A Year Before Jev, CUHK Team Proposed Probabilistic Parallel Generation

In September 2026, Jev flooded screens, and discussion continued for several days. According to a Machine Heart report, this model directly outputs judgments and probabilities: whether a piece of information should be retained, which action to choose next, whether a task is complete—the developer provides text or program state and defines the question and options, and the model returns structured results such as probabilities and scores for the program to use.

Tasks of this kind usually do not require long answers. Jev adopts a judgment-oriented output approach and supports processing multiple questions in parallel for the same input, forming a clear distinction from token-by-token text generation and making the many frequent small decisions in Agents a natural application scenario. The focus of the spread therefore lands on how AI enters software's decision process more directly.

It should be noted that the above description of Jev's capabilities and demonstrations comes from the Machine Heart report and is treated as a source_claim; this article does not introduce third-party verification beyond that report.

The Machine Heart report states that the paper "From Samples to Scenarios: A New Paradigm for Probabilistic Forecasting" by Professor Xu Qiang's team at the Chinese University of Hong Kong was made public on September 24, 2025, and has now been accepted by ICLR 2026; TimePrism is the validation model designed for this paradigm.

Common probabilistic forecasting methods express possibility through sampling: repeatedly generating future trajectories, then estimating the distribution from the sample set. More samples usually mean more sufficient statistical information, but generation and processing costs also increase. The "probabilistic scenarios" paradigm proposed by TimePrism writes the learning objective directly as {scenario, probability}; the model learns both representative future scenarios and explicitly learns the probability distribution over these scenarios; at inference time, a single forward computation can yield the entire set of results in parallel.

This means probability is no longer reflected only by the frequency of a large number of repeated samples. Each scenario carries its own weight, and the decision system can directly calculate the expected cost of different plans under these scenarios, or reserve resources for costly situations. The report uses backup power arrangement as an example: a single power generation forecast is not enough; the system needs to know what possible futures exist and how likely each is.

TimePrism's core learnable structure consists of 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 different real futures, and the probability branch learns the respective weights of these scenarios. The team deliberately adopted this concise design to test a key hypothesis: if the learning objective and output form are properly designed, a simple structure can also perform complex probabilistic forecasting well.

According to the Machine Heart report, the paper compared two main metrics on five benchmark datasets, and TimePrism achieved nine best results out of ten comparisons, outperforming various complex probabilistic forecasting models, including diffusion models, Flow models, and Transformer-based models, with very low computational cost. This result provides experimental support for the core idea that "changing the learning paradigm can allow a minimal structure to unleash stronger capability."

This research approach continues the consistent exploration of corresponding author Xu Qiang's team. In 2023, the team published "Are Transformers Effective for Time Series Forecasting?", challenging the Transformer forecasting methods of the time with a simple linear model, sparking intense discussion; the report states that as of September 2026, its Google Scholar citations have exceeded six thousand. This time, to validate the probabilistic scenarios paradigm, they deliberately chose a simple linear architecture and modeled scenarios and probabilities separately.

Putting the two works together, the most noteworthy common point is redesigning the model's output according to decision needs. Options and their probabilities, future scenarios and their weights, can all directly enter the judgment and computation of subsequent programs, so probability becomes a kind of interface between the model and the decision system. Both also embody the value of parallel processing: Jev can answer multiple judgment questions in parallel around the same input, reducing token-by-token generation overhead; TimePrism outputs an entire set of future scenarios and probabilities at once, eliminating repeated sampling during inference.

The differences are equally clear. Jev is oriented toward natural language and program state, usually with the developer defining the question or candidate options, and then the model making a judgment; TimePrism is oriented toward future uncertainty, and even "what candidate futures exist" is left to the model to learn. The former focuses on making judgments about a given question, while the latter focuses on generating future pictures with probabilities. This shows two different levels of generality: Jev uses a unified judgment interface to handle various semantic tasks, while TimePrism proposes a learning paradigm that can continue to be expanded.

Explicit probabilities, parallel output, and orientation toward decisions are the design ideas shared by the two paths. The report believes that the potential of probabilistic scenarios is worth exploring in broader tasks, such as robots comparing different motion trajectories, planning systems considering multiple subsequent states, and resource scheduling weighing different demand scenarios; in the future, it could also be combined with stronger backbone networks, more complex scenario representations, and planning and optimization methods. These are the report's outlook, not verified conclusions.

Jev's rise has brought new attention to this route. Looking back at the September 2025 paper, the "from sampling to scenarios" proposed by Xu Qiang's team at CUHK has already unfolded the same key proposition in probabilistic forecasting: let AI see multiple possibilities at once and hand them to the next step of decision-making. The concise starting point provided by TimePrism is: first clarify what information decisions need, then design how the model learns and delivers that information.

What needs to be distinguished is this: the paper's publication time, ICLR 2026 acceptance, the three linear layer structure, and nine best results out of ten comparisons come from Machine Heart's account of the paper and are treated here as source_claim; the description of Jev's capabilities likewise comes from that report. Whether the two represent an inevitable convergence of the same technological trend is an inference in this article based on the report, not a confirmed fact.