probability and typed Noul/Score outputs unchanged; prepared-memory decision cost is linear in K.Replacing option computations with identical cached entries also leaves outputs unchanged.Source: §3.4 · p. 6

3.5 Direct Supervised Learning

  • The primary open-set training loss is joint cross-entropy on the candidate-and-null distribution, with separate closed-set candidate cross-entropy and stage-dependent Noul and Score auxiliaries.The supplied recipe uses rank-64 LoRA with projections and the decision head.Source: §3.5 · pp. 6–7

4 Experiments

The experiments evaluate decision quality, structural properties, and efficiency using controlled permutation, batching, and cache tests. They report results across typed decisions, null detection, candidate scaling, and cross-request reuse.

4.1 General and typed decision quality

  • Jeva leads seven of eight full-coverage accuracy slices, exceeds 90% on BANKING77 and MASSIVE at both candidate counts, and reaches 78.30% on Typed Decisions versus InternDecision-4B’s 78.70%.For Noul, Jeva has the lowest Brier (0.0184) and MAE (0.0939); for Score, it has the lowest MAE (0.2315), while its RMSE is 0.3222 versus InternDecision-4B’s 0.3208.Source: §4.1 · pp. 7–8
  • Jeva records zero flips across tested candidate orders and identical original and all-order accuracy of 76.67%; baseline flip rates range from 17.00% to 26.17%.The evaluation uses 600 Choice examples in original, reversed, and up to three seeded random orders, aligning candidate identities.Source: §4.1 · p. 8

4.2 Null detection

  • Jeva achieves 98.66% null-detection AUROC on 3,080 matched BANKING77 requests, 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.Source: §4.2 · p. 8

4.3 Candidate scaling and batching

  • At K = 2,048, Jeva achieves 69.88% accuracy versus JevK5’s 67.47% at 28 s; length-aware batching reduces Jeva p50 latency from 94.459 to 12.040 s.Figure 8 reports accuracy and coverage across six disjoint candidate-count cohorts; unsupported slices are excluded from accuracy comparisons.Source: §4.3 · p. 8, Figure 8 · p. 8
  • Jeva remains correct at every size for 58 of 83 nested-set queries, versus 48 for JevK5 and 7 for CLM; its all-size accuracy is 69.88%, 12.05 points above JevK5.The nested candidate sets expand from 32 to 2,048 options.Source: Figure 7 · p. 7, §4.3 · p. 8

4.4 Reuse across states and candidate sets

  • At full option reuse, Hybrid reduces MLDR p50 latency from 2,102.0 to 75.1 ms, with a median paired speedup of 27.86×; decisions do not change across cache conditions.For shared banks, Typed Decisions p50 falls from 93.1 ms to 46.6 ms with Set caching, while 99.2% of candidate-row encodings are avoided.Source: Figure 9 · p. 8, §4.4 · p. 9

5 Conclusion

Jeva is presented as a system-one architecture for decisions, linking set-dependent judgment with reusable option representations. The conclusion highlights broad decision capability, null detection, stability under reordering, and cache savings.

5 Conclusion

  • Jeva is described as a native architecture specifically designed for system-one decision making.Its large-set results are presented as evidence for the practical value of candidate-indexed readout.Source: §5 · p. 9
  • Jeva models decisions as an evolving state reading reusable option memory, connecting permutation symmetry, variable candidate counts, explicit null probability, and caching.The trained Jeva-4B combines broad decision capability with strong null detection, stable predictions under reordering, and savings from hybrid cache.Source: §5 · p. 9
  • The conclusion identifies hybrid caching and candidate-indexed readout as practical aspects of Jeva’s decision architecture.The conclusion does not state that every system-one task will benefit equally.Source: §5 · p. 9

A Derivations and Structural Properties

The derivations characterize candidate reading as entropy-regularized evidence allocation. They distinguish intermediate attention weights from final probabilities and explain why separate temperatures govern different stages.

A.1 Entropy-Regularized Candidate Reading

  • The attention weights are the unique entropy-regularized allocation of attention over candidate evidence.The derivation establishes the result using the nonnegativity of KL divergence.Source: §A.1 · p. 11
  • Candidate-read weights are intermediate evidence-aggregation weights, not final decision probabilities.The state may gather evidence from several options before committing to one.Source: §A.1 · p. 11
  • Separate learned temperatures control within-block evidence aggregation, final candidate competition, and the answer-availability gate.Their numerical values need not be directly comparable because the corresponding similarity scores have different scales.Source: §A.1 · p. 11

A.2 Geometry and Single-Block Expressivity

The appendix explains how candidate evidence can alter another option’s score without updating the stored option vectors. It also distinguishes coverage decisions from candidate competition as the set grows.

A.2 Geometry and Single-Block Expressivity

  • A single decision block can represent set-dependent preferences that an independent scorer with set-independent tie-breaking cannot realize.The proposition establishes that the same request can uniquely select A in one candidate set and B in another.Source: §A.2 · p. 11
  • Evidence selected from one candidate can affect another candidate’s score through learned geometry without updating either candidate.The multi-head mixer computes the weighted message without materializing a K × K interaction matrix.Source: §A.2 · pp. 11–12

A.3 Coverage and Candidate Count

  • At fixed final scores, adding options below the existing maximum leaves coverage unchanged, removing a compulsory log K contribution to the coverage decision.Set refinement may still change scores, while competition redistributes probability across options; coverage and candidate scoring are assigned to explicit components.Source: §A.3 · p. 12

A.4 Detailed Reuse and Symmetry Proofs

The appendix details how Jeva reuses option-side computation across requests and preserves decisions under candidate reorderings. It also shows how the joint loss separates coverage supervision from candidate ranking.

A.4 Detailed Reuse and Symmetry Proofs

  • Cached option representations and layer-specific key/value projections remain reusable across requests, but each request still requires encoding and full decision-module evaluation.The result concerns reuse of computation, not reuse of an answer.Source: §A.4 · p. 12
  • Prepared option memories reduce per-request decision cost to O(L(d^2 + Kd + df)), while dense K-option output still requires Ω(K) writes.For f = 4d, the decision cost is O(L(d^2 + Kd)); this is a computation bound, not an accuracy extrapolation.Source: §A.4 · p. 12
  • Reordering candidates together with their metadata preserves the decision computation and readouts; the symmetry concerns reorderings, not changes in multiplicity or set composition.Numerical tolerances and tied maxima are treated separately in the appendix.Source: §A.4 · pp. 12–13

A.5 Joint Cross-Entropy Identity

  • The joint cross-entropy supervises coverage without a relative-ranking target on all-null examples, while answer-present examples supervise both coverage and competition.The population log-loss decomposition is given in Appendix B.4.Source: §A.5 · p. 13

B Method Proofs and Implementation Details

These proofs characterize a single-block representational capability, how temperature and added candidates affect aggregation, and how joint loss separates coverage from competition.

B.1 The Single-Block Construction in Higher Dimensions

  • A single-block construction in decision representation space shows that an independent scorer cannot reverse the preference between two retained options when the surrounding set changes.A shared coverage factor also cancels from the ratio between their probabilities.Source: §B.1 · p. 13

B.2 Temperature and the Sensitivity of Candidate Aggregation

  • For fixed logits, attention entropy is nondecreasing in temperature.The appendix derives the temperature sensitivity through a variance expression.Source: §B.2 · p. 13

B.3 A Quantitative Bound for Added Candidates

  • The bound on changes to unchanged-option scores depends on added candidates’ total attention mass and local FFN sensitivity; extending it across blocks requires bounds on incoming-state changes.The argument assumes bounded option-value norms and a Lipschitz bound for the residual FFN on the relevant segment.Source: §B.3 · p. 14

B.4 Joint Log Loss and the Meaning of Null

  • Excess expected joint log loss decomposes into coverage error and conditional competition error, with the competition term weighted by ρ⋆.When ρ⋆ = 0, the competition term vanishes and q⋆ may be chosen arbitrarily.Source: §B.4 · p. 14

B.5 Symmetry, Cache Identity, and Numerical Evaluation

The appendix specifies conditions for symmetry and cache equivalence, computational costs, implementation settings, type-specific losses, and the training curriculum.

B.5 Symmetry, Cache Identity, and Numerical Evaluation

  • Permutation invariance applies to the set of maximizing candidates, but selecting the first tied index can change the chosen member; exact arithmetic and unchanged cache configuration are assumptions.Floating-point reductions or batching changes can perturb nearly tied scores, and parameter updates require invalidating cached memories.Source: §B.5 · p. 14

B.6 Detailed Computational Cost

  • Prepared memories remove the 2Kd^2 option-projection cost from each request, while attention and aggregation remain linear in K and no K^2 interaction tensor is constructed.The final readout costs Kd; separate assembled-set entries may add storage beyond the atomic-bank count.Source: §B.6 · p. 15

B.7 Encoding and Numerical Configuration

  • The implementation uses independent request and option templates, explicit length limits, and a Jeva-4B decision core with one matrix block and eight heads.The inspected core has width 768 and contains 13,370,117 parameters with its two projections.Source: §B.7 · p. 15

B.8 Type-Specific Losses and Optimization

  • For Noul, the specified losses include squared error against the positive-candidate target and, for hard labels, a margin term; soft labels omit that margin.The margin computation clamps probabilities in FP32 without clamping the reported probability; the described loss setting applies to valid numeric rubrics with K ≥ 2.Source: §B.8 · p. 15

C.1 Training curriculum

  • Training proceeds through staged typed-decision, reasoning, open-set, retention, and large-pool phases, with supervision combining hard labels, soft probabilities, null outcomes, and rubric targets.Reported stage counts are training-set sizes and include replay where applicable.Source: §C.1 · pp. 15–16

C.2 General evaluation and typed targets

The evaluation section defines the general-suite tasks and metrics, then describes matched null-detection and candidate-order tests.

C.2 General evaluation and typed targets

  • The general suite covers Typed Decisions, Banking77, MASSIVE, and three Mind2Web slices, with candidate counts and task composition specified for each.Typed Decisions contains 600 Choice, 600 Noul, and 800 Score examples; Banking77 and MASSIVE include answer-present and answer-withheld requests.Source: §C.2 · p. 16
  • General accuracy uses first-argmax correctness, including null on open-set requests; Noul Brier and Score MAE/RMSE assess their respective typed outputs.Unsupported requests are documented separately in coverage results.Source: §C.2 · p. 16

C.3 Null operating points and candidate permutations

  • The null comparison uses 3,080 matched Banking77 K = 16 requests, while order tests measure decision flips and aligned-distribution total variation across candidate permutations.The null set has 2,492 answer-present and 588 answer-absent cases; order tests use 600 Choice examples and up to three seeded random permutations.Source: §C.3 · p. 16

C.4 Large-set support and bounded option encoding

Jeva completes the full 500-request candidate-count sweep with higher aggregate accuracy than recursive JevK5 selection, while the surrounding experiments specify cache comparisons, typed inputs, open-set adapters, and reuse across new sets.

C.4 Large-set support and bounded option encoding

  • 78.40% aggregate accuracy for Jeva exceeds JevK5’s 64.00% across the ToolRet sweep, with Jeva directly scoring the full candidate set in one decision-head pass.Jeva covers all 500 requests, reaching 86.90% at K = 32 and 69.88% at K = 2048; candidates include weak or unjudged negatives, and 389 requests are multi-positive.Source: §C.4 · p. 16

C.5 Cache conditions, timing, and numerical agreement

  • The cache ablation compares Off, Set, and Hybrid conditions across 171 episodes and 2,188 request identities, measuring 46 novel-set requests at each of 0%, 50%, and 100% option reuse.Priming is timed separately and excluded from request latencies; exact-set hits are zero for primed novel sets.Source: §C.5 · p. 16

C.6 Input Templates and Typed Examples

  • Candidate inputs omit the current request, other candidates, and position identifiers; Noul returns affirmative-candidate probability, while Score computes an expectation over rubric values.For Score, rubric values specify the output scale rather than being added as encoder fields.Source: §C.6 · p. 17

C.7 Evaluation Adapters and Cache Examples

  • Open-set adapters estimate answer availability with an auxiliary coverage question or an appended null candidate, keeping availability distinct from candidate selection.The evaluated distribution includes null alongside the original candidates; some adapters add no new trained prediction head.Source: §C.7 · p. 17
  • Hybrid caching reuses option vectors and layer-specific projections after a new ordered-set miss, but model-weight or encoder-setting changes require cache invalidation.Even with cached option memory, Jeva encodes the current request and computes new output probabilities.Source: §C.7 · p. 18

Related papers

Loading 2610.18773v1…