ImageEval 2026 · Task 1b — Hallucination Detection (English + MSA)

System artefacts for the Ayn-VQA hallucination-detection subtask at ImageEval 2026, ArabicNLP 2026.

Given a culturally grounded image and three statements — exactly one grounded, the other two plausible hallucinations — label each true / false.

The idea the system is built on

The ranking metric is Contrastive Instability (CI): of the items with at least one correct label, the fraction not fully correct. Lower is better.

The release guarantees exactly one grounded statement per item. Emit exactly one true and a wrong pick still gets the third statement right, so every item enters the CI denominator and

CI = 1 - top-1 accuracy

Task 1b is therefore a 3-way choice, not three independent true/false judgements. Measured on identical per-statement scores, switching from an independent 0.5 threshold to exactly-one-true decoding moves CI from 0.1767 to 0.0520 — a 0.125 absolute gain from the decoding rule alone.

Results

system track split CI ↓ combined acc ↑
Official baseline (Qwen2.5-VL-3B) EN devtest 0.3133 0.6840
Official baseline MSA devtest 0.4900 0.5080
Qwen3-VL-8B · MCQ · argmax EN devtest 0.0480 0.9520
Qwen3-VL-8B · MCQ · argmax MSA devtest 0.1140 0.8860
bilingual blend MSA devtest 0.0540 0.9460
ensemble-6 MSA devtest 0.0460 0.9540
ensemble-8 both dev 0.0340 0.9660

Ensemble members (dev accuracy): qwenP3-en 0.958, perstmt-en 0.948, llama90b-en 0.940, nemo12-en 0.924, llama90b-msa 0.920, qwen-msa 0.914, nemo12-msa 0.896, gemma-en 0.882.

What worked, with evidence

lever effect (dev n=500)
exactly-one-true decoding 0.1767 → 0.0520 on identical scores
permutation averaging (3 orders) Qwen MCQ 0.054 → 0.042
8-model equal-weight ensemble 0.054 → 0.034
a larger VLM Llama-3.2-90B scored 0.060, worse than 4-bit Qwen3-VL-8B at 0.054

Model generation beat model scale. All comparisons use a paired bootstrap over identical items: the ensemble is significantly better than plain Qwen MCQ (+0.0221, 95% CI [+0.0100, +0.0360]) but not significantly better than permutation-averaged Qwen alone (+0.0081, [-0.0060, +0.0220]).

Dataset observations

  • Tasks 1a and 1b are the same items — identical ids, and 1a's correct option index equals 1b's grounded index on 3000/3000 train and 500/500 dev.
  • English and MSA share gold — verified 3500/3500, so one prediction file is valid on either leaderboard.
  • Image-blind length prior: picking the longest statement gets 0.540 top-1 on dev (chance 0.333). Ties must break to one index; counting every tied index inflates this to 0.81.
  • Grounded position is uniform (0.319 / 0.336 / 0.344) — no positional shortcut.

Contents

submissions/ Codabench archives · predictions/ prediction CSVs and per-item score audits · metrics/ metric panels and LaTeX tables · figures/ paper figures (PNG+PDF) · config/ prompts and run configuration · notebooks/ the end-to-end Colab pipeline.

Data and licence

Built on QCRI/AynVQA-ArabicNLP26, CC BY-NC 4.0, non-commercial research only. This repo contains no dataset content: no images, no audio, no statements, no gold labels.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support