Third-party researchEditorial overview of third-party preprints

October 2026 Jev Research Overview

A source-backed overview of six independent, third-party Jev preprints covering theory of mind, application security, adversarial robustness, medicine, and decision bias.

Type Safe Editorial summarizes paper-reported findings from six non-peer-reviewed preprints submitted or revised from October 13 through October 16, 2026. Type Safe has not independently reproduced these results, and incompatible tasks and metrics are not merged into a single score or global ranking.

Publisher: Type Safe
Published or submitted: October 19, 2026
Last verified: October 19, 2026

Type Safe catalogs sources and evidence boundaries. Provider datasets and third-party preprints do not become an independent leaderboard.

Key facts

Papers covered: 6 independent preprints
Evidence dates: October 13–16, 2026
Review status: Third-party, not peer reviewed
Editorial treatment: No pooled score or ranking

Task and scope

Six independent Jev preprints spanning 11,117-question theory-of-mind evaluation, application decision hijacking, text and vision robustness, controlled adversarial judging, 15 medical datasets, and decision-bias experiments.

Research attribution · 6 papers covered

Six independent third-party preprints summarized by Type Safe Editorial.

Paper-reported findings

Jev-ToM compares five typed models and seven generative baselines on 11,117 English questions from nine theory-of-mind benchmarks. The paper reports 73.2% source-balanced accuracy for Jev; outcomes and confidence behavior vary by task.
The JPDH study covers 228 applications and 2,273 program-level test cases. The paper reports decision flips in 1,853 cases, an 81.5% decision flip rate; 1,827 of those flips were classified as potentially unsafe, or 98.6% of flips. These figures are paper-reported and not independently reproduced.
DecisionRobust evaluates 30 text and 17 vision model configurations under perturbation and prompt-injection attacks. The paper reports that relative robustness changes by task and attack, while label-consistency checks materially change measured attack success.
Same Attack, Different Judge compares Jev, LLM judges, and purpose-built classifiers across four task types under a controlled adversarial setup. Results are mixed and task-dependent, so the paper does not establish a universal model ranking.
The medicine study covers 15 medical datasets and 8,742 instances. Its October 16 v2 reports stronger results on text tasks than numerical tasks and identifies limitations in abstention behavior.
The bias case study analyzes 41,500 responses plus 15,000 additional responses. It reports sensitivity to irrelevant numbers and option ordering, including 54% conjunction-rule violations on adapted tasks.

Code and data availability · Code

Artifact availability differs by paper. This overview links the original arXiv records and does not certify every linked repository or implementation.

Code and data availability · Data

Datasets and sample counts are reported by their respective papers; Type Safe has not assembled or validated a combined dataset.

Method and usage

Each study is summarized from its original arXiv record, with scope, reported measurements, and qualifications kept with that source.
The overview separates application security, robustness, medical, social-reasoning, and bias findings because their tasks and metrics are not directly comparable.
Dates reflect the arXiv submission history available on October 19, 2026; the medicine paper includes its October 16 v2 revision.

Limitations and evidence boundaries

All six studies are third-party preprints and were not peer reviewed on the verification date.
Type Safe has not independently reproduced the experiments, audited every artifact, or validated the reported measurements.
The studies use different tasks, samples, attacks, baselines, and metrics, so their results cannot support one pooled score or global ranking.
Preprints can change after publication; readers should check each linked arXiv record for later versions.