FIRST AUTHOR · NEUROCOMPUTING · SUBMITTED, NOT ACCEPTED
Where do fusion gains come from in low-resource multimodal fake news detection? Operating points, archive artifacts and frozen CLIP baselines
The revised study asks where a score improvement comes from before attributing it to a new model capability.
The revised manuscript has been submitted to Neurocomputing. This status follows my update on 24 October 2026; it does not imply acceptance or a particular review stage.
What the revision examines
The manuscript reanalyzes saved predictions from 344 training runs. FCP is the token-level fusion model being examined against simple baselines with shared frozen CLIP features, splits and seeds; it is not presented as a new architecture.
01 / Compare the operating point
Some low-shot GossipCop Macro-F1 advantages track a change in the decision operating point; a higher score alone does not establish a stronger representation.
02 / Audit the data
Archive markers and near-duplicate content can confound evaluation on recollected news datasets.
03 / Account for the full budget
The manuscript limits training labels but uses a fully labeled validation set. The proposed work will count selection and calibration labels too.
What the evidence does not establish
- The study covers two English recollected news datasets, with limited seeds and small test sets. It does not establish a universal conclusion about multimodal fusion.
- The archived outputs are hard predictions. AUROC, score calibration and post-hoc threshold retuning cannot be recovered from those labels alone.
- Low-resource training still uses a fully labeled validation set for model selection; this is not an end-to-end few-label budget.
- KICP is linked as related implementation. This portfolio update neither reruns the benchmark nor declares the repository a complete archive of all revised-paper experiments.
How this informs my research plan
For multimodal emotion recognition, I propose counting training, model-selection and calibration labels together; auditing groups and near duplicates; and saving per-example scores and versions. Multiclass transfer, selective prediction and continual updates are proposed studies, not results of this paper.
Source: the revised manuscript supplied by the author (21 pages, preprint footer dated 14 October 2026) and the author’s status update of 24 October 2026. The footer date is not asserted to be the submission-system date.