HalluLens: Evaluation of Free-form Hallucinations

• HalluLens is a comprehensive evaluation framework for free-form hallucinations in LVLMs, identifying ungrounded outputs in open-ended visual descriptions.
• It employs a multi-phase protocol including free-form prompting, automated object extraction, ensemble filtering, and precise metric computation.
• The approach uses rigorous metrics—precision, recall, and F0.5—to assess hallucination rates, showing its superiority over structured benchmarks.

HalluLens is a comprehensive object-based evaluation framework and metric suite for quantifying free-form (“Type I”) hallucinations in large vision-LLMs (LVLMs), particularly those emerging in unconstrained textual generations grounded on visual input. Unlike prior benchmarks that focus on structured prompts or targeted yes/no/object presence queries, HalluLens-type protocols differentiate hallucination phenomena arising in genuinely open-ended LVLM responses and introduce robust automated assessment methodologies fit for large-scale empirical evaluation.

1. Definition and Motivation

Free-form hallucinations in LVLMs are semantically plausible outputs that contradict the semantically or visually grounded content of an input image. Type I hallucinations appear when a model is prompted in an unconstrained, generative format (e.g., “Describe this image in detail”), resulting in invented or misplaced objects not grounded in the image. HalluLens, as instantiated by THRONE (“Text‐from‐image Hallucination Recognition with Object‐probes for open‐eNded Evaluation”), targets deficiencies in prior benchmarks where improvements on Type II hallucinations (structured queries with closed outputs) do not tightly correspond to gains in free-form settings (Kaul et al., 2024). This motivates the design of new metrics and procedures specifically for open-ended hallucination detection.

2. Benchmark Protocol and Methodological Principles

HalluLens-type benchmarks operate in four principal phases:

1. Free-form Prompting: The LVLM receives a neutral, open-ended description prompt (e.g., “Describe this image in detail”), generating an unconstrained response $R_i$ per image $i$.
2. Automated Object Extraction: For each target object class $c$ from a predefined set $C$, the system generates yes/no questions using external public LLMs (e.g., ensemble of FLAN-T5 variants) with $R_i$ as context, phrased as “Is there a [c] in this image?”.
3. Ensemble Filtering and Voting: A set of $N$ LMs and $M$ prompt variants produce $k=N \cdot M$ binary answers per $(i, c)$. Unanimous voting sets the predicted existence $\hat{Y}_{i,c}$; ambiguous cases are ignored in metric aggregation.
4. Precision-oriented Evaluation: Ground-truth object matrices $i$0 are constructed from image-level annotations (e.g., COCO, Objects365). Evaluation proceeds via strict object-wise precision, recall, and $i$1-$i$2 metrics that emphasize penalizing hallucinations.

This protocol automatically and reproducibly quantifies the prevalence of object hallucinations in free-form LVLM generations, supporting scalable and impartial comparison (Kaul et al., 2024).

3. Formal Metric Suite and Core Definitions

Key evaluation metrics in HalluLens frameworks reflect object-level existence verification, prioritizing the avoidance of hallucinations. Let $i$3, $i$4, $i$5 describe true positives, false positives, and false negatives across all image/class pairs after discarding “ignore” cases.

• Overall Precision and Recall

$i$6

• Class-wise Scores (averaged over object classes)

$i$7

• Balanced F-score (principal: $i$8 to emphasize precision over recall)

$i$9

• Object-level Hallucination Rate: $c$0, capturing the fraction of predicted objects that are false.

The $c$1 score serves as the primary performance indicator, as it strongly penalizes ungrounded object invention (Kaul et al., 2024).

4. Why Object-based Free-form Evaluation is Essential

Empirical findings substantiate that gains in conventional structured benchmarks (Type II hallucinations), such as POPE-C (exhaustive yes/no object verification), do not correlate with improvements in free-form hallucination metrics. For instance, cross-benchmark analysis on recent LVLMs establishes only weak Spearman rank correlation ($c$20.4 between THRONE and POPE-Complete, $c$30.2 for prior POPE), indicating that mitigation strategies successful for structured mistakes cannot be assumed to suppress open-ended hallucinatory output. The error structure of open-form hallucinations is thus distinct, and HalluLens-type protocols are indispensable for robust LVLM evaluation (Kaul et al., 2024).

5. Dataset Construction and Experimental Design

Automatic object-based hallucination evaluation typically leverages richly annotated image corpora:

• COCO 2017 validation: 5,000 images, 80 classes, exhaustive multi-object annotation.
• Objects365: Subsets of $c$45,110 images, 365 classes, chosen to be out-of-training domain for most LVLMs.

Each model is tested in batch using one output per image. The ensemble AQA process for all classes per image produces hundreds of thousands of LM calls, though the approach is tractable at scale due to automation and parallelism (Kaul et al., 2024). Model line-up includes Adapter-v2, InstructBLIP, Otter-Image, MiniGPT-4, mPLUG-Owl, LRV-Instruction, LLaVA, and other widely used open LVLMs.

6. Quantitative Results and Interpretation

Representative THRONE results (COCO; $c$5):

Model	$c$6
Adapter-v2	68.7
InstructBLIP	76.1
MiniGPT-4	75.5
LLaVA-v1.3	76.5
LLaVA-v1.5	66.8
LLaVA-Mistral	77.5

Despite state-of-the-art architectures, even the best models hallucinate $c$720% of objects in free-form outputs ($c$8). The metrics generalize to other domains (Spearman $c$9 across COCO/Objects365). Auditing indicates a qualitative error rate for THRONE ($C$04.3%) is markedly lower than that of alternative benchmarks such as CHAIR (~8.8%), demonstrating its accuracy (Kaul et al., 2024, Fang et al., 2024).

7. Hallucination Mitigation, Baselines, and Future Work

A lightweight and effective mitigation strategy is the object enumeration task baseline, wherein the LVLM is prompted to explicitly list present/absent objects and their locations. For instance, LLaVA-v1.5, when trained with COCO-based enumeration, increased $C$1 by up to +18.1 percentage points. Inference-time enumeration and negative class sampling both contribute to improved robustness for both Type I and II hallucinations (Kaul et al., 2024).

Recent advances such as uncertainty-guided dropout decoding further reduce object hallucination rates by masking visually “surprising” tokens at inference time, yielding additional relative improvements in precision and overall robustness on THRONE, CHAIR, and MMBench without sacrificing recall (Fang et al., 2024).

Future directions proposed include extension beyond object existence (attributes, relations, actions), semi-supervised protocols for images without exhaustive ground-truth, integration of model-internal self-consistency checks, and explicit bias/ethics auditing.

---

HalluLens-style methodologies, epitomized by THRONE, have become foundational tools for the evaluation and mitigation of free-form hallucination errors in LVLMs. They provide reproducible, scalable, and precise metrics essential for decoding model reliability in unconstrained, real-world deployment (Kaul et al., 2024, Fang et al., 2024).