MAGER evolves meta-paths so frozen LLMs spot fake news
Eight researchers let genetic agents evolve meta-paths through news cascades, aiming to let frozen language models judge veracity without fine-tuning.
In short
MAGER runs a multi-agent genetic search to evolve the meta-paths that compress a story's propagation graph into something a frozen language model can actually reason over.
At a glance
- Preprint arXiv:2609.18597v1, submitted September 16, 2026
- Eight authors; the listing states acceptance at ACM MM 2026 as an oral paper
- MAGER is a multi-agent genetic evolution framework that discovers meta-paths automatically instead of hand-designing them
- Second component: graph in-context learning retrieves semantically and structurally similar cases as prompt demonstrations
- The abstract page names no datasets, base models or scores — accuracy cannot be checked from it
The claim behind MAGER is narrow and testable: a language model that is never fine-tuned can judge a story's veracity from how it spread, provided the graph is cut down to the right few paths first. The framework searches for those paths with a genetic loop run by multiple agents, then hands the resulting subgraph to the frozen model.
Why propagation is hard to feed to an LLM
Supervised graph neural networks are the established way to read diffusion patterns. They work, at the cost of large labeled corpora and brittle transfer to new events or platforms. Language models bring reasoning that transfers, but a raw cascade is the wrong shape of input for them — the paper separates that into a modality mismatch and plain information overload once the edge list grows.
Meta-paths as the compression step
A meta-path is a short typed template — user to post to reply, for instance — that selects a slice of the graph rather than all of it. Picking good templates is normally a matter of domain expertise. MAGER treats it as a search problem instead: agents propose candidates, score them, recombine the survivors and repeat, with the fitness signal tied to how well a language model reasons over what comes out.
Retrieval on top
The second piece is graph in-context learning. Rather than fixed few-shot examples, the system pulls in cases that resemble the target both semantically and structurally and places them in the prompt as demonstrations. The stated target is a standalone detector that holds up in data-efficient settings, not a module bolted onto a trained classifier.
What we could not verify
The abstract reports substantial gains but names no benchmark, no dataset and no base model, and the listing carries no results table. That leaves the size of the improvement, the models it was measured on and the comparison against supervised baselines unchecked from this source. We read the abstract page only, not the full text.
FAQ
What is a meta-path in this paper?
A short typed template through the graph, such as user to post to reply. It selects a compact, informative slice instead of passing the entire propagation tree to the model.
Does MAGER fine-tune the language model?
No. The paper describes frozen models as the detector; what gets optimized is the set of meta-paths and the retrieval of demonstration cases.
How accurate is MAGER, and where is it published?
Accuracy cannot be judged from the abstract page: it lists no dataset, baseline or score. The entry records acceptance at ACM MM 2026 as an oral paper, and the linked version is preprint v1 from September 16, 2026.
