Robust Poverty LLMs that Support Poverty Elimination
Jul 8, 2026
In this session, Robert Krueger and Esther Mao of Worcester Polytechnic Institute (WPI) presented research focused on developing language models capable of better understanding multidimensional poverty and supporting families, frontline workers, and public policymakers.
The central question was whether current generative models could be improved using data specifically related to poverty, development programs, international investment, and, most importantly, information defined by families themselves through the Poverty Stoplight.
HOW LANGUAGE MODELS LEARN
Large language models, known as LLMs, are a type of artificial intelligence capable of understanding and generating text. Their basic function is to predict the most likely word within a sequence.
To learn, they use large volumes of digital information, including books, academic articles, news reports, blogs, videos, government reports, and programming repositories.
The quantity of data matters, but so does its quality. Models learn the patterns, associations, and biases present in the information used to train them.
“The quality and quantity of the data can determine a model’s performance.”
When the available knowledge is incomplete or represents only certain perspectives, the responses produced by the model will also have limitations.
WHY CURRENT MODELS DO NOT UNDERSTAND POVERTY SUFFICIENTLY
Robert explained that generative models depend on knowledge that has been written down, digitized, and published.
This excludes a large volume of community experiences, orally transmitted knowledge, and locally developed solutions. As an example, he noted that when asked about the history of science and technology in sub-Saharan Africa, a model may begin its account with the arrival of Europeans, ignoring centuries of earlier knowledge.
Something similar happens with poverty. Much of the available data was produced by researchers, public officials, organizations, and funders. Families rarely participate directly in defining what it means to live in poverty or what their priorities should be.
Models can answer where poverty exists or what its causes are according to experts, but they have greater difficulty understanding:
• What families consider to be priorities.
• How different deprivations are connected.
• Which resources already exist within a community.
• Which solutions are viable in each territory.
• Which interventions have previously succeeded or failed.
They may also reproduce stereotypes. For example, some systems depict poverty in the United States primarily through images of African American people.
“If the context used to train the model is incorrect, the responses may also be incorrect.”
A MULTIDIMENSIONAL PERSPECTIVE DEFINED BY FAMILIES
Traditional assessments often analyze specific dimensions such as income, housing, health, or education. However, poverty is multidimensional, and its different components affect one another.
Robert mentioned the case of evaluators who considered housing to be a priority need for families in Rwanda. When they examined Poverty Stoplight data, they found that the families themselves did not necessarily place it among their immediate priorities.
This does not mean that housing is unimportant. It means that an effective intervention must recognize each family’s decisions, capabilities, and circumstances.
The research seeks to incorporate this perspective through data created and defined directly by people.
SOURCES USED TO TRAIN THE MODEL
The team combined four main sources of information:
• DHS and MICS surveys: These databases contain information from millions of households and individuals on health, education, housing, basic services, and other indicators used to measure multidimensional poverty.
• Development Experience Clearinghouse: This database contains documentation on USAID projects dating back to 1961, including identified problems, interventions, results, and lessons learned.
• International investment data: The team incorporated information on resources allocated to development programs by the United States, the United Kingdom, the World Bank, and the International Monetary Fund.
• Poverty Stoplight: The most innovative addition was Poverty Stoplight data. These records reflect assessments conducted by families themselves and organized through red, yellow, and green indicators.
“To better understand poverty, we need to listen to the people experiencing it.”
TRAINING AND INITIAL TESTING
The team began with a basic model that had no specialized knowledge of poverty. They then fine-tuned it using the different datasets.
During training, the model received approximately 80 percent of each dataset to learn patterns. The remaining 20 percent was then used to assess whether it could apply what it had learned to new situations.
Using multidimensional poverty data, the model received descriptions of families and had to classify them as extremely poor, poor, or not poor.
Using USAID documents, it had to analyze a development situation and suggest an intervention. The response was compared with the action that USAID had actually implemented.
Using Poverty Stoplight data, the model had to predict whether a family was red, yellow, or green for a specific indicator based on its other indicators and characteristics.
PRELIMINARY RESULTS
In tests involving USAID projects in the Philippines, the model was able to identify the general direction of the interventions. However, its responses still lacked some of the detail and sensitivity found in the original projects.
The team clarified that it had used a limited number of documents. It expects the model to improve when trained on a larger dataset.
Using Poverty Stoplight data, the model was able to identify relationships between household characteristics and different indicators. It could also predict the color of an indicator with some degree of success based on information from the others.
This could help identify patterns that are difficult to observe manually within very large databases.
The researchers emphasized that these findings are still exploratory. They do not constitute definitive validation, and more information, statistical testing, and evaluation are needed before the model can be used for consequential decisions.
POSSIBLE APPLICATIONS
A specialized model could support different levels of the poverty elimination process.
For families A family could ask what other people with similar conditions and priorities have done, or which actions tend to support progress across several indicators.
It could also receive information on the sequence in which other families addressed particular deprivations.
The tool should not make decisions for the person, but rather offer options to support reflection.
For frontline workers Facilitators and social workers could use the model to summarize conversations, identify patterns, and suggest additional questions.
This could be particularly useful for professionals managing large caseloads and working under pressure. However, any recommendation would need to be reviewed because the system may make mistakes.
For organizations and governments The model could compare needs expressed by communities with the interventions receiving funding.
Robert presented the example of an institution that decides to build a bridge to stimulate a local economy even though the community already uses a boat service and considers access to electricity or drinking water more urgent.
A tool connected to household priorities could help identify these differences before resources are invested.
MODELS ADAPTED TO EACH TERRITORY
Participants emphasized that a single global model may not correctly understand the languages, expressions, conditions, and priorities of every community.
Esther explained that it is possible to develop models adapted to specific regions, provided that sufficient high-quality local information is available.
The first step would be to define which questions the model should answer and what data would be required. An organization in Hawaiʻi, for example, could build a tool based on the specific realities of its islands and communities.
Once implemented, the system could continue learning through user feedback. However, its responses should never be treated as absolute truths.
A TOOL WITHIN A HUMAN PROCESS
The session emphasized that AI can suggest patterns, questions, or alternatives, but it should not replace human decisions.
Predictive models can make mistakes and reproduce inequalities. This is particularly sensitive when working with social services, vulnerable populations, or decisions that directly affect people’s lives.
The tool must be used transparently and should allow families, social workers, and organizations to question or reject its recommendations.
“Technology is a tool, not an authority.”
QUESTIONS AND ANSWERS
• How was the model’s performance evaluated statistically?
A participant asked which statistical method was used to determine that the suggested interventions resembled the actual ones.
Esther explained that the test involving projects in the Philippines had not yet undergone a rigorous statistical evaluation. The sample included approximately 50 documents, and the results were reviewed qualitatively by the team.
She acknowledged that it would be necessary to expand the sample and develop more robust methods for measuring the model’s performance.
• How could it specifically help a family?
The example was raised of a family with limited income that must decide between buying clothing and addressing a dental need.
Robert explained that the model could analyze what other families in similar circumstances had done and show which decisions were associated with subsequent progress.
It could also consider the order in which certain actions facilitated improvements in other indicators. The final decision would remain with the family.
• Could it support social workers with large caseloads?
A participant asked whether the model could analyze support conversations and identify needs that a social worker had not noticed.
Robert explained that it could summarize multiple interactions, identify patterns, and suggest new questions. This could support professionals working with constant emergencies and large volumes of information.
However, the suggestions would need to be reviewed carefully because the model could misinterpret a conversation.
• Is it possible to build truly local models?
Participants from Hawaiʻi and Paraguay asked whether an LLM could understand differences between territories, languages, and cultures.
Esther responded that it is technically possible, but depends on having sufficient high-quality local data. It is also necessary to clearly define the problem the model is intended to solve.
Local models could learn from user feedback, but their development would require collaboration among organizations, communities, and technical specialists.
• How can people be prevented from placing too much trust in its responses?
In response to a question about truth, empathy, and responsibility, Esther noted that users must understand that AI is only one tool within a broader set of resources.
Organizations must explain its limitations and avoid presenting it as an absolute source of truth. Its role should be to generate new questions and options, not replace human judgment.
Get Involved
Interested in attending future virtual and in-person events from the Global Stoplight Network? Subscribe to our monthly newsletter here.
You can learn more about the upcoming Cerrito Forum 2026, the annual international global development event that brings together global stakeholders committed to the elimination of poverty from October 12th - 16th by visiting cerrito.com.py.
Back to all webinars