Exploring System One AI Models Beyond Jev: 15 Outstanding Models

But first of all there are two completely opposite types of models – System One and 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 can help with decisions requiring several dependent steps, but adds latency and computation.

The distinction concerns how computation is used, not whether one model is intelligent and the other is not. Hybrid approaches can answer straightforward cases directly and reserve extra reasoning for uncertain ones.

System 2 has had much of the spotlight, with researchers scaling reasoning at inference time. Now direct decision-making is getting a closer look. Here are recent papers exploring its architectures, modalities, and computational trade-offs. →

Jeva encodes the request and each option separately, then updates the request against cached option representations. Shared scoring supports variable-sized candidate sets, symmetric aggregation preserves option-order consistency, and a separate coverage estimate detects when no candidate fits. Reusable option encodings reduce repeated work across requests with overlapping alternatives.

Main idea: Handling changing candidate sets and missing answers.

Prism-1 scores each candidate in an isolated attention block with shared positional indices, preventing its scoring tokens from seeing competing options. A softmax converts independent scores into probabilities. This makes reordering irrelevant and preserves the relative odds between existing options when another is added, imposing a specific choice-consistency rule.

Main idea: Removing option-order and rival-option effects

Bongard uses a T5Gemma 2 encoder to read the state and instructions once, then shares that encoding across decoder branches answering different questions. A decision head returns candidate probabilities without text generation. Training progresses from supervised judgments to semantic relationships and action outcomes, teaching the shared representation through several forms of feedback.

Main idea: Sharing evidence across multiple judgments

Verdict-MQ encodes a document or case once, then answers several typed questions from that representation. A structured attention mask keeps the state independent of the questions, prevents questions from influencing one another, and makes option order irrelevant. This shares input processing across judgments such as urgency, ownership, and whether a reply is needed.
Research paper: Verdict: Encode Once, Decide Many

Main idea: Answering multiple questions from one shared state

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 sequentially, while Joint inference handles multiple questions together, avoiding repeated processing of the same video history.

Main idea: Reusing video computation across questions

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 use coarse-to-fine classification over value bins and regions. Each classification step requires one forward pass, replacing generated spatial answers with structured predictions.

Main idea: Predicting spatial relations, values, and points

AudioJev scores supplied answers directly from a waveform and question, retaining acoustic evidence that transcripts can lose. Training pairs original and reordered candidate lists, supervises both, and aligns their probability distributions using symmetric KL divergence. This reduces option-order sensitivity while keeping inference to one scoring pass, without an order ensemble.

Main idea: Making decisions from sound without transcription

Intern-Decision takes text and optional images, with decision placeholders inside a complete JSON structure. Its training objective supervises those placeholders, while inference scores every field in one causal forward pass using the backbone’s language-model head. The 0.8B, 2B, and 4B models avoid generating and parsing a full JSON response.

Main idea: Answering several structured fields in one pass

MetaEncoder separately embeds natural-language requests and candidate answers, 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 both small closed-set decisions and retrieval over millions of alternatives without jointly encoding every request–candidate pair.

Main idea: Scaling from a few options to millions of candidates

FastJEV reuses context states, shares matching candidate prefixes, and prunes candidate-path layers according to their effect on relative scores. It preserves full context encoding and all candidates without retraining. The method reduces computational depth, but prefix-sharing overhead can increase measured latency, exposing the difference between doing less computation and finishing faster.

Main idea: Reducing repeated candidate computation

SanSi repeatedly applies shared Transformer layers and reads candidate probabilities after each loop. Training supervises different loop budgets, allowing decisions after an early pass or further hidden-state refinement. This is System 1.5, not single-pass inference: additional loops cost computation, and confidence can keep increasing after accuracy gains level off.

Main idea: Adding internal computation through loops without generating reasoning

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. The family connects direct decisions with selective deliberation.

Main idea: Reserving explicit reasoning for uncertain cases

System X adds a text-reply option to a direct decision readout. When that option wins, or confidence falls below a threshold, the model generates a response from the existing KV cache. This hybrid connects structured selection with conversational replies while avoiding a second prompt-prefill step.

Main idea: Reusing decision computation for a text response

FunDM trains a small decision head to match a frozen language model’s answer probabilities using 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.

Main idea: Distilling decisions into a shorter computation path

This model places 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 select which information to request instead of immediately choosing an answer from incomplete evidence.

Main idea: Estimating when clarification would improve a decision

System One AI models make decisions by directly scoring or selecting 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.

System One focuses on direct decisions using the available evidence, while System Two uses additional computation to reason through a problem, explore alternatives, or check an answer. System One is useful for fast, bounded decisions; System Two is better suited to problems that require multiple dependent steps. Some architectures combine both.

AI agents make many small decisions, such as choosing tools, routing requests, checking conditions, or interpreting observations. Generating a text response for each decision introduces decoding and parsing overhead. System One models can return structured decisions directly, making them an interesting option for these tasks.

Not necessarily. Direct scoring removes autoregressive answer generation, but actual latency depends on the architecture, candidate set, input size, and implementation. Some optimizations reduce computation without improving wall-clock speed, as the FastJEV research illustrates.

Yes. Recent approaches extend direct decision-making beyond text. 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 possible outcomes.

Some can allocate additional computation when a direct decision isn't sufficient. 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. These approaches explore the space between immediate decisions and extended reasoning.

Different architectures approach uncertainty differently. Jeva estimates whether any candidate fits, StartLux-Decision can reserve additional reasoning for uncertain decisions, and LAVOIR estimates whether asking for missing information would improve the result. However, a model returning probabilities doesn't automatically mean those probabilities are well calibrated.

Thank you for reading! [Sign-off — e.g. “Share this list with a friend who is still deploying models by hand.”]