AI News Feed
Market watch
Samsung Working on Fix for Home Up Good Lock Backup BugTrump-Xi Summit Puts AI Safety Talks on Agenda as Freeze Debate and Senate Limits Draw ScrutinyAI Takes Over Entry-Level Grunt Work, Forcing Companies to Rethink HiringGooglebook OS Demo Shows Mac-Like Gestures on Android FoundationTencent's WorkBuddy Enters Leapmotor, Cutting AI System Development Time by About 90%AIsphere Launches PixVerse R2, Bringing LLM-Style Scaling to Real-Time World Models
Samsung Working on Fix for Home Up Good Lock Backup BugTrump-Xi Summit Puts AI Safety Talks on Agenda as Freeze Debate and Senate Limits Draw ScrutinyAI Takes Over Entry-Level Grunt Work, Forcing Companies to Rethink HiringGooglebook OS Demo Shows Mac-Like Gestures on Android FoundationTencent's WorkBuddy Enters Leapmotor, Cutting AI System Development Time by About 90%AIsphere Launches PixVerse R2, Bringing LLM-Style Scaling to Real-Time World Models
Companies

Nokia Open-Sources AnyJev as Decision AI Splits Between Jev and Decitron

Nokia's applied research team has open-sourced AnyJev, a training-free Python library that turns open LLMs into calibrated decision models. It borrows Jev's typed decision interface and adds cyclic-shift and prior corrections, while QbitAI contrasts Jev's instant judgments with Decitron's future simulations.

AI News Feed · 2026-09-23 17:08

AnyJev borrows its interface from Jev, the System One decision model TypeSafe AI launched in September 2026. Users give it a typed question and get back a decision with a probability they can threshold. That probability is read from the model's next-token distribution; nothing is generated, parsed, or trained. The library supports three question types: a choice question that picks one of K options, a noul question that is yes or no, and a score question that places the answer in one of several ordered bins. It installs from PyPI, ships under Apache-2.0, and has transformers and vLLM backends with shared-prefix scoring.

The Nokia research team flags two flaws in the common shortcut of restricting the next token to option labels and reading the scores. First, the answer can change when options are reordered. Second, the probabilities are not calibrated. The causes are prior bias, where a model favors some labels such as "Yes" over "No" regardless of input, and position bias, where it favors certain slots in the option list.

AnyJev addresses these with two levels. L0, with zero labels and on by default, applies cyclic shifts: for a question with K options, the list is shown in K rotations so every option appears in every position once, and the results are combined in log space as a geometric mean. If position bias is additive in logit space, L0 removes it exactly. L0 also uses batch calibration by default, keeping a running mean of predicted distributions on real inputs and dividing it out at strength 0.75; correction starts after eight items. L0 costs K prefills per decision, batched over a shared prefix, which is about 0.25 seconds per decision at batch 32 on one H100 with K equal to 20. L1, for 100 to 500 labels per question, adds temperature scaling on top of L0. The fitted values are saved as a small JSON artifact. L1 reshapes confidence but does not change the ranking of answers.

On Qwen3-8B with BANKING77, a 20-way task with 300 test items, MarkTechPost reports that raw logits produce a 0.230 flip rate when options are reversed, an accuracy of 0.747, a calibration error of 0.240, and 7.7 percent auto-decidable traffic at 5 percent error. AnyJev L0 lowers the flip rate to 0.073, raises accuracy to 0.803, lowers calibration error to 0.184, and raises auto-decidable traffic to 46.3 percent. AnyJev L1, which requires 100 to 500 labels, produces a 0.077 flip rate, 0.807 accuracy, 0.095 calibration error, and 52.0 percent auto-decidable traffic. The repository says L0 reduced order flips on all nine model and task rows tested. On a typed-decisions set, Qwen3-32B with L1 reached an ECE of 0.036, compared with 0.144 published for Jev. On accuracy, the fine-tuned Laya still leads. The full ablation table covers Qwen, OLMo, Granite, Phi and Mistral models. Wu says the team tried AnyJev on an internal Nokia routing problem and saw promising results.

For use, AnyJev can be installed with pip install anyjev[hf] and run through a Hugging Face backend, or served by starting vLLM with prefix caching and pointing a VLLMBackend at it.

The release lands amid a broader debate over what "decision AI" means. QbitAI reported that Jev, from the OpenAI-background team TypeSafe AI, focuses on immediate judgments for software and agents, with the slogan "Decisions, not strings." Its outputs are choices, scores, and probabilities that software can execute directly. Jev turns traditional if/else logic into an AI version: instead of hard-coding "if A, then B," a program can run B when the model judges A's probability to be above 80 percent. Its goal is speed, low cost, stability, and direct embedding in software and agents.

Decitron, released in June 2026 by Zhongke Wenge, which some media have called the "first stock in decision large models," takes another route. It models reality as states, agents, actions, and outcomes, then uses world models, multi-agent simulation, game theory, and solving to compare paths. QbitAI describes the split this way: Jev makes a decision an instantaneous judgment; Decitron turns a decision into a simulation of the future. One solves "what to choose now," while the other asks "after choosing, what happens next." Both explicitly quantify uncertainty—Jev for the current judgment, Decitron for future evolution—and both try to make decision-making an independent AI computation paradigm rather than a byproduct of text generation.