PaperlayerView exact revision on arXiv

Source-linked AI summary

Seeing is Believing: Mitigating Hallucination in Large Vision-Language Models via CLIP-Guided Decoding

Ailin Deng, Zhirui Chen, Bryan Hooi

arXiv:2402.15300v2cs.CVcs.AIcs.CLcs.LGcs.MM

TL;DR

Object hallucinations in LVLMs threaten reliability, while CLIP’s ability to detect them in open-ended generation remains underexplored. The paper introduces training-free CLIP-Guided Decoding, which uses image-text similarity to guide decoding; experiments report lower hallucination metrics while maintaining similar response lengths and generation quality.

  • Problem

    Object hallucinations threaten LVLM reliability and practical utility, while CLIP’s ability to identify hallucinations in open-ended generated text remains underexplored.

    Source: §1 · p. 1

  • Method

    CLIP-Guided Decoding uses CLIP as an external guide to compare generated text with the image during decoding, without additional training.

    Source: §1 · p. 2

  • Results

    CGD achieves lower CHAIR scores than other baselines while maintaining similar response lengths, and performs on par with other approaches on COCO caption-quality metrics.

    Source: §5.2 · p. 8

  • Takeaways & Limitations

    CLIP scores can guide sentence-level decoding, including when reusing CLIP models already used in an LVLM.

    Source: §1 · p. 2

Abstract

from arXiv · show

Large Vision-Language Models (LVLMs) are susceptible to object hallucinations, an issue in which their generated text contains non-existent objects, greatly limiting their reliability and practicality. Current approaches often rely on the model's token likelihoods or other internal information, instruction tuning on additional datasets, or incorporating complex external tools. We first perform empirical analysis on sentence-level LVLM hallucination, finding that CLIP similarity to the image acts as a stronger and more robust indicator of hallucination compared to token likelihoods. Motivated by this, we introduce our CLIP-Guided Decoding (CGD) approach, a straightforward but effective training-free approach to reduce object hallucination at decoding time. CGD uses CLIP to guide the model's decoding process by enhancing visual grounding of generated text with the image. Experiments demonstrate that CGD effectively mitigates object hallucination across multiple LVLM families while preserving the utility of text generation. Codes are available at https://github.com/d-ailin/CLIP-Guided-Decoding.

1 Introduction

Object hallucinations threaten LVLM reliability and practical utility. The paper introduces training-free CLIP-Guided Decoding, motivated by CLIP’s stronger sentence-level hallucination signal, and reports reduced hallucination while preserving text-generation utility.

1 Introduction

  • Object hallucinations—descriptions containing non-existent objects—undermine LVLM reliability and practical utility and can mislead users.The introduction highlights safety-critical settings such as robotics and medical image analysis.Source: §1 · p. 1
  • Existing mitigation approaches use instruction tuning, model-internal signals, or external tools, but can entail annotation costs, unreliable signals, or engineering burden.The introduction notes that relying only on internal information may be insufficient and potentially unreliable.Source: §1 · p. 2
  • CGD directly compares generated text with the image, using CLIP as an external guide during decoding without additional training.The approach aims to improve correspondence between text and image.Source: §1 · p. 2
  • Whether CLIP can detect hallucinations in open-ended LVLM-generated text, including from models with CLIP vision encoders, remains underexplored.Prior CLIP work had focused mainly on image-text pairwise comparisons.Source: §1 · p. 2
  • CLIP scores are stronger and more robust hallucination indicators than token likelihoods, especially for later caption sentences, where hallucinations are more likely.The later-sentence pattern is reported across multiple LVLMs.Source: §1 · p. 2
  • CGD mitigates hallucination while preserving overall generation quality and utility.The introduction reports empirical effectiveness alongside preserved generation quality.Source: §1 · p. 2

2 Related Works

LVLMs remain prone to hallucination, especially in open-ended generation. Prior work explores mitigation through tuning and model-internal or external signals, while the paper positions CLIP-guided decoding as a distinct alternative.

2.1 Large Vision-Language Models

  • LVLMs remain susceptible to hallucination, particularly in open-ended generation, despite advances in vision-language task capabilities.The paper describes hallucination mitigation in recent LVLMs as comparatively unexplored.Source: §2.1 · p. 3, §2.2 · p. 3

2.2 Hallucination in VLMs

  • LVLM mitigation work includes instruction tuning and internal likelihood or hidden-state signals, while CLIP’s role in detecting open-ended hallucinations remains underexplored.The passage notes potential unreliability of internal states and engineering complexity in external-model approaches.Source: §2.2 · p. 3

2.3 Decoding Strategies for Mitigating Hallucinations

  • Prior decoding strategies use internal states or contrast outputs; this paper instead uses CLIP to guide generation toward visually accurate content.The cited approaches include LVLM methods based on attention patterns or varied image inputs.Source: §2.3 · p. 3

3 Preliminaries

The section defines sentence-level measures and compares sentence likelihood with CLIPScore for detecting hallucinations across LVLMs and datasets. Later sentences are more prone to hallucination, while CLIPScore is a stronger detector and more stable across datasets.

3 Preliminaries

  • Responses are modeled as autoregressively generated sequences of sentences conditioned on an image and text input, with sentence likelihood measuring each sentence given prior sentences.Source: §3 · pp. 3–4
  • CLIPScore measures image-text association as cosine similarity between normalized image and sentence features.Source: §3 · p. 4

3.1 Hallucination Analysis

  • The analysis defines sentence-level hallucination ratio R(i), first-time hallucination ratio Rfirst(i), and AUROC for detection performance.Source: §3.1 · p. 5
  • Hallucinations increase in later sentences, and the pattern remains after removing error propagation, suggesting it may partly reflect diminishing attention to visual inputs.Source: §3.1 · p. 5
  • Sentence likelihood has generally weak hallucination-detection performance, which diminishes for later sentences.Source: §3.1 · p. 5
  • Sentence likelihood scores are consistently higher on COCO than on out-of-domain NoCaps, raising concerns about relying on them alone for real-world generalization.Source: §3.1 · p. 6
  • CLIPScore distinguishes hallucinated from correct sentences across models and datasets, performs well across sentence positions, and is more stable across COCO and NoCaps than sentence likelihood.Source: Figure 2 · p. 4, Figure 3 · p. 5, §3.1 · p. 6

4 CLIP-Guided Decoding

CLIP-Guided Decoding (CGD) uses CLIP-based reliability scoring to prioritize visually grounded responses while retaining the LVLM’s sentence-sampling process. It applies this guidance during sentence-level decoding, scoring complete sentences rather than partial token sequences.

Guided Sentence Generation

  • Sentence-level decoding lets CGD apply CLIP scoring to complete sentences instead of incomplete words or phrases.Source: §4 · p. 6

Reliability Scoring

  • CGD combines normalized LVLM likelihood with CLIP guidance over generated sentences to score candidate responses for visual grounding.The hyperparameter α controls the mix: α = 0 uses only likelihood, while α = 1 indicates preference for higher CLIPScore.Source: §4 · p. 6

Guided Sentence Generation

  • At each step, CGD samples M possible next sentences for each candidate, then retains the top N candidates by sentence-level hallucination score.Source: §4 · p. 7
  • Decoding repeats until candidates reach EOF or the maximum output length, and the highest-scoring candidate is returned.Source: §4 · p. 7

5 Experiments

Across captioning datasets, CGD lowers CHAIR scores while keeping response lengths similar to baselines, and COCO generation quality remains on par with other approaches. Experiments also examine open-ended VQA, scoring components, hyperparameters, and the guiding model.

5.1 Experimental Setup

  • Hallucination evaluation used 500 COCO and near- and out-of-domain NoCaps samples per set and run; open-ended VQA used MM-Vet.The image prompt was “Describe this image in detail.”Source: §5.1 · p. 7
  • Evaluation paired CHAIR metrics with generation-quality measures and compared CGD with six decoding baselines, using SigLIP as its default guiding model.Baselines included greedy, Nucleus and TopK sampling, DoLa, VCD, and OPERA; results were averaged over three runs.Source: §5.1 · p. 7

5.2 Experimental Results

  • CGD consistently achieves lower CHAIR scores than other baselines while maintaining similar response lengths.The evaluation covers COCO and near- and out-of-domain NoCaps.Source: §5.2 · p. 8, Table 1 · p. 8
  • On COCO, CGD performs on par with other approaches across generation-quality metrics.The evaluation includes CIDEr and SPICE.Source: §5.2 · p. 8, Table 2 · p. 8
  • On MM-Vet, CGD surpasses greedy decoding on most tasks, including recognition and OCR, but underperforms on spatial awareness.The authors attribute the spatial-awareness result likely to CLIP models’ limitations in relational understanding.Source: Figure 4 · p. 8, §5.2 · p. 9
  • Without CLIP guidance, performance is close to greedy decoding, while sentence likelihood alone has relatively limited effectiveness in reducing hallucination.The ablation controls α, with α = 1 removing sentence likelihood and α = 0 removing CLIP guidance.Source: §5.2 · p. 9, Figure 5 · p. 9
  • CGD outperforms the baseline with reduced candidate and sampling budgets, while larger N and M generally improve performance at increased inference cost.Its inference efficiency is reported as on par with VCD and better than OPERA, alongside better hallucination mitigation.Source: §5.2 · p. 9, Table 3 · p. 9
  • Using LLaVA-1.5’s own CLIP ViT/L-14 vision encoder as the guide still outperforms the baseline.The authors suggest this may indicate that vision-language alignment fine-tuning can compromise original vision capabilities.Source: §5.2 · p. 9

6 Conclusion

The conclusion reports severe object hallucination in later sentences and identifies CLIPScore as a stronger, more robust indicator than token likelihood. It presents training-free CLIP guidance during decoding as mitigating hallucination while preserving generation quality.

6 Conclusion

  • Table 4 is identified as a sensitivity analysis of the CLIP-guidance model.Source: Table 4 · p. 9
  • The study finds severe hallucination in later sentences and CLIPScore a stronger, more robust indicator than token likelihood.Source: §6 · p. 9
  • The training-free approach effectively mitigates hallucination while preserving generation quality.CLIP is integrated as external guidance during decoding.Source: §6 · pp. 9–10

A Analysis Detail

The analysis compares sentence-level hallucination measures across sentence positions and finds that CLIPScore generally detects hallucinations better than sentence likelihood, especially later in captions. The algorithm takes an image and prompt and selects a response from candidate continuations.

A Analysis Detail

  • Figures 6 and 7 present hallucination and first-time hallucination ratios by sentence index on COCO and out-of-domain NoCaps.Source: Figure 6 · p. 15, Figure 7 · p. 15
  • CLIPScore generally outperforms sentence likelihood for hallucination detection, especially in later sentences, across three LVLMs and COCO and NoCap.Figure 8 reports AUROC, with 50% representing random guessing.Source: Figure 8 · p. 16

B Algorithm

  • The algorithm takes an image-text input, LVLM, CLIP model, candidate limit, sample size, and weight, then returns a response.It retains the top N candidates and returns the candidate maximizing F(c).Source: B Algorithm · p. 16

C Experimental Setup

Experiments use InstructBLIP, mPLUG-Owl2, and LLaVA-1.5, which differ in how they connect vision and text features. The models use specified image representations and a shared default maximum of 500 new tokens.

C Experimental Setup

  • The evaluation uses InstructBLIP, mPLUG-Owl2, and LLaVA-1.5 as representative LVLMs.Source: §C.1 · p. 16
  • InstructBLIP uses a Q-former with 32 image tokens, while mPLUG-Owl2 and LLaVA-1.5 use linear projections with 256 or 576 image tokens.Source: §C.1 · p. 16
  • All methods use a default maximum of 500 new tokens, and the models’ image encoders are pretrained with image-text contrastive objectives such as CLIP or EVA-CLIP.Source: §C.1 · p. 17

C.2 Datasets

The experimental setup draws on COCO, NoCaps, and MM-Vet and evaluates hallucination, generation quality, and open-ended answers with distinct metrics. It compares common decoding strategies with three methods proposed to mitigate hallucinations.

C.2 Datasets

  • COCO experiments use samples from the Karpathy split; NoCaps validation images test models on categories less seen or unseen in COCO training.NoCaps validation is used because test-set ground-truth captions are unavailable.Source: §C.2 · p. 17
  • MM-Vet evaluates six vision-language capabilities and their 16 integrations using 200 images and 218 questions.Source: §C.2 · p. 17

C.3 Metrics.

  • CHAIR measures the ratio of mentioned objects absent from ground-truth labels at sentence and instance levels, using 80 COCO classes and 90 coarse NoCaps categories.The NoCaps categories are mapped from 600 fine-grained classes.Source: §C.3 · p. 17
  • Generation quality is assessed with response length, object coverage, BLEU, METEOR, ROUGE-L, CIDEr, SPICE, and CLIPScore.Coverage is the proportion of correctly identified objects relative to gold objects; CIDEr uses weighted n-gram similarity and SPICE parses semantic propositions.Source: §C.3 · pp. 17–18
  • CLIPScore measures textual-image association using cosine similarity of normalized CLIP features, while MM-Vet open-ended answers receive average LLM soft-grading scores from 0 to 1.Source: §C.3 · p. 18

C.4 Baselines.

  • The comparisons include greedy, Nucleus, and TopK decoding, plus DoLa, VCD, and OPERA.All methods use temperature 0.2; Nucleus uses top-p 0.7 and TopK uses top-k 5.Source: §C.4 · pp. 18–19

C.5 Implementation Detail.

The implementation uses SigLIP by default to guide decoding, with TopK sampling and fixed algorithm settings; ablation and sensitivity studies focus on LLaVA-1.5.

C.5 Implementation Detail.

  • SigLIP (Zhai et al., 2023) is the default CLIP-guiding model, with variant CLIP models examined in the sensitivity study.Source: §C.5 · p. 19
  • Generation uses TopK sampling under the same setting as the TopK sampling baseline, and reported performance averages three runs.Source: §C.5 · p. 19
  • Default algorithm hyperparameters are N = 3, M = 3, and α = 0.99; ablation and sensitivity studies focus on LLaVA-1.5.Source: §C.5 · p. 19

D.1 Ablation Study

The section reports a scoring-component ablation and illustrative cases of hallucination mitigation on COCO and VQA responses on MM-Vet.

D.1 Ablation Study

D.2 Case Study

  • COCO examples compare original LLaVA-1.5 descriptions with the method's descriptions for the same image, marking hallucinated text in red.Source: §D.2 · p. 19, Table 6 · p. 20
  • MM-Vet examples pair each image and question with original and method responses, ground-truth answers, response accuracy scores, and required capabilities.Source: §D.2 · p. 19, Table 7 · p. 21
Loading 2402.15300v2…