Here we describe the high-level algorithms associated with each model.
We use multiple models for each prediction we make. A model is associated with a specific task: 1) Is this media AI-generated or not, or 2) Is this media likely true or false. Here we describe the high-level algorithms associated with each model. The models that were developed at Georgetown will be shared over the next few months. If you have an open model that you think would perform well against our existing models, send us an email so we can test it.
We have models for Audio and Images. Video models are coming soon.
EFFORT takes the singular value decomposition (SVD) of each linear weight matrix in the self-attention blocks of every transformer layer, the top-r principal components are frozen to keep the pretrained semantic knowledge. Only the residual subspace is trained to learn forgery cues.
This model is a binary real/fake classifier on top of adapted CLIP features. The focus is semantic-level features with forgery discrimination confined to an orthogonal, low-capacity subspace, rather than explicit frequency or noise-artifact modeling.
This particular model is based on the Mixture-of-Experts (MoE) concept.
It dynamically selects and activates the relevant subset of parameters based on the characteristics of the input data.
The training objective uses a loss function combined to four different forms of the losses. The first one is the BCE (binary cross-entropy) loss fused for the prediction. The next term is again a BCE loss but for each expert but for each over its available samples. Now the third term determines the load balance for each of the experts in the router/decoder. Then we finally penalize the orthogonal and singular values.
This module features two LLM-based fact-checking systems, PLAN and ADAPT, that are designed to verify textual claims using evidence retrieved from the web.
Both systems are built on open-weight LLMs from the Qwen model family. Both are connected to the same live web search engine to retrieve the necessary information before evaluating a claim. They differ in how they are trained to search for evidence and reach their verdicts.
Each system returns one of three verdicts:
PLAN is a structured pipeline that uses multiple specialized LLMs, each trained to perform a specific task in the fact-checking process.
PLAN follows a five-stage verification process:
1 - Search query generation
2 - Search evidence retrieval
3 - Evidence filtering to remove low quality information
4 - Check evidence sufficiency and supplement new evidence
5 - Verify claim
The verifier model is finetuned to evaluate the claim’s truthfulness using the available evidence.
ADAPT is fact-checking agent trained with reinforcement learning to search for evidence and reason about what it finds to verify the claim
Unlike PLAN, ADAPT does not follow a verification workflow with fixed steps. It dynamically alternates between three actions to investigate a claim: think, search, verify.
ADAPT is trained using reinforcement learning where the model learns to imprve its search and reasoning strategies using feedback on its own attempts.