Evolved Meta-paths Let Frozen LLMs Judge Fake News
A multi-agent genetic search breeds the graph paths a frozen language model reads, trading labeled training data for compressed propagation evidence.
In short
An eight-author team proposes MAGER, a framework that evolves meta-paths through a propagation graph so a frozen LLM can reason about a story's veracity without any task-specific training.
At a glance
- Preprint arXiv:2609.18597v1, submitted September 16, 2026, listed under cs.AI.
- Eight named authors, among them Erik Cambria and Litian Zhang.
- The abstract page states acceptance at ACM MM 2026 as an oral presentation.
- Code address given as github.com/SenticNet/MAGER; not independently checked.
- No datasets, baselines or measured results appear on the abstract page.
How a story spreads — who reposts it, in what order, answering whom — is one of the sharper signals for spotting a fabricated claim. A preprint posted September 16, 2026 argues that an off-the-shelf language model can read that spread pattern itself, provided something first compresses the graph into a few readable paths.
Why raw graphs break the prompt
Today's detectors are mostly supervised graph neural networks, which need large labeled corpora and, the authors argue, travel poorly to unfamiliar cases. Pasting the propagation graph into an LLM instead does not fix it. A node-and-edge structure is not the modality the model was trained to consume, and a real cascade contains far more nodes than a prompt can usefully hold.
Evolution as a feature selector
MAGER hands that compression job to a population of agents. Instead of a researcher hand-picking path templates such as user-to-repost-to-user, the framework generates candidates, scores them and recombines the survivors across generations. What comes out is a subgraph small enough to prompt with and, ideally, still carrying the evidence. A retrieval step then supplies worked examples that resemble the case both in wording and in shape, all while the model's weights stay untouched.
What the source does not establish
The abstract page claims substantial gains for frozen LLMs in data-efficient settings. It names no benchmark, no competing system and no number, so the size of that gain — and which models it holds for — cannot be verified from the page we read. The same page reports acceptance at ACM MM 2026 as an oral presentation and points to github.com/SenticNet/MAGER for code; we checked neither claim against a second source.
The part that generalizes
Strip away the fake-news framing and the contribution is a translation trick: turning network topology into something a text model can actually consume. Fraud rings, supply chains and citation networks pose the same mismatch. Whether evolved paths survive outside the authors' own setup is a question for replication with published figures.
FAQ
What is a meta-path?
A short template describing a walk through a graph, such as user to repost to user. MAGER searches for useful templates automatically instead of relying on hand-designed ones.
Does MAGER fine-tune the language model?
No. Per the abstract, the model stays frozen; adaptation comes from the evolved paths and the retrieved in-context examples rather than from updated weights.
Has this paper been peer reviewed?
The arXiv page states acceptance at ACM MM 2026 as an oral presentation. We report that as the page's own claim; it was not confirmed against a conference listing.
