In 2026, the SHROOM-Visions shared task was launched at the UncertaiNLP Workshop co‑located with EMNLP to address hallucinations in large vision‑language models. The task builds on the SHEEP dataset and asks participants to detect and classify fine‑grained hallucination spans in image‑conditioned text generation across four languages (Chinese, English, French, Italian) using a five‑class taxonomy. The competition attracted 27 teams and over 600 system submissions, with top systems achieving character‑level, label‑conditioned, and IoU scores of 0.58, 0.46, and 0.51 respectively, surpassing baselines by 30‑40 points.
By Ra\'ul V\'azquez, Aman Sinha, Chuyuan Li, Claudio Savelli, Eduardo Cal\`o, Emilio Raimond, Stella Frank, Hengyu Luo, Flavio Giobergia, Vincent Segonne, Lorenzo Vaiani, J\"org Tiedemann, Timothee Mickus
In 2026, we held the fourth iteration of the SHROOM Shared Task series: SHROOM-Visions (\textbf{S}hared-task on \textbf{H}allucinations and \textbf{R}elated \textbf{O}bservable \textbf{O}vergeneration...
SpanCalib-VLM is a hybrid system for detecting hallucinated text spans in Vision‑Language Models. It combines a multimodal sequence tagger (XLM‑RoBERTa‑Large + SigLIP) with a fine‑tuned generative VLM (Qwen3.5‑4B‑SHROOM‑SFT) and uses a Union‑Calibrated Fusion strategy to re‑score candidate spans. On the SHROOM‑Visions English evaluation split, the ensemble achieves a Pearson calibration correlation of 0.41, an overall IoU of 0.39, a clean‑response IoU of 0.91, and a detection accuracy of 70.7%.
By Amanuel Gizachew Abebe, Yasmin Moslem
Detecting hallucinations in Large Vision-Language Models (LVLMs) requires both accurate span localization and well-calibrated confidence scores. Fine-tuned generative VLMs excel at identifying halluci...
ReWEIGH the Evidence is a training‑free decoding technique that calibrates token‑level ordinal visual evidence to reduce hallucinations in large vision‑language models. It aggregates vocabulary ranks across visual positions, compares candidates to a token‑specific reference derived from unlabeled images, and applies a bounded penalty only when evidence falls below this reference. Experiments on four 7B backbones show up to a 21.3% reduction in hallucinated object mentions while largely preserving or improving descriptive and general performance, with minimal added latency.
By Jihae Jeong, Junha Choi, Hwanjo Yu
HALDETECT is a system developed for the English hallucination-detection track of ImageEval 2026, where the task is to identify the single visually grounded statement among three culturally plausible options. The approach treats the problem as a contrastive decision, outputs the answer before an explanation, and bases reasoning on colour/texture, shape/form, and context. The best model fine‑tunes Qwen2.5‑VL‑7B‑Instruct with 4‑bit QLoRA, freezes the vision encoder, and achieves a Contrastive Instability score of 0.035 on the test set, placing third among eight teams.
By Syed Mohaiminul Hoque, Md Sakhawat Hossain
The paper examines six inference-time hallucination mitigation methods applied to three large vision-language models across four benchmarks, including MMStar. It finds that reducing hallucination rates often comes at the cost of lower informativeness—such as decreased object recall, visual coverage, and response detail—and that gains on hallucination benchmarks do not consistently translate to improved performance on fine-grained perception and reasoning tasks. The authors argue that current evaluation protocols may overstate progress by favoring conservative generation, and propose that hallucination mitigation should be assessed as a trade-off among faithfulness, informativeness, and overall capability.
By Mehrdad Fazli, Sina Mansouri, Mohit Marvania, Ziwei Zhu
The paper describes a system for the SHROOM-Visions 2026 shared task on character-level VLM hallucination detection. It combines a small 4‑B‑parameter VLM fine‑tuned as a per‑token classifier that uses a two‑token feature from its hidden states with a large ~400‑B zero‑shot VLM judge at prediction time, both leveraging OCR of visible in‑image text. Using synthetic hallucination data from the large model for ensemble diversity and validation‑based selection of feature layer, training data, and OCR grounding, the entry achieved competitive results across multiple languages.
By Eli Schwartz
The article surveys hallucination issues in Large Vision‑Language Models (LVLMs), a type of multimodal foundation model that blends visual data with large language models. It categorizes hallucination causes into model architecture and data quality, presents a taxonomy of mitigation strategies, and critically evaluates existing evaluation benchmarks from both discriminative and generative viewpoints. The survey also outlines open challenges and future research directions to improve LVLM reliability and trustworthiness.
By Yinghao Guo, Wei Lan, Wenyi Chen, Qingfeng Chen, Shichao Zhang, Shirui Pan, Huiyu Zhou, Yi Pan
arXiv:2606. 06959v1 Announce Type: cross Abstract: Hallucination detection is essential for the reliable deployment of large language models (LLMs).
By Xinyi Li, Zhen Fang, Yongxin Deng, Jinyuan Luo, Hongnan Ma, Changdae Oh, Zijing Shi, Shanshan Ye, Hanchen Wang, Shu-Lin Chen, Yadan Luo, Mengyue Yang, Sean Du, Sharon Li, Ling Chen
The paper investigates why vision‑language models like LLaVA‑1.5‑7B hallucinate objects in captions and proposes a targeted fix. By ranking attention heads whose image attention drops around hallucinated words, the authors identify 32 key heads and apply a head‑sliced LoRA adapter plus an inference‑time grounding controller. On COCO images, this combined method reduces hallucinated captions from 37% to 23% and hallucinated object mentions from 15.6% to 9.6%, while also lowering object recall.
By Armaan Sandhu, Abhilasha Senapati, Hima Kammachi
The paper reports a submission to the SHROOM-Visions shared task, aiming to detect and classify hallucinated character spans in vision‑language model outputs across four languages. The authors use multiple fine‑tuned vision‑language models as independent annotators, combine their predictions via character‑level majority voting, and also investigate activation probes. Their method achieved first place in three of the four languages and consistently ranked on the podium for all languages and metrics, with analysis showing that model disagreement mirrors human annotator disagreement.
By Toqeer Ehsan, Nico Penttil\"a, Richard Schmidt, Arash Hajikhani, Victoria Palacin