Nokia Open-Sources AnyJev: A Training-Free Layer That Turns Any Open LLM Into a Calibrated Decision Model
MarkTechPost Asif Razzaq ● Covered by 18 sources
Nokia open-sourced AnyJev, a no-training layer that turns open LLMs into decision models. It cuts bias and boosts calibrated yes/no or choice answers.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Nokia’s applied research team has put AnyJev out in the open, and the pitch is refreshingly specific: take an open LLM and make it behave like a typed decision engine, not a chatty paragraph machine. The library is in Python, ships on PyPI under Apache-2.0, and hooks into transformers and vLLM. It does one job well. You hand it a typed question, it gives back a decision and a probability you can threshold.
The trick is that AnyJev does not generate free-form text at all. It reads the model’s next-token distribution and uses that to score fixed answers. That covers three cases: choice questions with K options, yes-or-no questions, and score questions that bucket an answer into ordered bins. Nokia says the appeal is obvious in production settings where teams need a single answer, not a mini-essay.
The research team’s complaint is also obvious if you have spent any time around logit hacks. Many systems simply force the model to pick from the option labels and call it a day. AnyJev tries to fix two problems with that approach: the answer can flip when the options are reordered, and the resulting probabilities are poorly calibrated. Nokia points to prior bias, where a model likes labels such as “Yes” over “No” regardless of the prompt, and position bias, where certain slots in the list get favored.
AnyJev’s L0 level attacks both. It shows the same question in every cyclic rotation, then combines the results in log space so each option gets the same treatment. It also applies batch calibration, starting after 8 items, using a running mean of predicted distributions and dividing it out at strength 0.75. Nokia says L0 costs K prefills per decision, batched over a shared prefix, and lands at about 0.25 seconds per decision at batch 32 on one H100 when K equals 20.
L1 adds temperature scaling for larger label sets, from 100 to 500 labels, without changing the ranking. On Qwen3-8B with BANKING77, that matters: flip rate when options were reversed fell from 0.230 with raw logits to 0.073 with AnyJev L0, and auto-decidable traffic at 5% error rose from 7.7% to 52.0% with L1. The repo also reports better calibration on several other models, and Nokia says it has already tried the tool on an internal routing problem with promising results.
My take — AI-written commentary, not fact-checked reporting
This is the rare AI release that sounds less like a demo and more like plumbing. Good: open models should be made useful for boring decision work, not just costume themselves as chatbots. Also good: any method that embarrasses raw logits for being biased and uncalibrated deserves a seat at the table, preferably before the next vendor markets “reasoning” as a personality trait.
Read more about this at: MarkTechPost