TimePrism: Probabilistic Parallel Generation Predates Jev by a Year

Background: Jev's Viral Moment and TimePrism's Prior Work

In September 2026, the AI model Jev gained widespread attention for its ability to output judgments and probabilities directly, enabling developers to define questions and options and receive structured probabilistic results for decision-making. Jev's design emphasizes explicit probabilities, parallel output for multiple questions, and a decision-oriented interface, moving away from token-by-token text generation.

However, a year earlier in September 2025, a research team led by Professor Qiang Xu at the Chinese University of Hong Kong published a paper titled "From Samples to Scenarios: A New Paradigm for Probabilistic Forecasting" (arXiv:2509.19975), which has been accepted at ICLR 2026. The paper introduces TimePrism, a model built to validate the proposed "probabilistic scenarios" paradigm.

TimePrism: From Samples to Scenarios

Traditional probabilistic forecasting relies on generating many samples to estimate a distribution. TimePrism changes the learning target to directly output a set of representative future scenarios, each with an explicit probability weight: {scenario, probability}. This allows a decision system to compare expected costs across scenarios or reserve resources for low-probability high-impact outcomes.

The paradigm is illustrated in the paper's Figure 1, showing a single forward pass producing multiple future trajectories with associated probabilities.

Architecture: Three Linear Layers for Scenarios and Probabilities

To test whether a simple structure can realize the new paradigm, TimePrism uses only three parallel linear layers: two branches generate future scenarios, and one branch learns the corresponding probabilities. During training, the scenario branches learn to cover diverse true futures, while the probability branch learns their weights.

The authors deliberately chose this minimal design to verify the hypothesis that a well-defined learning objective and output format can enable simple architectures to perform complex probabilistic forecasting.

Experimental Results

On five benchmark datasets, TimePrism was compared against multiple strong baselines including diffusion models, flow models, and Transformer-based models. Using two main metrics, TimePrism achieved the best result in nine out of ten comparisons, with very low computational overhead.

Figure 4 in the paper shows qualitative comparisons: TimePrism's probabilistic scenarios (top row) display both high-probability (thick blue) and low-probability (thin red) futures after similar histories, while the baseline TACTiS-2 relies on sampling (bottom row).

Connection to DLinear and Research Philosophy

This work continues the research philosophy of Professor Xu's team. In 2023, they published "Are Transformers Effective for Time Series Forecasting?" (known as DLinear), which used a simple linear model to challenge Transformer-based forecasting methods, garnering over 6,000 citations on Google Scholar by September 2026. That paper prompted the field to re-examine the relationship between task assumptions and model architecture.

TimePrism deliberately adopts a simple linear architecture and separate modeling of scenarios and probabilities, echoing the DLinear approach of using minimal structure to validate a core idea.

Jev vs. TimePrism: Convergent Design Principles

Both systems share three key design principles:

Explicit probabilities as the interface between model and decision system.

Parallel output : Jev answers multiple judgment questions in parallel; TimePrism outputs the full set of scenarios and probabilities in one forward pass, avoiding repeated sampling.

Decision-oriented design : The output format is shaped by what downstream decision-making requires.

In both cases, the computation is reorganized around the final needed result.

Differences and Broader Implications

Jev operates on natural language and program state, with developers defining questions or candidate options for the model to judge. TimePrism addresses future uncertainty, letting the model learn both the candidate futures and their probabilities. Jev provides a unified judgment interface for diverse semantic tasks; TimePrism proposes an extensible learning paradigm for joint scenario-probability learning.

The article suggests the probabilistic scenarios paradigm could benefit robotics (comparing motion trajectories), planning systems (considering multiple subsequent states), and resource scheduling (weighing demand scenarios).

Conclusion

TimePrism offers a clean starting point: first clarify what information decisions need, then design how the model learns and delivers that information. Jev's popularity has brought renewed attention to this line of research. The CUHK team's September 2025 paper already articulated the core proposition: let AI see multiple possibilities at once and hand them to the next decision step.

Paper: https://arxiv.org/abs/2509.19975 Code:

https://github.com/Fifthky/TimePrism