Daily papers

Grouped by first public publication date.

2026-10-08

One Word Opens the Gate: The Option-Channel Attack on Typed Decision Models as Agent Guardrails — Separates fail-open and fail-closed errors in open decision-model guardrails; misleading option names can reverse decisions when labels enter the model input.
Specialized Decision Models vs. General-Purpose LLMs: Benchmarking Jev Across Knowledge, Reasoning, and Multilingual Tasks — Compares Jev with 19 LLMs on 13 benchmarks; strong knowledge scores coexist with a marked weakness on mathematical word problems.
Can Decision Models Understand Stance? Evaluating Jev Against General-Purpose LLMs — Jev matches GPT-5.6 on English VAST stance labels but trails stronger models on Chinese conversational stance, especially favor-versus-against distinctions.
Can Jev be Your Q or Policy in Reinforcement Learning? — Uses frozen Jev as a policy reference, exploration judge and replay rater; gains depend on the learning role and supplied information.
Adversarial Cues in Decision Models Used as Judges: The Role of Request Presentation — A colon edit increases false acceptance of explicitly wrong final answers under sorted request keys, while insertion-order requests reject both variants.
TypedBench: A Benchmark for Calibration, Framing Sensitivity, and Cost in System One Decision Models — Uses policy-driven generators to test calibration, framing, abstention and cost; good error ranking does not ensure calibrated or evidence-sensitive confidence.
FastJEV: Understanding Redundancy for Compact JEV Inference — Combines context-state reuse, candidate-prefix sharing and layer pruning for OmniJev; compact execution does not consistently reduce measured latency.
MetaEncoder: Exploring the Limit of Bi-Encoders for Multimodal System One Decision Making with Natural Language Interface — Trains a multimodal bi-encoder with request-to-candidate contrastive learning; including options in requests improves closed-set decisions while retaining cached candidate embeddings.
Can a System-One LLM Perform Knowledge Tracing When Few or No Learners Are Logged? — Tests cold-start knowledge tracing with reader swaps and matched typed inputs; Jev benefits mainly from its model prior, while supervised methods catch up with more learners.

2026-10-07

JevForest: Path Voting for Budgeted Feature Acquisition — Uses tree-path voting to select semantic questions under a feature budget; fewer sequential questions cost more than batching in the tested pilots.
Where Can a Decision Model Diagnose HVAC Faults? Reasoning Demand, Physical Representation, and Robustness Under Shift — Separates physical representation from reasoning in fault diagnosis; engineered features improve consistency, but stable shift performance does not imply superior absolute accuracy or useful detection.
System Switch: When Should a Fast Decision Model Stop and Think? — Tests confidence-based deferral from fast decision models to a slow visual reasoner; offline gains do not translate into Doom level completion.
Visual Jev Rewards: Reference-Bound Verification for Multi-Subject Image Generation — Trains a local Qwen verifier for joint reference identity and image conditions, then averages binary probabilities as an image-generation reward.

2026-10-06

Calibrated Decisions Are Not Calibrated Probabilities: An Exact-Target Audit of Jev and Three Open Decision Models — Audits exact probability targets and human disagreement: normalized Noul better tracks stated base rates, while per-task calibration narrows the gap on human-vote distributions.
From Probabilities to Decisions: Search and Multi-Teacher Distillation with Jev — Distills Jev and Qwen pairwise probabilities into small evaluators; matched-budget chess gains replicate, while retrieval gains remain uncertain.
Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts? — Benchmarks agents that build reusable artifacts for large workloads; most runs lose quality relative to zero-shot calls, while selected artifacts approach Jev at lower projected API cost.
Same-Number Citation Swaps: Stress-Testing Jev as a Financial Evidence Judge — Holds arithmetic and operand values fixed while swapping financial citations; Jev misses some wrong-role evidence and rejects some equivalent evidence, with column rendering changing the trade-off.
Token-Efficient Multi-Agent Collaboration via System One-Guided Computational Division of Labor — Uses a Laya controller, compact evidence reader and bounded task menus to coordinate LLM workers; reports lower GPT token use and end-to-end latency across seven benchmarks.
When Plans Change Answers: Formalizing Cost-Accuracy Optimization for Semantic Queries — Formalizes contribution-weighted semantic-query quality; v2 adds confidence-centric skipping of unscored tuples once predicted output targets are guaranteed.
Benchmarking System One Models in Online Moderation — Separates moderation rules, retrieved precedents and answer inventories across five benchmarks; Jev benefits from precedents in matched settings, while Laya often changes its operating point without better discrimination.
SanSi: A Looped Typed Decision Model for System 1.5 Thinking — Trains typed readouts after repeated backbone loops; eight loops improve accuracy at fixed model shape, but cost more compute and calibrate worse than three loops.
Readout Stability in Prefill-Only Decision Models:Zero-Label Prediction and Inference-Time Compute Allocation — Tests cached first-pass rankings under candidate-menu changes and compares re-asking with model cascades; ranking stability supports computation reuse but does not itself identify unlabeled accuracy.

2026-10-05

ufakzeka-karar: An Open Turkish Typed-Decision Model with Order-Invariant Option Scoring — Scores Turkish answer options independently at shared positions to remove order effects; supervised training beats the tested REINFORCE variant, while temperature calibration degrades on held-out question types.
JEV versus LLMs: Accuracy, Cost and Calibration on Seven Political Science Replications — Replicates seven annotation and scaling studies: Jev is faster, matches or approaches comparison models on several tasks, and has no price advantage over the reported batch-rate baseline.
SoK: Semantic Decision Engines in Network Control Loops — Audits 139 network-control paper families and separates decision latency from verified service completion; queueing and coverage checks can reverse admission decisions.
GraphDecide: Benchmarking System One Models on Graph Tasks — Benchmarks graph structure, graph-text evidence and sequential optimization across fourteen model-interface configurations; Jev benefits from heuristic proposals but gains over fixed rules depend on the task.

2026-10-04

Hidden Risks of Jev: An Empirical Study of Security, Privacy, and Dual Use — Studies input manipulation, constrained-output privacy and defensive detection; official Jev API experiments and controlled NanoJev training experiments expose distinct risks.

2026-10-03

System One Models for Wireless Decision-Making:Applications and Performance Evaluation — Measures the quality/latency trade-off in antenna selection and RAN slicing: Jev responds faster, while utility gains depend on the task and execution costs.

2026-10-02

Benchmarking Candidate Coverage in Typed Decision Models — Pairs present and omitted reference labels at matched candidate counts; Jev and Laya show task-dependent detection versus false-rejection trade-offs.
To Jev or Not? Evaluating the Accuracy and Efficiency of Structured Decision Models for Hate-Speech Moderation — Compares six decision-model configurations for hate-speech moderation; supplied definitions and question decomposition do not consistently improve agreement with labels.
SecJev: Bringing Security Expertise to System One Decision Models — Specializes a shared candidate scorer with security LoRA adapters; strong in-domain gains coexist with limited latency gains over one-token generation and transfer false alarms.

2026-10-01

LLM2Jev: LLMs Are Already Jev-Style Decision Models -- When and How to Fine-Tune Them — Studies when existing LLMs already support Jev-style decisions.
HydroJEV: A one-second, training-free screen for cyber-attack and fault attribution in water distribution networks — Uses Jev and a rule gate to triage simulated water-network alarms, reducing the workload passed to an LLM reviewer.
Code Owns the Simulation, Jev Owns the Evaluation — Separates evaluating supplied consequences from simulating missing ones.
Jev-IDS: System One Models for Network Intrusion Detection — Evaluates typed intrusion decisions with limited labels on an NSL-KDD pilot; the study does not establish performance on live network traffic.
Beyond Answer Confidence: A Controlled Audit of Self-Knowledge in a Black-Box Decision Model — Audits missing knowledge with paired evidence interventions.

2026-09-30

Tron-1B: Fast, Calibrated Typed Decisions with a Set-Attention Option Head — Introduces a bidirectional encoder with a set-attention option head; benchmark-trained results are compared with a zero-shot hosted service.
AnyJev Technical Report — Reads option probabilities from pretrained LLMs and corrects label and position bias with prior normalization and cyclic option rotations.
JevSpawn: Adaptive Agentic Inference through Compositional Action Spaces — Explores compositional action spaces for fast agent inference.
OmniMed-Jev: Calibrating LVLM Confidence for Trustworthy Medical Multimodal Decisions via System One — Studies candidate-conditioned multimodal decisions against a matched generative baseline, reporting improved calibration with task-dependent accuracy trade-offs.
A First Glance at Jev for Network Traffic Classification: Accuracy, Processing Time, and Cost — Jev is faster than the tested LLM but trails trained trees on traffic classification.
When the Right Answer Is Missing: An Arithmetic-Dependent Rejection Bottleneck in Jev — Finds an arithmetic rejection bottleneck despite an explicit fallback option.
Bongard: Training Machine Intuition — Trains an open System One model with outcome feedback.
OpenJev-RLCD: A Working RLCD Implementation — Implements a third-party calibrated-decision reinforcement objective.
More Choices, Fewer Decisions: Ordinal-Scale Bias in JEV-like Direct-Decision Models — Measures ordinal scale compression as candidate counts grow.

2026-09-29

Confident Where People Disagree: A preregistered, bias-corrected test of whether TypeSafe AI’s Jev lowers its confidence when humans disagree, on ChaosNLI — Tests confidence under human disagreement with bias-corrected calibration.
Benchmarking System One decision models against trained classifiers and language models for automated decision gates — Compares typed models, trained classifiers and LLM readouts under matched requests.
Evaluating and Benchmarking the System One Model Jev — Evaluates Jev across public classification, routing and reasoning datasets. Calibration and selective prediction vary with the task and decision interface.
Chinese-Jev: Bringing System One Model to Chinese-Language Tasks — Develops a System One model for Chinese-language tasks.

2026-09-28

A Noul Log Does Not Identify the Policy — Shows why marginal Noul probabilities do not identify a conjunction policy.
Calibrated to Whom? Persona and Language Effects on Cultural Values in JEV — Audits persona and language effects despite highly repeatable answers.
Dyad: Extending Large Language Models with Native Typed Decision-Making — Adds an action-description encoder to an LLM and evaluates frozen-backbone and joint reinforcement learning for typed decisions.
Koa-action: Fast and Consistent Structured Decision Making with Generative LLMs — Uses atomic output tokens for fast structured decisions.
Mnemon: Raw Records, Fast Judgments, Slow Thoughts — Combines raw records with fast judgments and slower reasoning.
Jev thinks "I don't know'', but doesn't say it: Introducing Sys1Cal-v1 Dataset for Probability Calibration — Introduces Sys1Cal-v1 with probabilities known by construction, exposing differences between Choice, Noul and Score outputs.
The Argument and the Letterhead: Source-Position Coherence in AI Evaluation — Separates source attribution from the quality of a fixed argument.
NavJev: Efficient Vision-Language Navigation via Action-Centric Visual Compression and Discriminative Action-Semantic Memory — Studies action-centric representations and decision memory for navigation.
JevVibe: Efficient Classification-Guided Secure Code Generation — Uses Jev to classify code weaknesses and guide a repair model. Repairs improve results under the benchmark’s insecure-code detector.
JEV as a Judge for Agent Trace Security: An Empirical Comparison with Generative LLM Judges — Compares Jev with generative judges on agent-trajectory security benchmarks, examining detection quality, valid outputs and inference cost.
When Does Selection Replace Extraction? A Pre-Registered Test of Agent Memory with a Typed Decision Model — Tests raw-turn selection against extraction under controlled memory budgets.
Decision Readouts for Text-Mediated Video Anomaly Detection: An Exploratory Evaluation of Jev and Qwen — Compares anomaly decision readouts under fixed video-derived evidence.

2026-09-27

Probability Contracts: Accuracy, Coherence, and Decisions Across LLM Interfaces — Links exact posteriors, equivalent requests and action-sensitive loss.
Jev in Medicine: A Benchmark Evaluation — Audits Jev 1.13 accuracy, calibration, selective prediction and unanswerable-question handling across four medical benchmarks.
Do System One Decisions Add Up? A Study of Probabilistic Coherence — Compares direct and hierarchical decisions on matched items.
JET: Justification Evaluation in Transformer — Tests local candidate-likelihood inference with shared computation.
Laya as a Typed Probabilistic Assessor: An Independent Reproduction and a Preregistered Study of Calibration and Selective Escalation — Audits calibration and selective escalation in an open decision model.
Type-Safe Decision Frameworks for Agentic 5G Control: A Theory-Driven Testbed Characterization of Where They Can Be Applied — Maps when typed decision gates are applicable to network control.
COGNIT-Guard: Calibrated Standalone Direct-Decision Guardrails with Heterogeneous CPU-NPU Confidence Cascading under Explicit Latency and False-Positive Constraints — Combines a calibrated CPU gate with a Laya CPU-NPU cascade for prompt safety, measuring false positives, latency and out-of-domain transfer.
You Only Edit Once: Incentivizing In-Context Capability of LLMs via Local Demonstration Refinement — Uses typed decisions in local demonstration refinement.
Evaluating System One Models for Agent Security Decisions: Reliability, Calibration, and Selective Automation — Compares security judgments across attack groups and tests confidence-based escalation. Aggregate calibration can hide groups with more missed attacks.
Beyond Calibration: Do a Typed-Decision Model's Probabilities Obey the Probability Axioms? — Tests logical probability coherence without needing class labels.

2026-09-26

PACT: Pairwise-Anchored Calibrated Tuning for Single-Token Typed Decisions — Uses paired evidence and permutation constraints to improve typed-readout robustness.
Typed Decision Models: An Early Evidence Audit and Evaluation Checklist — Audits the first research wave and its evaluation practices.

2026-09-25

Jev at the Agent Authorization Boundary: Evaluating TypeSafe’s Decision Model on Allow, Hold, and Deny — Evaluates allow, hold and deny against authored policy specifications.
JevAdvBench: A Benchmark and Black-Box Attacks for Reinforcement Learning for Calibrated Decisions Models — Introduces an adversarial benchmark that compares single-field perturbations with repeated clean requests, separating attack effects from output noise.
JevSoup: System-One Routing for Training-Free LoRA Composition — Routes and combines LoRA experts using typed choices.
LAVOIR: Teaching a Single-Pass Decision Encoder When and What to Ask with Amortized Value of Information — Learns which missing information is worth asking for.

2026-09-24

How far can a commercial decision model’s probabilities be trusted? A calibration audit of Jev against open and general-purpose classifiers — Audits task-specific calibration and label-wording sensitivity.
JevOut: Natural Context Can Flip Decision Models — Optimizes ordinary-looking context to redirect initially correct decisions.
Jev in the Wild: A Data-Driven Analysis of the Jev Model's Functionality, Applications and Ecosystem — Maps decision interfaces and adoption across 2,170 public projects.
Jev-Mobile: Jev as an Executor for Mobile GUI Agents — Separates VLM planning from typed GUI execution.
JEV vs. LLMs as Rubric Judges: Cheaper, Faster, and Wrong in the Same Places — Correlated errors limit the accuracy benefit of judge cascades.
Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures — Evaluates zero-shot detection of ten alignment-failure families.
From Text Decisions to Pixels: An Study of Jev-Style Visual Choice Model — Studies visual choice readouts with matched generation baselines.
Calibrated Decision Models for Autonomous Penetration-Testing Harnesses: JEV and Laya as System One Decision Layers for LLM-Driven Pentest Agents — Explores typed decision layers for pentest finding validation and triage.
Harness Tokenomics: A Router for the Enterprise Agentic Control Plane — Models routing economics with session-level cache effects.

2026-09-23

Jev in Practice: A Composable Python Toolkit for TypeSafe’s System One Decision Model — Provides composable tooling, calibration utilities and recorded API experiments.
Decision Hijacking: Prompt Injection Attacks on Jev's Typed Probabilistic Decisions — Measures injection-driven probability shifts and validated targeted decisions.
NumericJev: Jev-like LLM Numerical Decoding with Multiway Decision Trees — Uses multiway decision trees for numerical readout.
Same Scores, Different Decisions: Evaluating JEV and Language Models for Legal Document Understanding — Shows that aggregate accuracy and repeated agreement hide item-level failures.

2026-09-22

Typed Decisions at the Edge: A Privacy-Preserving Hybrid Architecture for Everyday Decision Support — Studies a privacy-oriented typed interface and deterministic fallback.
When a Judgment Layer's Self-Reported Fields Lie: Cost, Latency and the Failure Boundary of Three Judgment Layers on the Same Items — Audits judgment-layer costs, latency and self-reported fields; separates access-layer defects from model errors and tests paired error complementarity.
Type-Safe Is Not Error-Free: A Constrained Decision Head Follows the Option Name, Not the Rubric Bound to It — Tests whether option names override their attached rubrics.
JEV-as-a-Judge: Accept When Confident, Escalate When Unsure — Studies confidence-based acceptance and escalation for model judging.
REFLEX with Jev for Efficient Selective Control in LLM Agents — Tests selective agent control and near-valid action alternatives.
Visual Jev: Accurate and Efficient Decisions from Shared Visual Context — Compares visual decision heads with matched generation baselines. Shared computation improves efficiency, while dedicated heads do not consistently improve accuracy.

2026-09-21

Universal Fractal Natural Language Decision Map: Real-Time Edge Triage Across Heterogeneous Domains — Reports an alternative edge decision implementation and benchmark comparison.
Jev for Scientific Decisions: Evaluating Semantic Choices and Their Consequences — Separates semantic choices, intermediate calculations and final labels to reveal errors that a final-answer score can hide.
Evaluating Decision Models for Text Annotation in Computational Social Science — Compares decision models with LLMs on text annotation; confidence-based routing helps on some tasks but high confidence can conceal task-specific errors.
Calibrated Decisions at Scale: Converting Police Crash Narratives into Probabilistic Crash Variables with a System One Model (Jev) — Audits Jev probabilities against blinded human labels and tests recalibration, probability-grid resolution and review-budget allocation.
Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents — Places a typed decision controller inside agent memory.
Open-Jev Judgments on CallScreenBench: Calibrated One-Pass Scam Screening with a Small Language Model — Evaluates an open typed readout for scam-call screening.

2026-09-20

this-that-model-1.0: A typed decision model that decides in 30 ms, for a millionth of a cent — Provides an open model with a native typed decision interface.

2026-09-19

Fast Intent-Driven Service Orchestration with Jev for 6G Edge Networks — Studies how decision latency affects bounded service orchestration.
Replacing Large Language Models with Jev Decision Models for Low-Latency Edge Service Orchestration — Measures typed decisions in deadline-constrained service admission.

2026-09-17

Calibration Does Not Compose, Types Destroy Vagueness: The Hidden-Markov and Fuzzy Primitives Missing from System-One Decision Models — Analyzes how latent regime shifts and repeated thresholds can invalidate composed decision pipelines, using formal assumptions and synthetic experiments.