PaperlayerView exact revision on arXiv

Source-linked AI summary

Hallucinatory Image Tokens: A Training-free EAZY Approach on Detecting and Mitigating Object Hallucinations in LVLMs

Liwei Che, Tony Qingze Liu, Jing Jia, Weiyi Qin, Ruixiang Tang, Vladimir Pavlovic

arXiv:2503.07772v2cs.CVcs.LG

TL;DR

Object hallucination compromises LVLM reliability, while how visual input contributes to it remains insufficiently understood. EAZY detects and mitigates hallucinations by zeroing candidate hallucinatory image tokens. On LLaVA-1.5 detection, it improves precision of HOs over the best baseline and outperforms training-free methods across most evaluation metrics.

  • Problem

    Prior work has largely focused on language-side biases, leaving how visual input contributes to object hallucination insufficiently understood despite its costs to LVLM reliability and real-world applicability.

    Source: §1 · p. 1

  • Method

    EAZY is a training-free method that detects HOs by removing object-specific top-K candidate HITs, then rectifies responses by zeroing aggregated candidates.

    Source: §1 · p. 2, §7 · p. 9

  • Results

    15.09% improvement in precision of HOs over the best baseline on LLaVA-1.5 detection; EAZY outperforms training-free methods across most evaluation metrics.

    Source: §1 · p. 3, §5.2.2 · p. 7

  • Takeaways & Limitations

    Zeroing HITs can remove hallucinated objects while minimally affecting real objects, supporting object-hallucination detection and mitigation.

    Source: §1 · p. 2, §7 · p. 9

    Logit Lens alone cannot capture all HITs, limiting current interpretability-based identification and motivating alternate tools to refine the approach.

    Source: §3.2 · p. 4

Abstract

from arXiv · show

Despite their remarkable potential, Large Vision-Language Models (LVLMs) still face challenges with object hallucination, a problem where their generated outputs mistakenly incorporate objects that do not actually exist. Although most works focus on addressing this issue within the language-model backbone, our work shifts the focus to the image input source, investigating how specific image tokens contribute to hallucinations. Our analysis reveals a striking finding: a small subset of image tokens with high attention scores are the primary drivers of object hallucination. By removing these hallucinatory image tokens (only 1.5% of all image tokens), the issue can be effectively mitigated. This finding holds consistently across different models and datasets. Building on this insight, we introduce EAZY, a novel, training-free method that automatically identifies and Eliminates hAllucinations by Zeroing out hallucinatorY image tokens. We utilize EAZY for unsupervised object hallucination detection, achieving 15% improvement compared to previous methods. Additionally, EAZY demonstrates remarkable effectiveness in mitigating hallucinations while preserving model utility and seamlessly adapting to various LVLM architectures.

1. Introduction

Object hallucination undermines LVLM reliability, and prior work has largely focused on language-side causes rather than how visual input contributes. The paper links hallucinations to highly attended image tokens and introduces EAZY to detect and mitigate them by zeroing candidate tokens.

1. Introduction

  • Object hallucinations compromise LVLM reliability and limit broader real-world applicability, motivating work to improve their accuracy and trustworthiness.Source: §1 · p. 1
  • Earlier work has emphasized language-side biases and interventions, leaving how visual input influences hallucination as an open question.The paper turns to visual input as a source of object hallucinations.Source: §1 · p. 1
  • The Figure 1 case shows that removing three image tokens eliminates hallucinated apples and oranges while revealing the real kiwis.Source: Figure 1 · p. 2
  • Hallucinatory objects can attend to visually related regions without a true image anchor, while zeroing top-attended candidate tokens can make them disappear.For example, a nonexistent baseball glove may focus on the baseball; removing two or three HITs can eliminate a hallucinated object while leaving real objects largely unaffected.Source: §1 · p. 2
  • EAZY achieves a detection rate up to 80% and surpasses state-of-the-art training-free methods.It identifies hallucinated objects, then zeroes aggregated candidate HITs from those objects to rectify the response.Source: §1 · p. 2

2. Connecting Generated Object Token with Its Image Anchor via Attention

The paper uses token-to-token attention to connect generated object tokens with image regions, then evaluates where those connections are strongest. Its analysis points to middle layers, especially layer 15, where attention is more spatially distributed than in typical layers.

2.1. Evaluating Object Localization Performance among Model Layers.

  • Attention patterns in many layers are noisy, but middle-to-late layers reveal clearer links between generated object tokens and visual grounding regions.The finding motivates using layerwise attention to locate where the model extracts object information.Source: §2.1 · p. 3
  • OL-VAR measures attention on an object's bounding-box region relative to all image-token attention from a generated object token.The study uses it to assess token-to-token object localization across decoder layers.Source: §2.1 · p. 3

2. Connecting Generated Object Token with Its Image Anchor via Attention

  • The analysis evaluates object localization through attention relationships between generated object tokens and their corresponding image regions.The model is represented as a vision encoder, vision-text connector, and transformer-based text decoder; LLaVA-1.5 maps one image to 576 image tokens.Source: §2.1 · p. 3, §2 · p. 3

2.1. Evaluating Object Localization Performance among Model Layers.

  • Generated object tokens assign more than 55% of visual attention to object anchors in layers 10–25, with layer 15 obtaining the highest OL-VAR.The analysis averages OL-VAR over 500 MSCOCO images for LLaVA-1.5 and LLaVA-Next.Source: §2.1 · p. 3, Figure 2 · p. 3
  • Layer 15 shows multiple distinct attention spots linking generated text tokens with image tokens, whereas layer 32 typifies the attention-sink pattern seen in most layers.The contrast supports the finding that object information is extracted mainly in early-middle to later-middle layers.Source: Figure 3 · p. 3, §2.1 · p. 4

3. Object Hallucination from Image Tokens.

The paper links object hallucinations to a small set of highly attended image tokens, then tests whether zeroing those tokens reduces hallucinations while preserving real objects. Its analysis also identifies a limitation of Logit Lens for finding all such tokens.

3.1. Identify Hallucinatory Image Tokens

  • Three tokens in the papaya region had the same effect as zeroing all 14 high-attention tokens, and Logit Lens projected them as “apple” and “orange.”The paper names tokens of this type Hallucinatory Image Tokens (HITs).Source: §3.1 · p. 4, Figure 4 · p. 4

3.2. Quantifying the Impact of Zeroing Out HITs

  • 86% of Hall-COCO images improved after zeroing the top five candidate HITs; images with hallucinations fell from 100% to 31.73%, and 74.17% of hallucinations disappeared.Figure 5 reports these changes under the same greedy decoding.Source: Figure 5 · p. 4, §3.2 · p. 5
  • Hall-COCO contains 200 MSCOCO images selected because human annotation identified them as consistently inducing object hallucinations.Source: §3.2 · p. 4
  • The paper cautions that Logit Lens alone cannot capture all HITs and identifies alternative interpretability tools as future work.Source: §3.2 · p. 4

3.3. Real Objects are Robust to Zero-Out

  • At K = 8, about 80% of hallucinated objects were removed, while only about 10% of real objects were affected; removed real objects tended to be small or peripheral.The comparison counts objects remaining in responses after zeroing the top-K image tokens attended by each object token.Source: §3.3 · p. 5

4. Detecting and Mitigating Object Hallucination by HIT Zero-outs

EAZY detects object hallucinations by testing whether generated objects disappear after their most-attended image tokens are zeroed. For mitigation, it combines candidate HIT sets for detected hallucinations and generates a response with those image tokens zeroed.

4. Detecting and Mitigating Object Hallucination by HIT Zero-outs

  • EAZY is an automatic approach to detecting and mitigating LVLM object hallucinations by zeroing out hallucinatory image tokens.Source: §4 · p. 5

4.1. Object Hallucination Detection

  • EAZY detects hallucinations by extracting object tokens, zeroing each object's top-K attended image tokens, and classifying objects absent from the updated response as hallucinations.Objects that remain in the updated response are classified as real.Source: §4.1 · p. 5

4.2. Mitigating Object Hallucination

  • For mitigation, EAZY estimates candidate HITs from an initial response, identifies objects that disappear after zeroing, and unions their candidate sets for final zeroing and generation.Figure 6 illustrates the process with hallucinated apples and oranges disappearing while kiwis are revealed.Source: §4.2 · pp. 5–6, Figure 6 · p. 6

5. Experiment

Experiments compare EAZY with training-free baselines on object-hallucination mitigation and detection, then test its effect on general model capability and design choices. EAZY leads the reported mitigation and detection comparisons, while ablations examine token selection and zero-out strategies.

5.1.1. OH Mitigation Setting

  • Mitigation experiments use LLaVA-1.5 and Shikra at 7B scale and LLaVA-Next at 8B scale, comparing EAZY with six training-free methods.The baselines include greedy and beam search, Dola, VCD, OPERA, and SID.Source: §5.1.1 · p. 6

5.1.2. OH Detection Setting

  • Detection is evaluated on LLaVA-1.5 with Hall-COCO, treating real objects as positive and hallucinated objects as negative instances.The reported metrics are accuracy, precision, recall, and F1 score.Source: §5.1.2 · p. 6

5.2.1. Evaluation Result on OH Mitigation

  • On CHAIR, EAZY surpasses baselines by a large margin across three LVLMs; on POPE, it has the best average performance across random, popular, and adversarial settings.POPE's advantage is less pronounced than its CHAIR results, which the authors attribute to the binary question format and prompts that already name the object.Source: Table 1 · p. 7, §5.2.1 · p. 7, Table 2 · p. 7

5.2.2. Evaluation Result on OH Detection

  • 16.21% accuracy, 16.48% precision of real objects, 15.09% precision of HOs, 8.60% ROC AUC, and 9.37% PR AUC improvements over the best baseline.The evaluation uses LLaVA-1.5 on Hall-COCO; Figure 7 presents precision-recall and ROC curves for EAZY and baselines.Source: Figure 7 · p. 6, Table 3 · p. 7, §5.2.2 · p. 7

5.3. Evaluation on MLLM benchmarks.

  • EAZY maintains LLaVA-1.5's general capability on the MME and MMBench benchmarks.The reported evaluation is on LLaVA-1.5-7b.Source: §5.3 · p. 7, Table 4 · p. 8

5.4. Ablation Study

  • LLaVA-1.5 achieves its best CHAIR performance at K = 5, while Shikra does so at K = 3.The comparison uses greedy decoding performance as the reference.Source: Figure 8 · p. 8, §5.4 · p. 8
  • Zeroing out the same image tokens at the LLM input, ViT output, or original-patch stage effectively mitigates object hallucination.The authors infer that the associated visual biases originate from the image itself and are not strongly associated with the ViT or linear projector.Source: Table 5 · p. 8, §5.4 · p. 8
  • Zero-out outperforms average, random, and neighbor token replacements on images with fewer hallucinations, no hallucinations, and removed hallucinations.The comparison is on Hall-COCO with K = 5.Source: Table 6 · p. 8, §5.4 · p. 8

6. Related Work

Prior work attributes object hallucination to training-data and language biases, while inference-time methods address it through decoding or anchor-token strategies. This paper instead identifies visual bias as a cause of most object hallucinations.

  • Prior explanations of object hallucination include training-data and language biases, including object-frequency and co-occurrence patterns and language-model prior knowledge.Some work also reports that LVLMs prefer text tokens and attend to neighboring text when generating new tokens.Source: §6 · p. 8
  • Inference-time mitigation includes contrastive decoding and methods targeting sequence-supervision-induced anchor tokens, avoiding costly retraining or fine-tuning.The passage describes contrastive decoding as comparing original logits with outputs from processed images.Source: §6 · pp. 8–9
  • This work identifies visual biases as the cause of most object hallucinations and proposes zeroing out hallucinatory image tokens.Source: §6 · p. 9

7. Conclusion

The paper identifies high-attention Hallucinatory Image Tokens as correlated with generated object hallucinations and reports that zeroing them removes hallucinations while minimally affecting real objects. It presents EAZY as a training-free method for detecting and mitigating object hallucinations, with strong performance in both tasks.

  • A small number of high-attention image tokens are directly correlated with hallucinated objects in generated text.Source: §7 · p. 9
  • Replacing Hallucinatory Image Tokens with zero embeddings effectively removes hallucinated objects while minimally affecting real objects.Source: §7 · p. 9
  • EAZY is a training-free method for automatically detecting and mitigating object hallucinations, with reported strong performance in both tasks.Source: §7 · p. 9

A. Hall-COCO: Benchmarking for Object Hallucination Mitigation Evaluation

Hall-COCO was curated from MSCOCO to provide images that reliably induce hallucinations, with human-annotated objects and corresponding high-attention image tokens. The appendix also specifies evaluated models, metrics, token-selection settings, and an iterative EAZY variant.

A. Hall-COCO: Benchmarking for Object Hallucination Mitigation Evaluation

  • Hall-COCO contains 200 MSCOCO images selected to consistently induce hallucinations, with human-identified hallucinated objects and their top five attention-scored image tokens.GPT-4o first detected hallucinations in LLaVA outputs; for multiple hallucinated objects, their token sets were combined by union.Source: A. Hall-COCO: Benchmarking for Object… · p. 11

B.1. Model Details.

  • The evaluated models are LLaVA-1.5, Shikra, and LLaVA-Next, with the passage describing their respective architectures and capabilities.Source: §B.1 · p. 11

B.2. Evaluation Metric of OH Mitigation

  • CHAIR measures object hallucination in image captioning at sentence and image levels, while POPE probes object presence with yes-or-no questions under three sampling settings.POPE's settings are random, popular, and adversarial sampling; the experiment reports average accuracy and F1 over the three modes on MSCOCO.Source: §B.2 · p. 11

B.3. Implementation Detail

  • Candidate HITs are selected from attention scores at layer 15 for LLaVA-1.5 and LLaVA-Next, and layer 18 for Shikra.The selected top-K counts are five tokens for LLaVA-1.5 and LLaVA-Next and three for Shikra.Source: §B.3 · p. 11

B.4. Better OH Mitigation via Iterative Hallucinatory Objects Detection

  • EAZY's iterative variant tests each object's candidate HITs separately, checking whether that object disappears; it generally improves CHAIR performance on MSCOCO.Separate testing avoids interference between candidate HITs associated with different text object tokens.Source: §B.4 · pp. 11–12

B.5. Additional Evaluation Result of OH Detection

EAZY maintains strong object-hallucination detection across LLaVA-Next and Shikra, while its real-object detection is less pronounced on those models. The section also reports inference-time, token-removal, and layer-choice results.

B.5. Additional Evaluation Result of OH Detection

  • Accuracy improves by 7.65% on LLaVA-Next and 12.98% on Shikra, while PR(OH) gains are 46.98% and 23.56%, respectively, over the best baseline.EAZY also has the highest F1 scores: 80.12% on LLaVA-Next and 76.61% on Shikra.Source: §B.5 · p. 12, Table 9 · p. 13, Table 10 · p. 13
  • EAZY’s real-object detection performance is less pronounced on LLaVA-Next and Shikra, where image-token transformations may reduce preserved visual information.Shikra maps 576 patch embeddings to 256 image tokens; LLaVA-Next uses AnyRes and requires more tokens to be removed than LLaVA-1.5.Source: §B.5 · p. 12

C. Efficiency Analysis

  • EAZY has 50% lower inference time than OPERA under the stated LLaVA-1.5 7B CHAIR evaluation setting.The comparison averages inference time over 500 images on NVIDIA A6000 GPUs.Source: C. Efficiency Analysis · p. 12

D. Statistic on Image Token Zeroing-out

F. Visual Bias is the Main Cause of Object Hallucination.

The authors examine visual bias as one contributor to object hallucination by adding Gaussian noise to weaken image information. Their analysis estimates that visual bias accounts for most hallucination variation, while language prior and other factors account for smaller shares.

F. Visual Bias is the Main Cause of Object Hallucination

  • The analysis distinguishes New Hallucinations, which appear after image noise is added, from Stable Hallucinations, which persist after noise.Gaussian noise is used to weaken visual information while amplifying language-prior and other influences.Source: F. Visual Bias is the Main Cause of… · p. 14
  • After 100 noise steps, New Hallucinations substantially outnumber Stable Hallucinations, whose proportion continues to decline.The authors interpret this pattern as indicating that visual bias is a key factor in OH variation and generation.Source: F. Visual Bias is the Main Cause of… · p. 14
  • At 300 noise steps, visual bias accounts for 80% of OH variations, language prior for 8.6%, and other factors for 11.4%.The authors report that 20% of HOs are stable, and 43% of those stable HOs cannot be eliminated by EAZY.Source: F. Visual Bias is the Main Cause of… · p. 14, Figure 13 · p. 14

G. Hallucinatory objects are looking at image regions visually close to them.

Hallucinatory objects attend to visually related image regions rather than a true image anchor, whereas real objects tend to attend to their own anchors. The section also describes cases where EAZY can remove real objects and summarizes its HIT-based approach.

G. Hallucinatory objects are looking at image regions visually close to them

H. Failure Case Analysis

I.1. Summary of Key Findings

  • EAZY replaces top-K candidate HITs with zero embeddings to suppress hallucinated objects while preserving real-object information.The authors characterize HITs as a small subset of image tokens whose removal can significantly mitigate hallucinations.Source: §I.1 · pp. 15–16

I.2. Limitations and Future Work

The authors identify limitations in explaining HITs and in handling some real objects, and outline directions for future work. Case studies illustrate EAZY’s object corrections and occasional improvements to image descriptions.

I.2. Limitations and Future Work

  • Existing interpretability tools such as Logit Lens are insufficient to fully explain and identify all HITs, motivating further investigation and improved distinction from top-K candidates.The stated goal is to improve EAZY’s efficiency in zeroing out image tokens.Source: §I.2 · p. 16
  • EAZY may erroneously remove real objects that are small, split into multiple tokens, or located at image boundaries.Proposed future directions include repeated inference, uncertainty estimation, and external vision models to improve identification.Source: §I.2 · p. 16

J. Case Study

  • Case studies show EAZY removing hallucinated objects and, in some examples, correcting object counts or inaccurate scene details.One example reports that zeroing out a single image token corrected cat and cup counts while removing hallucinated objects.Source: Figure 20 · p. 17, Figure 21 · p. 17, Figure 22 · p. 18

Related papers

Loading 2503.07772v2…