Vision-language models describe images fluently and, too often, wrongly. SHROOM-Visions 2026 asks systems to mark exactly which characters of such a description are hallucinated, and of what kind, in Chinese, English, French, and Italian. Working on this task we find that how a system reads predictions out of a model matters as much as which model it uses: the metric is a rank correlation against soft, multi-annotator labels, and aligning the prediction pipeline to it is worth as much as a larger encoder. We treat the problem as per-character sequence tagging. Starting from a fine-tuned XLM-RoBERTa tagger, we (i) formulate the task with one independent probability head per hallucination class, which matches the labelled and unlabelled correlation metrics; (ii) reconstruct spans in a way that respects the scorer’s tie-breaking behaviour; and (iii) tune per-class decision thresholds with cross-validated Bayesian search. These changes lift the score substantially with no change to the underlying model. Keeping this output layer fixed, we then compare encoders: a text-only tagger with added visual features, and a fine-tuned vision-language model. The vision-language model, adapted with low-rank updates and read out through the same five heads, is our strongest system in every language and improves most on the labelled score, which is the part of the task that needs the image. It also transfers cleanly from validation to test, in contrast to the tuned text ensembles, whose threshold choices overfit the validation split.

We describe the Dynamos submission to ImageEval 2026 Task 1b, image-grounded hallucination detection: given an image and three statements, exactly one of which is grounded, a system must label each true or false. The official ranking metric, Contrastive Instability (CI, lower is better), reduces exactly to one minus the accuracy of picking the grounded statement once a system always outputs a single true label, which removes an entire failure mode (marking zero or several statements true) at no cost. Building on this reduction, we score each statement with an open vision-language model and always output the highest-scoring one as true. Starting from zero-shot Qwen2.5-VL, we low-rank fine-tune the model with a selection objective, train multiple random-seed adapters, and average their scores. Our main finding is that the strongest single model is not the strongest ensemble member: a 32B adapter trained on more data performs substantially better on the labelled devtest split but generalises worse than smaller 7B adapters to the blind test set, yet combining the two model families improves over either pure-family ensemble on both languages, despite the 32B family generalising worse on its own. Our final submitted systems reach CI 0.033 in English (2nd of 8 teams on the official leaderboard) and 0.056 in Modern Standard Arabic (3rd of 6 teams), both well below our zero-shot baselines of 0.058 (English) and 0.110 (MSA).

For the ArGuard 2026 shared task on harmful content detection in Arabic memes (Task A) and LLM prompts (Task B), we fine-tune Arabic transformer encoders across all four subtasks. Memes additionally use a vision tower (ViT-B/16 for the binary subtask, CLIP ViT-L/14 for the fine-grained one) late-fused with MARBERTv2; ensembling choices and decision thresholds are tuned per subtask throughout. Two methodological choices account for much of our gains: selecting checkpoints by development macro-F1 rather than loss, and aligning the B2 training target with the evaluation distribution. Our largest single effect, however, was diagnostic. The B1 blind test set is approximately class-balanced while its training data is ~80/20, so our dev-tuned threshold collapsed macro-F1 from 0.9557 on dev to 0.6345 on test. We recovered the test class balance algebraically from the aggregate metrics the platform reported to all participants; re-thresholding the same saved probabilities, with no retraining, restored 0.8628. Because this uses only aggregate leaderboard feedback (within the task rules), the B1 figure is not directly comparable to teams that did not use the channel and is not a generalisation estimate. This correction is also not general: on the weaker multi-label classifier (A2), stronger prior matching reduced performance in every setting we tested, so matching the test prior is not necessarily F1-optimal. We additionally report a catalogue of quantified negative results, including external-data augmentation with AraSafe. Final test macro-F1: A1 0.7869, A2 0.3463, B1 0.8628, B2 0.7622.