Top-one accuracy
12,000 evaluation decisions from test and out-of-distribution files.
MoLeMo Lab / research project
MoJev
Typed decisions.
Calibrated probabilities.
MoJev supports up to 1M tokens of context and multimodal text and image inputs. A caller-defined schema turns this evidence into a probability distribution over the permitted outcomes—in one forward pass.
One passmany typed decisionsMoLeMo Lab contact@molemo.org
Research overview01 / The decision architecture
From an open-ended state
to a finite decision.
MoJev shares the state across schema fields, isolates candidate branches with TreePacked Attention, and reads out a distribution through a typed decision head.
Inside the method
Architecture figure from the MoJev preprint. Open for the full-resolution vector diagram.
Expected calibration error
Accuracy and confidence measured on the same evaluation decisions.
Per-decision inference
Average per decision at batch 8, with no autoregressive output tokens.
Context length
Supports up to one million tokens of context.
01 / The mechanism
One state.
Many possible outcomes.
Every decision is defined at request time. MoJev scores the named candidates, normalizes their utilities within each field, and returns a probability law in the caller's schema.
Define the decision
Supply a state, a question, and the allowed values for a Choice, Noul, or Score field.
Share the state
TreePacked Attention lets candidate branches read the same evidence without reading one another.
Return a distribution
The typed head converts candidate utilities into probabilities and a schema-bound result.
TreePacked Attention
Parallel branches.
Clear boundaries.
Questions attend to the shared state. Each candidate attends to the state, its own question, and its own tokens. This visibility pattern evaluates multiple fields in one packed backbone pass.
Read the method
Learning the probabilities
RLCD defines the typed decision problem. CDPO combines grouped preference learning with Brier scoring of observed outcomes; the packed predictions feed the loss in the same training pass.
02 / Experimental evidence
Measure the distribution.
Then use it.
The preprint tests decision quality, sequential search, visual evidence, and uncertainty. Select an experiment to see its figure and reported result.
Prediction & uncertainty
Know when a decision is reliable.
MoJev reaches 93.23% top-one accuracy and 0.79% expected calibration error on 12,000 decisions. Ranking decisions by normalized predictive entropy exposes high-error cases for selective prediction.
0.9750Error-detection AUROC on the test subset
Selective-prediction figure: retained coverage levels on a 400-example test subset and an OOD split.
Closed-loop optimization
Choose the next reaction to run.
On a measured direct-arylation grid, MoJev-guided search reaches 94.97% mean best yield after 38 experiments, versus 90.14% for random search.
+4.83 ppMean best-yield difference after 38 experiments
Five paired seeds on the Shields et al. direct-arylation benchmark. Each run begins with eight identical experiments, then evaluates three new conditions per round. Bands show ±1 sample standard deviation.
Multimodal intervention
Change the image. Keep the question.
A tabby photograph replaces a gray control while the question and candidate set stay fixed. For the four-outcome choice, predicted P(cat) rises from 0.207 to 0.773.
+56.6 ppChange in predicted cat probability
Two- and four-outcome distributions use the same visual intervention. Photograph: Alvesgaspar / Wikimedia Commons, CC BY-SA 3.0.
Uncertainty modeling
Measure how the evidence resolves uncertainty.
MoJev returns a probability for every candidate. Normalized predictive entropy measures how concentrated that distribution is: 0 when one outcome has all the probability, and 1 when all outcomes are equally likely.
H_{\mathrm{norm}}(p) = -\frac{1}{\log K}\sum_{i=1}^{K} p_i\log p_i
For the gray control and the cat photograph, compute entropy over the same candidate set and compare it alongside P(cat). The probability tracks support for the cat; entropy tracks uncertainty across all K outcomes.
03 / Where distributions matter
From prediction
to intervention.
A calibrated distribution is useful beyond selecting the top candidate: it supports belief updates, information acquisition, and decisions under a specified mechanism.
Causal inference
Reason about a change in the world.
Potential outcomes, interventions, and sequential evidence share a distribution-valued decision interface.
Drone-control mechanism
Probe which variables move risk.
With the scene fixed, changing collision thresholds or braking parameters reveals the response of the predicted risk distribution.
04 / Research resources
Read. Run. Inspect.
The paper, implementation, trained checkpoint, and MoJev-Mix data are available from the MoLeMo Lab repositories.
01 / PREPRINTRead the paperMethod, results, and appendices↗
02 / CODEExplore the implementationRuntime, examples, and evaluation↗
03 / MODELOpen the checkpointMoJev on Hugging Face↗
04 / DATASETExplore MoJev-MixTraining and evaluation mixture↗
Citation
Cite the preprint
@misc{molemo2026mojev,
title = {MoJev: Schema-Conditioned Causal Decision Learning with Large Language Models},
author = {{MoLeMo Lab}},
year = {2026},
note = {Preprint},
url = {https://github.com/MoLeMo-Lab/mojev/blob/master/paper/mojev-preprint.pdf}
}