Published in JMLR · 2026 · 27(229)

COAT / Causal Representation Assistant

Discover variables
in unstructured data.

COAT discovers measurable attributes in reviews, news and other observations. Turn them into variables for your analysis.

Language models propose and measure candidate factors. Statistical feedback guides further discovery, causal-structure analysis and effect estimation.

One review, three measured attributes
“Very sweet apples, a rich aroma, but rather small.”

Define what each attribute means, then apply the rule to every review.

SweetnessVery sweet
AromaStrong
SizeSmall

COAT in one minute

From reviews to measured variables, then statistical feedback.

Transcript

Both reviews say the apples are sweet. So why are the ratings different? Perhaps something is missing.

First, a language model proposes variables, such as sweetness,

and defines how to measure them.

Statistical methods can then analyze relationships,

rather than letting the language model decide what causes what.

But analysis is not the end. Looking back, we notice a clue: bruising.

We add a measurement and analyze again.

Observations, variables, values, and statistical feedback form a loop.

The same approach can discover factors in news,

measure climate variables with tools,

or account for customer differences when estimating how emails affect website visits.

This is COAT-Learn: language models and statistics working together, from observations to analysis.

Examples

From observations
to measured factors.

Define an attribute, then apply the same measurement rule to every observation.

Interactive example

Observation

Rating: 5/5
A rich apple aroma, very sweet taste, and small apples that fit in my lunch box.

Measurement rule

Sweetness

2

Sweetness explicitly described in a review.

0
Explicitly not sweet
1
Mildly sweet
2
Very sweet

The highlighted passage supports this level under the displayed rule.

Inspect the measurement table

Select a value to inspect its observation and rule. Try R03, then select Aroma.

One rule across observations makes a column.
SampleSweetnessAromaSize
R01
R02
R03

Observation

Paired target: diagnosis class
Select a region to inspect its measurement rule.

Measurement rule

Local brightness

1

A visual attribute in a specified observation region.

0
No clear local bright area
1
A marked local bright area

Select a region to inspect its measurement rule.

Inspect the measurement table

Select a value to inspect its observation and rule.

One rule across observations makes a column.
SampleLocal brightnessShape difference
M01
M02

Observation

Example news
The company announced a new cloud technology platform and described its research team's work.

Measurement rule

Technical progress

1

Distinguish routine maintenance from a new technical development.

0
Routine maintenance
1
A new technology or technical breakthrough

The highlighted passage supports this level under the displayed rule.

Inspect the measurement table

Select a value to inspect its observation and rule.

One rule across observations makes a column.
SampleTechnical progressMarket expansion
N01
N02

Define the region and quantity, then use code to calculate each measurement.

Quantity
Near-surface air temperature
Region
Niño3 · 5°S–5°N, 150°W–90°W

Measurement definition / illustrative Python

field = archive["air_temperature"].sel(
    latitude=slice(-5, 5),
    longitude=slice(210, 270),
)
factor = field.mean(("latitude", "longitude"))
Niño3 region in the tropical Pacific · Natural Earth map.Niño3
Niño3 region in the tropical Pacific · Natural Earth map.
One value for each month / °C
JanFebMarAprMayJun
26.827.227.928.227.727.4

How COAT works

Statistical feedback guides
the next factor proposal.

COAT links factor discovery with statistical analysis in a feedback loop.

  1. 01

    Propose

    A chat model reads selected observations and defines candidate factors.

  2. 02

    Measure

    An annotation model applies the same definitions across observations.

  3. 03

    Assess

    Conditional-independence tests ask what information remains after other variables are considered.

  4. 04

    Refine

    Statistical feedback directs the next proposal toward information the current factors leave unexplained.

View the paper's framework figure
Paper figure showing factor proposal, annotation, statistical causal discovery and feedback.

Analysis goals

Choose the result
your question needs.

Discover factors, explore structure, or estimate effects.

COAT-MB

Which attributes jointly inform a rating?

Markov blanket

Discover a compact set of factors that jointly inform a target.

Input
One numeric target + paired text.
Output
Factor definitions and a selected set.
View inputs and Python interface
COAT-PAG

What structure connects my variables?

Partial ancestral graph

Discover additional factors to refine the relationships among variables.

Input
Existing numeric variables + paired text.
Output
An expanded variable set and a PAG.
View inputs and Python interface
COAT-ADJ

How much would an action change an outcome?

Adjustment sets and effect estimates

Find adjustment variables and estimate the average effect of an action.

Input
A 0/1 treatment, numeric outcome + paired text.
Output
Graph and adjustment analysis.
View inputs and Python interface

Apply existing factor definitions to your data, or analyze a measured table and graph.

Annotate existing factorsAnalyze an existing graph

Get started

Configure models. Run an analysis.

01 / MODELS

Configure models

Main model

Model source

Model

Reasoning effort

Connection settings

Annotation model

Model type

LLM connection

Save model connections · Shell
02 / ANALYSIS

Run an analysis

Goal

Data

Advanced settings

Number of annotation tasks processed at once; actual request counts are also affected by model connection rate limits.

Run analysis

Research cases

Climate measurements
and news factors.

See how COAT explores climate and financial news.

Climate / ENSO

Which factors relate to next month's sea-temperature change?

A tool-using model defines regional measurements in NOAA reanalysis data. The study elicits 13 factors, recovers literature-consistent relationships and proposes a soil-temperature hypothesis for further investigation.

View the published graph
Paper ENSO graph with 13 climate factors and a sea-temperature-change target.
Read the JMLR paper

News / Microsoft

Which parts of news inform future returns?

The study separates innovation, product, legal and strategic information. Factors discovered in the first 200 trading days are examined with strategies on the following 400 days.

View the published graph
Paper graph showing four news factors and return rate.
Read the JMLR paper

The published work

Discovering and Reasoning of Causality in the Hidden World with Large Language Models

Chenxi Liu · Yongqiang Chen · Tongliang Liu · Mingming Gong · James Cheng · Bo Han · Kun Zhang

Chenxi Liu and Yongqiang Chen contributed equally. Corresponding author: Bo Han.

Journal of Machine Learning Research
27(229):1–95, 2026

The journal extension adds COAT-PAG, COAT-ADJ and their identifiability analysis to the NeurIPS 2024 framework.

Citation / official JMLR BibTeX
@article{JMLR:v27:25-2137,
  author  = {Chenxi Liu and Yongqiang Chen and Tongliang Liu and Mingming Gong and James Cheng and Bo Han and Kun Zhang},
  title   = {Discovering and Reasoning of Causality in the Hidden World with Large Language Models},
  journal = {Journal of Machine Learning Research},
  year    = {2026},
  volume  = {27},
  number  = {229},
  pages   = {1--95},
  url     = {http://jmlr.org/papers/v27/25-2137.html}
}