Source-linked AI summary
Vroom-Vroom at SHROOM-Visions: A Multi-Judge Committee for Detecting Hallucinated Spans in Vision-Language Outputs
Toqeer Ehsan, Nico Penttilä, Richard Schmidt, Arash Hajikhani, Victoria Palacin
TL;DR
The paper addresses span-precise detection and classification of hallucinations in multilingual vision-language outputs, where both model errors and human span annotations are difficult to resolve. It combines diverse fine-tuned VLM judges through character-level voting and explores activation probes. The committee ranks first in three of four languages and places on the podium for every language and metric, while probe behavior depends strongly on the source model and remains mechanistically uncertain.
Problem
Span-precise multilingual hallucination detection is difficult because VLMs produce unsupported visual claims and human annotators disagree about hallucinated characters and boundaries.
Method
The paper combines independent fine-tuned VLM judges through character-level voting and additionally trains probes on VLM internal activations.
Results
The committee ranks first in three of four languages and, with a two-vote threshold, achieves the best correlation and IoU across all four languages on the labeled test split.
Takeaways & Limitations
Model diversity reflects annotator disagreement, and corroborated character-level voting can improve detection over individual judges within the evaluated multilingual task.
Takeaways & Limitations
The committee may miss long hallucinations not detected by multiple judges, especially in longer English responses, and its vote threshold may not generalize beyond the evaluated settings.
Abstract
from arXiv · showhide
This paper describes our submission to the SHROOM-Visions shared task on detecting and classifying hallucinated character spans in vision-language model outputs across four languages. We employ several fine-tuned vision-language models as independent annotators and combine their span predictions through character-level majority voting, and additionally explore activation probes. The approach ranks first in three of four languages and places on the podium in every language and metric. Our analysis indicates that disagreement among diverse models tracks disagreement among human annotators.
1 Introduction
Vision-language models generate fluent but unsupported visual claims, making span-precise hallucination detection difficult because human annotations also disagree. This paper addresses the challenge with diverse model judges, character-level aggregation, and activation probes.
- VLMs produce several visual hallucination types, including invention, miscounting, relational, misinterpretation, and character-recognition errors.
- Character-level detection is difficult because human evaluators disagree about whether spans are hallucinated and where their boundaries lie.
- The paper combines fine-tuned VLM judges through character-level voting and trains probes on internal activations.
- The committee aggregates multi-model span predictions into character-level probabilities and examines whether model diversity reflects annotation disagreement.
- The strategy ranks first in three of four languages.
2 Related Work
Related work frames vision-hallucination detection around how ground truth is sourced and how hallucinations are defined. Existing approaches include reference-based checks, rollout consistency, VLM judges, and internal-representation analysis.
- Vision-hallucination datasets differ in ground-truth sourcing and hallucination definition, with human annotations capturing natural errors but requiring more annotation effort.
- Model-free detectors use reference captions under closed-world assumptions or between-rollout consistency as an uncertainty proxy.
- A larger body of work uses secondary VLMs either as judges or by analyzing their internal representations.
- Internal-representation studies include analyses of values, concepts, monitoring signals, and layer dynamics.
3 Methodology
The methodology uses diverse VLM judges and internal-state probes to identify hallucinated spans, then projects judge spans onto characters for agreement-based aggregation. Training and probing choices address annotation ambiguity, span structure, and cross-token information flow.
- EDA and training split: EDA found invention and mischaracterization comprise 80-88% of errors, while 80-86% of hallucination characters are marked by only one of three annotators.
- EDA and training split: Judges receive inline span annotations, three agreement-derived probability classes, and similarity-retrieved demonstrations; fine-tuned judges use distinct training sets to reduce correlation.
- VLM judges: Five judges from four VLM backbone families provide diverse inductive biases, using retrieved few-shot demonstrations or LoRA fine-tuning on inline-annotated data.
- Internal State Probing: Probes analyze internal activations from Qwen3-VL-4B and Qwen3-VL-8B, using transformer encoders and down projection to share information across tokens.
- Internal State Probing: Multi-stage probes add boundary, sample-level, and token-level heads, gate type predictions with sample and token logits, and incorporate residual-stream differences.
- Committee Aggregation: Committee aggregation maps each judge’s highest-confidence covering span onto characters, averages confidence across judges, labels characters by summed confidence, and retains runs receiving at least two votes.
4 Results and Discussion
The committee improves hallucination detection by aggregating diverse judges, with two-vote character-level voting achieving the strongest overall challenge-set performance. Model disagreement tracks annotator disagreement, while probes offer a lower-cost but generally weaker alternative.
- 4 Results and Discussion: The committee outperforms every individual judge on Cor and IoU, except English Cor_lbl, while votes ≥2 achieves the best Cor and IoU across all four languages.Votes ≥3 slightly improves Cor_lbl in three of four languages on the seen split but degrades every challenge-set language and metric.
- 4.1 Evaluation on the Challenge Set: With votes ≥2, the committee ranks first in French, Italian, and Chinese correlation metrics, second in English, and places on the podium for every language and metric.English performs worst because longer responses reduce annotator agreement and limit achievable scores.
- 4.2 Label-wise Analysis: The committee is strongest overall in label-wise analysis, especially on correlation metrics, and lowers MAEp by averaging confidence values into a finer range.Its IoU advantage comes from suppressing spans proposed by only one judge rather than marking more text; individual judges remain competitive for invention, OCR, and miscounting.
- 4.3 VLM Judge vs. Annotator Agreement: The mean number of human annotators marking a character rises from 0.11 at zero judge votes to 2.07 at five, with correlations of 0.39 by character and 0.51 by record.This agreement pattern motivates the two-vote threshold.
- 4.4 Committee vs. Probes: Probes match individual judges on IoU but trail the committee on correlation metrics, trading accuracy for lower inference cost.The committee requires five forward passes through large fine-tuned VLMs, whereas a probe adds a lightweight head to hidden states from one pass.
5 Conclusion
The paper presents a multilingual committee for detecting hallucinated spans by combining independent VLM judges through character-level voting. It ranks first in three of four languages, while activation-probe performance depends strongly on the source model and may reflect surface linguistic features rather than mechanistic insight.
- 5 Conclusion: The committee combines fine-tuned and few-shot judges’ unsupported-span predictions through character-level voting, retaining corroborated spans and averaging confidences for correlation scoring.Its contribution is a multi-model aggregation procedure for multilingual hallucination span detection.
- 5 Conclusion: The system ranks first in three of four languages.
- 5 Conclusion: Activation probes detect hallucinations from hidden states, but their results depend strongly on the model supplying the activations.The paper leaves open whether probes provide mechanistic insight or exploit surface linguistic features.
Limitations
The committee’s performance is constrained by judge coverage, image handling, threshold transfer, coarse confidence classes, and the limited language and judge set.
- Limitations: Long, fluent hallucinations missed by multiple judges can fall below the two-vote threshold, especially in longer English responses.
- Limitations: Images are passed without resolution normalization or aspect-ratio-aware transformation, which may affect grounding.
- Limitations: The voting threshold is tuned on the labeled test split and may not generalize across models, languages, or annotation schemes.
- Limitations: Three discrete agreement classes limit how precisely the system represents uncertainty.
- Limitations: The analysis covers only four task languages and selected judges, so aggregated predictions may not capture the full complexity of human disagreement.
Data and Code Availability
The paper provides access to its committee code and fine-tuned checkpoints, while the SHEEP dataset is distributed by the task organizers.
- The VLM committee code is available in a public GitHub repository.
- The fine-tuned v3 checkpoints are available through Hugging Face.
- The SHEEP dataset is distributed by the task organizers.
A SHEEP examples: penalizing elaboration and world knowledge
The SHEEP examples illustrate how the task handles hallucinated spans and labels them using an image-grounded detector prompt. The examples include cases involving elaboration, world knowledge, and visual-text errors.
- A SHEEP examples: penalizing elaboration and world knowledge: The SHEEP examples show penalization of factually correct elaboration when it introduces hallucinated attributes.
- A SHEEP examples: penalizing elaboration and world knowledge: The detector prompt takes an image, query, and response as input and assigns hallucinated spans to invention, mischaracterization, OCR, miscounting, or other.
- A SHEEP examples: penalizing elaboration and world knowledge: The examples also show penalization of factually correct world knowledge that is difficult to verify.
C.2 Training
Probe training uses specified optimization settings and weighted losses for hallucination, boundaries, sample-level, and token-level predictions.
- Probe training uses a learning rate of 1 × 10−4, batch size 24, and weight decay 0.01.
- The BCE-loss weights are 1.0 for hallucination type, 0.1 for boundary, 0.5 for sample-level, and 0.25 for token-level hallucination loss.
- The weighting prioritizes hallucination type loss over boundary, sample-level, and token-level losses.
D.1 Fine-tuning
The system fine-tunes diverse vision-language judges with LoRA and prompts them to identify, classify, and confidence-score hallucinated spans while preserving the original response exactly.
- D.1 Fine-tuning: The committee fine-tunes Gemma 4, Mistral-Small, Qwen3-VL, and Qwen3.6 judges on distinct data subsets, alongside one few-shot Gemma 4 judge.
- D.1 Fine-tuning: Each judge receives an image, prompt, and model response, then marks unsupported response spans with <hall> tags without changing any other characters.
- D.1 Fine-tuning: Hallucinations are classified as invention, mischaracterization, OCR, miscounting, or other, with labels required to be nonempty and mutually specified.
- D.1 Fine-tuning: Each span receives prob1, prob2, or prob3, representing approximately 0.3333, 0.6667, or 1.0 likelihood of agreement among three human annotators.
- D.1 Fine-tuning: The annotation procedure checks claims thoroughly, marks doubtful spans with prob1, and selects the smallest text carrying each error.