SpanCalib-VLM: Calibrated Hallucination Span Detection in Vision-Language Models

AI summary

Overview

Research area: Computer vision and multimodal machine learning, specifically hallucination detection in Large Vision-Language Models (LVLMs), with a focus on span localization and probability calibration.

Technical level: Intermediate. The paper assumes familiarity with sequence tagging, character-level spans, IoU, and calibration metrics such as Pearson correlation, but explains its architecture and training objective clearly enough for readers from adjacent areas.

Scope: The paper describes a hybrid system that pairs a discriminative multimodal sequence tagger with a fine-tuned generative VLM, evaluates it on the multilingual SHROOM-Visions shared task, and reports calibration, localization, accuracy, and latency results across English, French, Italian, and Chinese.

What This Paper Is About

The SHROOM-Visions shared task asks systems to find the exact character-level spans in a vision-language model's response that are hallucinated, and to attach a well-calibrated confidence probability to each one. The researchers observe a trade-off: generative VLMs find hallucinated spans well but are slow and overconfident, while discriminative sequence taggers are fast and better calibrated but miss spans. SpanCalib-VLM combines both, using the generative model to propose candidate spans and the tagger to re-score them with calibrated probabilities.

Key Contributions

1. A multimodal sequence tagger that combines XLM-RoBERTa-Large with SigLIP-2 vision features through cross-attention, equipped with three parallel heads for span detection, probability calibration, and error categorization.
2. Union-Calibrated Fusion, an ensemble algorithm that treats generative VLM spans as candidate proposals and re-calibrates them using the tagger's probability distribution, with weights w1 = 0.55 (tagger) and w2 = 0.45 (VLM binary mask).
3. Competitive results on SHROOM-Visions: Pearson 0.413, IoU 0.391, and a clean-response IoU of 0.913 (described in the contributions list as 91.3% clean detection accuracy), with overall detection accuracy of 70.7%.
4. Public release of model weights, code, and evaluation artifacts at the project's GitHub repository.

Main Findings

• Best combined calibration and localization on English: SpanCalib-VLM reaches Pearson 0.413 and overall IoU 0.391 on the SHROOM-Visions English evaluation split (N = 379), compared with 0.124 Pearson / 0.210 IoU for the BLIP baseline and 0.245 Pearson / 0.298 IoU for a fine-tuned MiniCPM-V.
• Discriminative tagging beats generative decoding on calibration: The multimodal tagger alone reaches Pearson 0.369 versus 0.285 for the fine-tuned generative VLM (Qwen3.5-4B SFT), which the authors attribute to direct MSE supervision of the calibration head against human annotator consensus ratios.
• Generative VLM wins on hallucinated-span recall: System 2 achieves hallucinated-sample IoU of 0.182 versus 0.155 for the multimodal tagger, while the full ensemble improves this to 0.196.
• Each component contributes measurably: The full hybrid exceeds weighted averaging by +0.032 Pearson (0.381) and intersection by +0.061 Pearson (0.352). Removing the MSE calibration loss causes the largest single drop, −0.095 Pearson (0.318). Removing the SigLIP-2 vision tower lowers IoU only slightly, from 0.391 to 0.388 (+0.003 attributed to the vision tower).
• Stable threshold behavior: Ensemble IoU stays at 0.391 for all thresholds in τ ∈ [0.25, 0.45], which the authors attribute to a bimodal fusion score distribution; at τ = 0.50 IoU falls to 0.317 and detection accuracy drops to 43.8%.
• Substantial speed advantage: The tagger runs at 0.19 s/sample (5.25 samples/s) versus 0.85 s/sample (1.18 samples/s) for standard Qwen-4B inference and 23.10 s/sample (0.04 samples/s) for chain-of-thought mode, making it 4.5 times faster than standard inference and 120 times faster than chain-of-thought.
• Strong cross-lingual results on the hidden test set: Against the task-provided baseline, SpanCalib-VLM scores 0.2418 Cor+Lbl / 0.3212 Cor / 0.2549 IoU on English (baseline 0.1549 on all three), 0.2885 / 0.3600 / 0.2994 on French (baseline 0.1679), 0.2918 / 0.3832 / 0.3109 on Italian (baseline 0.1635), and 0.3347 / 0.3752 / 0.3425 on Chinese (baseline 0.2554).
• Chinese performs best, with a tokenization explanation: The paper reports 0.437 overall IoU and 0.958 clean IoU on Chinese validation, outperforming base Qwen-3.5-4B by +0.169 IoU, attributing this to ideographic characters providing dense token representations that align directly with character boundaries.
• Large gains on non-English hallucinated spans: Hallucinated IoU improves by +0.182 on French (0.020 to 0.202) and +0.180 on Italian (0.027 to 0.207) over base models, which the authors say suffer from severe under-detection in those languages.
• Overfitting after epoch 2: Validation Pearson peaks at epoch 2 (0.369) while training loss continues to fall from 1.068 at epoch 1 to 0.613 at epoch 5 and validation loss rises from 0.923 at epoch 2 to 1.042 at epoch 5; the best-checkpoint strategy preserves the epoch-2 state.

Methodology in Plain English

The system takes three inputs: an image, the question prompt, and the vision-language model's text response. It then runs two independent subsystems.

System 1 is a discriminative sequence tagger. The prompt and response are concatenated into a text string and encoded by XLM-RoBERTa-Large, which outputs hidden states over up to 512 tokens. The image is encoded separately by SigLIP into 196 patch embeddings, which are projected into the text model's hidden space and merged into the text representations through a cross-attention layer. On top of this fused representation, three small heads run in parallel on every token: one predicts whether the token is part of a hallucination, one predicts a continuous confidence score, and one predicts which of five error categories applies. Training uses a combined loss of binary cross-entropy for span detection, mean squared error for calibration, and cross-entropy for categorization, with loss weights λ1 = 1.0 and λ2 = 0.5. Crucially, the MSE term supervises the confidence score directly against the human annotator agreement values rather than relying on the model's own generation probabilities.

System 2 is a generative VLM, Qwen3.5-4B, fine-tuned with LoRA (r = 16, α = 16, all-linear, BF16, 8-bit AdamW, learning rate 2.0 × 10⁻⁴, 3 epochs, max sequence length 2048 tokens) to output structured JSON spans. Its spans are converted into a binary character-level mask.

Fusion then works in three steps: collect candidate span boundaries from System 2, map System 1's token probabilities onto a character-level array by aligning subtoken offsets to characters, and combine the two into a single score P_ens(k) = 0.55 · P_SC(k) + 0.45 · M_VLM(k). The weights were chosen by grid search on the validation set, giving slightly more influence to the tagger because of its better calibration baseline. The tagger is trained for 5 epochs on NVIDIA A40 GPUs with AdamW (learning rate 1.5 × 10⁻⁵), 10% linear warmup, batch size 16, and 512-token maximum sequence length.

Why This Matters

Impact on research: The paper reframes hallucination detection as a joint localization-and-uncertainty problem rather than a pure detection problem, and shows that a discriminative head trained with an explicit regression loss produces better-calibrated probabilities than autoregressive decoding probabilities. The ablation results give a concrete measurement of how much each ingredient (fusion strategy, calibration loss, vision tower) contributes, and the dual-system framing connects to the broader System 1 / System 2 hybrid architecture literature.

Real-world applications:

• Medical or scientific image description systems, where a confidently wrong statement about an image is more dangerous than an uncertain one.
• Accessibility tools that narrate images to blind or low-vision users, where unflagged invented objects could mislead.
• Content moderation and fact-checking pipelines for AI-generated image captions at scale, where per-sample latency matters.
• Assistive or autonomous systems that read scenes and must know when to defer to a human instead of acting.

Industry relevance: The latency comparison gives a direct deployment argument: 5.25 samples/s versus 1.18 samples/s for standard generative inference and 0.04 samples/s for chain-of-thought. A calibrated confidence signal also enables selective prediction, letting a production system route only uncertain cases to expensive human review or to a larger model.

Future Directions

• Reducing false positives on correct but elaborate language, such as idioms and rare accurate visual details, which the tagger flags because it relies on surface-level distributional patterns rather than visual grounding.
• Extending beyond the 512-token input limit, since hallucinated content past the truncation boundary is currently undetectable.
• Improving character-level localization of hallucinated spans, where the paper reports hallucinated-span IoU of only 0.196 despite strong overall calibration.
• Developing detection methods that can recover hallucinations missed by both subsystems, since fusion cannot create a span that neither the tagger nor the generative VLM proposes, and revisiting threshold and hyperparameter selection beyond the English validation split used here.

Target Audience

Researchers and practitioners working on hallucination detection, multimodal model reliability, and uncertainty calibration, particularly those entering the SHROOM-Visions or related shared tasks. It is also useful for engineers building production systems where a vision-language model needs to report calibrated confidence and fast throughput, and for readers interested in hybrid discriminative-generative architectures.

Authors’ abstract

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 hallucinated text spans but suffer from overconfidence and high inference latency. Discriminative sequence taggers offer deterministic speed and superior calibration but exhibit conservative span recall. We present SpanCalib-VLM, a hybrid dual-system for the SHROOM-Visions Shared Task that combines a multimodal sequence tagger, consisting of XLM-RoBERTa-Large fused with a SigLIP vision encoder via cross-attention, with our fine-tuned generative VLM (Qwen3.5-4B-SHROOM-SFT). Through a Union-Calibrated Fusion strategy, candidate spans from the generative model are re-scored with calibrated probabilities from the sequence tagger. On the SHROOM-Visions English evaluation split, our ensemble achieves a Pearson calibration correlation of 0.41 and an overall IoU of 0.39, with a clean-response IoU of 0.91} and overall detection accuracy of 70.7%. We make our model weights and code publicly available.