Editorial illustration for Nokia's AnyJev Fixes LLM Biases for Decision Models Without Training
Nokia's AnyJev Removes LLM Bias Without Retraining
Nokia's AnyJev Fixes LLM Biases for Decision Models Without Training
Nokia's applied research team has put a new tool on PyPI called AnyJev, and it does something narrower than most LLM releases: it turns an existing open model into a decision-maker without touching its weights. No fine-tuning, no retraining, just a Python library that sits on top of models running through transformers or vLLM backends. The license is Apache-2.0, so anyone can pull it down and start using it today.
The problem AnyJev targets is specific. Production systems often don't need an LLM to write a paragraph. They need it to pick option A, B, or C, or answer yes or no, or slot a response into one of five ranked bins.
Nokia's library handles all three cases, borrowing its question format from Jev, the decision model TypeSafe AI launched in September 2026. Ask AnyJev a typed question and it hands back a choice along with a probability, pulled straight from the model's next-token distribution rather than generated as text.
That last detail matters more than it sounds. Reading probabilities off logits directly, a trick plenty of open projects already lean on, turns out to have real problems baked into it.
Nokia’s applied research team has open-sourced AnyJev, a Python library that turns an open LLM into a decision model. It needs no training. It targets a common production job: picking one answer from a fixed set instead of writing a sentence.
Why this matters
For anyone shipping LLMs into production classification jobs, AnyJev is worth a look precisely because it doesn't ask for anything. No fine-tuning run, no labeled dataset, no GPU cluster reserved for a week. You install it from PyPI, point it at an existing open model, and get calibrated scores back.
That's a real gap it's filling: prior bias and position bias are the kind of quiet failure modes that don't show up in a demo but wreck accuracy once you're routing thousands of real decisions a day. Nokia naming the two causes explicitly, and building L0/L1 levels around them, suggests this came out of actual production pain, not a research paper looking for an application.
We'd flag the obvious question before adopting it: how does AnyJev's shared-prefix scoring hold up on option sets larger than a handful of labels, and does the Apache-2.0 license mean Nokia expects outside contributors to harden the vLLM backend further. Worth testing against your own labeled eval set before trusting it blind. Watch for community benchmarks against fine-tuned classifiers over the next few months.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives OpenAI's GPT-6 Sol Outperforms Claude Opus at One-Tenth the Cost
From Archives OpenAI's GPT-6 Models Halve Prices but Show Minimal Performance Gains
Common Questions Answered
What is AnyJev and how does it work without fine-tuning?
AnyJev is a Python library developed by Nokia's applied research team that transforms an existing open LLM into a decision-making model without requiring any training or weight modifications. It sits on top of models running through transformers or vLLM backends, allowing users to install it from PyPI and immediately get calibrated scores for classification tasks without needing fine-tuning, labeled datasets, or GPU resources.
What specific production problem does AnyJev solve for LLM deployments?
AnyJev addresses the common production requirement of selecting one answer from a fixed set of options rather than generating free-form text responses. It specifically targets bias issues like prior bias and position bias that can silently degrade accuracy in production classification jobs when routing thousands of requests, problems that don't typically appear during demos.
What are the key advantages of using AnyJev over traditional LLM fine-tuning approaches?
AnyJev requires no fine-tuning runs, labeled datasets, or reserved GPU clusters, making it significantly more accessible for production deployments. Users can simply install the Apache-2.0 licensed library from PyPI, point it at an existing open model, and immediately receive calibrated decision scores, eliminating the infrastructure and data preparation overhead associated with traditional model retraining.
Which backends does AnyJev support for running LLMs?
AnyJev is compatible with models running through both transformers and vLLM backends, providing flexibility in how users deploy their existing open language models. This compatibility ensures that organizations can integrate AnyJev into their current LLM infrastructure without requiring backend changes or migrations.
Further Reading
- nokia-applied-research/AnyJev: Turn any LLM into a Jev-style decision model - GitHub
- Jev Projects Directory: 145 Tools, 103 Reviewed | MrJev - MrJev
- Research and data - What people are building with Jev - Made with Jev
- Jev GitHub repos: 195 open-source projects — Made with Jev - Made with Jev
- Jev model · GitHub Topics - GitHub
