From Text Decisions to Pixels
Paper introducing PixelJev, a native-image decision interface on small open multimodal models with few-shot adaptation.
This summary is based on how curated lists describe the project; its README was not available to review.
Overview
This arXiv paper extends typed decisions from text to images with PixelJev, a native-image decision interface built on small open multimodal models. According to the list summary, 64-shot adaptation raises accuracy on the Pets image dataset from 60.13% to 92.40%.
A key finding is that calibration does not follow accuracy gains: better accuracy after adaptation does not automatically mean better-calibrated probabilities. That matters for anyone thresholding image decisions. The paper itself was not reviewed here.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which class does this image belong to?(inferred) | — |
How it works
- Pose image classification as typed decisions
- Adapt with few labelled examples
- Measure accuracy and calibration
Strengths
- Large few-shot accuracy gains reported
- Calibration analysed separately from accuracy
Caveats
- List-only entry; paper not reviewed
- Calibration does not improve with accuracy
Getting started
See https://arxiv.org/abs/2609.29283.At a glance
- Type
- Paper
- Status
- research
Use cases and integrations
How curated lists describe it
PixelJev, a native-image decision interface on small open multimodal models; 64-shot adaptation lifts Pets accuracy from 60.13% to 92.40%, while calibration does not follow accuracy gains
AnotiaWang/awesome-decision-models
Frequently asked questions
What is PixelJev?
A native-image decision interface on small open multimodal models.
What accuracy gain is reported?
From 60.13% to 92.40% on Pets with 64-shot adaptation.
Does calibration improve too?
No, the summary says calibration does not follow accuracy gains.
Related projects
Research code for an image-and-text System One model that reads option-letter logits from a multimodal LLM, with training and calibration.
ProjectresearchPythonResearch toolkit that encodes an image once and answers several typed questions from a shared distribution, plus OpenJev-0.6B text models.
ModelresearchPythonFree hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
ProjectbetaPython- cbjev★ 4
Self-hosted encoder model giving Jev-compatible choice, score and noul answers in one pass, pitched as a better-calibrated Laya successor.
ProjectresearchPython Joke decision model trained to answer wrongly with total confidence: 1.9% accuracy at 0.957 mean confidence, via a Jev-style API.
ProjectresearchPython