Beyond Jev: 15 System One Decision Models That Turn Reasoning into a Single Forward Pass
System One models skip token-by-token answers and return a choice, score, or probability in one forward pass. A guide to 15 studies on dynamic candidates, multimodality, retrieval, and selective reasoning.
Not every AI task needs an answer. Many need a decision. A tool call, a routing choice, a safety check, a box to select, an event to flag — these are bounded choices, not essays. System One models are built for exactly that: they take a state, a question, and a set of candidates, then return a choice, score, or probability distribution directly, usually in a single forward pass.
This guide grew out of the attention around Jev by TypeSafe AI, which answers multiple typed questions in parallel and returns typed decisions. Jev is not alone. The 15 research efforts below extend the same idea in different directions: changing candidate sets, option-order consistency, shared evidence across questions, video and audio inputs, retrieval at millions of candidates, and selective reasoning when a direct decision is not enough.
For robot builders, the topic is less exotic than it first appears. A policy that maps an observation to an action, a VLA that picks among action chunks, a supervisor that decides whether to re-plan — all of them are decision problems wearing different clothes. The question this guide raises is whether those decisions should be generated token by token, or scored directly.
What a System One model actually does
The term covers models that make a choice or help a person or system make one. They use the available information to select, rank, or score possible actions or outcomes. A System One model takes a state, a question, and candidate answers, then returns a choice, score, or probability distribution without generating a text response.
This matters most when an agent makes many small decisions: selecting a tool, routing a request, checking a condition, interpreting an observation. Generating a response for each one adds decoding time and forces the surrounding software to parse free text back into a structured value. Direct scoring removes both steps. It does not guarantee a correct decision or a well-calibrated confidence, but it removes a layer of machinery that was never part of the decision itself.
System One versus System Two
System 1 makes a direct judgment from the available evidence. It suits bounded decisions where the possible outputs are already defined, such as choosing among tools or detecting an event. System 2 spends additional computation working through a problem, often generating intermediate reasoning, exploring alternatives, or checking a proposed answer. That helps with decisions that require several dependent steps, but adds latency and computation.
The distinction is about how computation is used, not about one model being intelligent and the other not. Hybrid approaches can answer straightforward cases directly and reserve extra reasoning for uncertain ones. Several models below are explicitly built around that boundary.
The 15 models at a glance
| Model | Main approach | What makes it different | Best suited for |
|---|---|---|---|
| Jeva | Candidate scoring | Handles changing candidate sets and detects when no option fits | Dynamic decision tasks |
| Prism-1 | Independent scoring | Avoids option-order and competing-option effects | Consistent multiple-choice decisions |
| Bongard | Shared encoding | Reuses evidence across multiple judgments | Multiple decisions about one state |
| Verdict | Multi-question scoring | Answers several typed questions from one encoded input | Document classification and workflow automation |
| VideoJev | Video decision model | Reuses video representations across questions | Video understanding and event detection |
| MJev-Spatial | Spatial scoring | Predicts relations, numerical values, and image coordinates | Visual and spatial reasoning |
| AudioJev | Audio decision model | Scores answers directly from audio without transcription | Audio understanding |
| Intern-Decision | Structured prediction | Scores multiple JSON fields in one forward pass | Multimodal structured decisions |
| MetaEncoder | Bi-encoder retrieval | Precomputes candidate representations for large-scale matching | Decisions over millions of candidates |
| FastJEV | Inference optimization | Reduces redundant candidate computation without retraining | More efficient JEV inference |
| SanSi | Looped decision model | Refines decisions through repeated internal passes | Adjustable computation without text reasoning |
| StartLux-Decision | Hybrid decision and reasoning | Uses additional reasoning when confidence is low | Decisions requiring selective deliberation |
| System X | Decision-to-text hybrid | Generates a text reply when direct selection is not appropriate | Agents needing both choices and responses |
| FunDM | Decision distillation | Learns a shorter decision path from a frozen language model | Lower-cost decision inference |
| LAVOIR | Value-of-information scoring | Predicts whether asking for missing information would help | Decisions under incomplete information |
Decision architectures
1. Jeva
Jeva encodes the request and every option separately, then updates the request against cached option representations. Shared scoring lets the candidate set change size, symmetric aggregation keeps the result independent of option order, and a separate coverage estimate detects when none of the candidates fits. Because option encodings are reusable, overlapping requests do not pay for the same work twice.
Links: Paper · GitHub · Project · Weights
2. Prism-1
Prism-1 scores each candidate inside an isolated attention block that shares positional indices, so the tokens scoring one option never see its competitors. A softmax turns the independent scores into probabilities. Reordering candidates becomes irrelevant, and adding a new option preserves the relative odds of the existing ones — a specific choice-consistency rule rather than a statistical accident.
Links: Paper
3. Bongard
Bongard uses a T5Gemma 2 encoder to read the state and instructions once, then shares that encoding across decoder branches that answer different questions. A decision head returns candidate probabilities without generating text. Training moves from supervised judgments to semantic relationships and action outcomes, teaching one shared representation through several kinds of feedback.
4. Verdict
Verdict-MQ encodes a document or case once, then answers several typed questions from that representation. A structured attention mask keeps the encoded state independent of the questions, stops questions from influencing one another, and makes option order irrelevant. Input processing is shared across judgments such as urgency, ownership, and whether a reply is needed.
Links: Paper
Multimodal extensions
5. VideoJev
VideoJev replaces autoregressive answer decoding with a shared head that scores temporal intervals and textual options against video evidence. It returns event spans or answer choices directly. Cached inference reuses visual computation when questions arrive one after another; joint inference handles several questions together, avoiding repeated passes over the same video history.
Links: Paper · GitHub · Project · Weights
6. MJev-Spatial
MJev-Spatial reuses pretrained query and key projections to score spatial candidates without adding scoring parameters. It selects supplied answers directly, while numerical estimates and image coordinates are produced through coarse-to-fine classification over value bins and regions. Each classification step is one forward pass, replacing generated spatial answers with structured predictions.
7. AudioJev
AudioJev scores supplied answers directly from a waveform and a question, keeping acoustic evidence that a transcript can lose. During training it pairs original and reordered candidate lists, supervises both, and aligns their probability distributions with symmetric KL divergence. Option-order sensitivity drops while inference stays a single scoring pass, with no order ensemble.
Links: Paper · GitHub · Weights
8. Intern-Decision
Intern-Decision takes text and optional images, with decision placeholders embedded inside a complete JSON structure. Training supervises those placeholders; inference scores every field in one causal forward pass using the backbone's language-model head. The 0.8B, 2B, and 4B variants avoid generating a full JSON response and parsing it afterwards.
9. MetaEncoder
MetaEncoder embeds natural-language requests and candidate answers separately, with images or video available as additional evidence. Contrastive training aligns the two representations, turning a pretrained multimodal decoder into a decision-making bi-encoder. Precomputed candidate embeddings support small closed-set decisions as well as retrieval over millions of alternatives without jointly encoding every request–candidate pair.
Reasoning and inference efficiency
10. FastJEV
FastJEV reuses context states, shares matching candidate prefixes, and prunes candidate-path layers according to their effect on relative scores. It keeps full context encoding and all candidates, and needs no retraining. The method reduces computational depth, but prefix-sharing overhead can raise measured latency — a useful reminder that doing less computation is not the same as finishing sooner.
11. SanSi
SanSi applies the same Transformer layers repeatedly and reads candidate probabilities after each loop. Training supervises different loop budgets, so a decision can be taken after an early pass or refined through further hidden-state computation. It is System 1.5 rather than single-pass inference: loops cost compute, and confidence can keep rising after accuracy has plateaued.
12. StartLux-Decision
StartLux separates single-pass Instant models from AutoThink models that generate reasoning when confidence is low. Instant uses bidirectional attention in full-attention layers. AutoThink combines reasoning supervision with reinforcement learning that rewards improved decision probabilities and penalizes excessive reasoning length, connecting direct decisions with selective deliberation.
Links: GitHub · Weights · Blog
13. System X
System X adds a text-reply option to a direct decision readout. When that option wins, or when confidence falls below a threshold, the model generates a response from the existing KV cache. The hybrid links structured selection with conversational replies while avoiding a second prompt-prefill step.
Links: Paper · GitHub · Weights
14. FunDM
FunDM trains a small decision head to match a frozen language model's answer probabilities with KL divergence, without correct-answer labels. It reads the question and options together through a shortened backbone, then scores candidates without the original language-model head. Its Qwen3-8B implementation uses 24 of 36 layers, trading a small accuracy loss for higher throughput.
15. LAVOIR
LAVOIR puts candidate pieces of missing information alongside answer options. One forward pass predicts both decision probabilities and the expected improvement from asking about each missing detail. Trained on rule-derived decisions and clarification outcomes, it can decide which information to request instead of answering from incomplete evidence.
Links: Paper
What this means for embodied AI
This section is RobotWorld analysis, not part of the original article.
Robot control has always been a System One problem at the bottom of the stack. A balance controller cannot spend 800 milliseconds generating a chain of thought; it has to map state to action now. What changes with these models is the layer above: the part that chooses a skill, selects an action chunk, decides whether a grasp has succeeded, or routes a perception query.
There are four direct mappings. First, action selection can be framed as candidate scoring over a discrete set of skills or action primitives, which removes token decoding from the inner loop. Second, VideoJev and MJev-Spatial map naturally onto robot perception: event spans, spatial relations, and image coordinates are exactly the outputs a manipulation or navigation stack needs. Third, SanSi and StartLux-Decision describe a credible safety architecture: answer fast when confidence is high, pay for deliberation when it is not. Fourth, LAVOIR points at active perception: asking for another view or a missing measurement before committing to an action.
None of this is automatic. The candidates still have to be enumerated. Continuous action spaces need discretization or a scoring head that can rank continuous proposals. Calibration is not guaranteed by returning a probability, and most of these papers evaluate on text, images, or audio rather than closed-loop robot control. The useful claim is narrower but still important: for bounded decisions, direct scoring can remove a decoding-and-parsing layer that robotics stacks have been paying for without needing it.
flowchart LR
S[Observation or state] --> E[Shared encoder]
Q[Task or question] --> E
C[Candidate actions or answers] --> H[Scoring head]
E --> H
H --> D{Confidence high enough}
D -->|yes| A[Select and act in one pass]
D -->|no| R[Extra loops or deliberate reasoning]
R --> A
A --> O[Outcome and new evidence]
O --> H
Limits and open questions
The first limit is structural: direct scoring assumes the options can be enumerated. That is fine for tool names, labels, event spans, and a fixed menu of skills. It is awkward for open-ended generation and for continuous control unless the action space is discretized or a separate proposal mechanism supplies candidates.
The second is calibration. A probability distribution is not automatically a trustworthy confidence estimate. Several models in this list attack the problem from different angles — Jeva estimates whether any candidate fits, StartLux reserves reasoning for uncertain cases, LAVOIR estimates the value of asking for more information — but none of them makes calibration free.
The third is that less computation does not always mean lower latency. FastJEV is the clearest example: it reduces computational depth while prefix-sharing overhead can raise measured wall-clock time. Anyone deploying these models should measure end-to-end latency, not just FLOPs.
The fourth is evaluation. Most of the current evidence comes from classification, visual question answering, document workflows, and audio understanding. Closed-loop robotics adds distribution shift, timing constraints, and failure costs that static benchmarks do not capture. The field still needs decision benchmarks where a wrong choice has consequences and where latency is part of the score.
FAQ
What are System One AI models?
They make decisions by directly selecting or scoring possible outcomes rather than generating a text answer token by token. Given a state, question, and candidate answers, they can return a choice, score, or probability distribution, often in a single forward pass.
What is the difference between System One and System Two AI?
System One focuses on direct decisions using the available evidence. System Two uses additional computation to reason through a problem, explore alternatives, or check an answer. System One suits fast, bounded decisions; System Two suits problems with multiple dependent steps. Some architectures combine both.
Why are System One models useful for AI agents?
Agents make many small decisions — choosing tools, routing requests, checking conditions, interpreting observations. Generating a text response for each decision introduces decoding and parsing overhead. System One models can return structured decisions directly.
Are System One models always faster than LLMs?
Not necessarily. Direct scoring removes autoregressive answer generation, but actual latency depends on architecture, candidate set size, input size, and implementation. Some optimizations reduce computation without improving wall-clock speed, as FastJEV illustrates.
Can System One models process images, video, and audio?
Yes. VideoJev scores video events and answer options, MJev-Spatial handles spatial predictions, and AudioJev makes decisions directly from audio. Other models combine text with images or video to select among outcomes.
Can System One models handle complex reasoning?
Some allocate additional computation when a direct decision is not enough. SanSi refines hidden states through repeated loops, StartLux-Decision can generate reasoning for uncertain cases, and System X can switch from direct selection to a text response.
How do System One models handle uncertainty?
Different architectures approach it differently. Jeva estimates whether any candidate fits, StartLux-Decision reserves additional reasoning for uncertain decisions, and LAVOIR estimates whether asking for missing information would improve the result. A model returning probabilities does not automatically mean those probabilities are well calibrated.
Source: Turing Post, "Exploring System One AI Models Beyond Jev: 15 Outstanding Models" by Alyona Vert., published October 26, 2026 (turingpost.com). This bilingual article is a complete translation and adaptation of the source, with an explicitly labeled RobotWorld analysis section. All links point to the original papers, code, and model weights cited by the author.