PaperlayerView exact revision on arXiv

Source-linked AI summary

IBD: Alleviating Hallucinations in Large Vision-Language Models via Image-Biased Decoding

Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, Jun Liu

arXiv:2402.18476v1cs.CV

TL;DR

LVLMs can hallucinate, with over-reliance on linguistic priors identified as a contributing factor. IBD contrasts ordinary and image-biased predictions, and the paper reports high effectiveness across multiple evaluation metrics.

  • Problem

    LVLMs generate information unrelated to or incorrect with respect to the image and text inputs, and over-reliance on linguistic priors is identified as a contributing factor.

    Source: §1 · p. 1

  • Method

    IBD contrasts predictions from an ordinary LVLM and an image-biased model to emphasize image-related information and uses dynamic adjustment for different vocabulary types.

    Source: §1 · p. 2, §5 · p. 8

  • Results

    Validation experiments report leading results across multiple metrics, and the conclusion describes IBD as highly effective in addressing LVLM hallucinations.

    Source: §1 · p. 2, §5 · p. 8

  • Takeaways & Limitations

    The paper presents IBD as a flexible approach with low extra parameter and data costs that may provide insights for future research.

    Source: §1 · p. 2, §5 · p. 8

    The authors do not develop targeted solutions for image-biased hallucinations and leave comprehensive assessment and corresponding solutions for future research.

    Source: §3.5 · p. 6

Abstract

from arXiv · show

Despite achieving rapid developments and with widespread applications, Large Vision-Language Models (LVLMs) confront a serious challenge of being prone to generating hallucinations. An over-reliance on linguistic priors has been identified as a key factor leading to these hallucinations. In this paper, we propose to alleviate this problem by introducing a novel image-biased decoding (IBD) technique. Our method derives the next-token probability distribution by contrasting predictions from a conventional LVLM with those of an image-biased LVLM, thereby amplifying the correct information highly correlated with image content while mitigating the hallucinatory errors caused by excessive dependence on text. We further conduct a comprehensive statistical analysis to validate the reliability of our method, and design an adaptive adjustment strategy to achieve robust and flexible handling under varying conditions. Experimental results across multiple evaluation metrics verify that our method, despite not requiring additional training data and only with a minimal increase in model parameters, can significantly reduce hallucinations in LVLMs and enhance the truthfulness of the generated response.

1. Introduction

LVLM hallucinations are a safety and reliability concern linked to over-reliance on linguistic priors, especially during long responses. IBD contrasts ordinary and image-biased predictions to favor image-related content; its validation reports leading results across multiple metrics.

1. Introduction

  • LVLM hallucinations include information unrelated to or incorrect with respect to the inputs, and excessive reliance on linguistic priors can divert generation from visual content.The introduction links this over-reliance to models continuing from previously generated text rather than drawing logical inferences from the image.Source: §1 · p. 1
  • Later segments of long generated responses are more susceptible to hallucinations, according to empirical observations (Zhou et al., 2023).Source: §1 · p. 2
  • IBD contrasts ordinary and image-biased model predictions to favor image-related information and suppress hallucinations associated with textual over-reliance.The image-biased model is formed by modifying attention weights without changing the original model’s parameters; the example contrasts ‘keyboard’ with ‘plant’.Source: Figure 1 · p. 1, §1 · p. 2
  • Statistical analysis finds correlation, in certain conditions, between tokens receiving the highest probability boost and correct tokens without hallucinations, informing dynamic adjustment for potential failure scenarios.The method is designed to handle different vocabulary types flexibly.Source: §1 · p. 2
  • IBD is presented as having low extra parameter and data costs compared with approaches requiring additional training data or considerable parameter expansion.Source: §1 · p. 2
  • IBD’s validation experiments report leading results across multiple metrics, demonstrating effectiveness in mitigating LVLM hallucinations.The introduction presents these as benchmark results against other advanced methods.Source: §1 · p. 2

2. Related Work

Prior work addresses LVLM hallucinations through fine-tuning, supplementary networks, and decoding strategies. The paper positions IBD as a decoding-focused alternative that avoids substantial additional data or parameters and addresses limitations of earlier contrastive decoding.

2. Related Work

  • Fine-tuning on fine-grained datasets and supplementary networks are prior approaches, but they typically require substantial additional training data or model parameters.Source: §2 · p. 2
  • IBD focuses on refining decoding rather than using extensive fine-tuning or supplementary networks, with the paper framing this as a lower-overhead route to reducing hallucinations.Source: §2 · p. 2
  • Traditional decoding strategies, including beam search, top-k decoding, and sampling, are described as prone to hallucinatory content, motivating research into alternative decoding methods.Source: §2 · p. 2
  • Existing LVLM contrastive decoding uses randomly generated noise that can introduce uncontrollable, unstable perturbations; the paper reports superior performance for its approach in Table 1.Source: §2 · p. 3

3. Method

The method contrasts ordinary and image-biased predictions to adjust next-token decoding, motivated by the risk that linguistic priors produce text-biased hallucinations. Its analyses identify when this contrast is useful and inform adaptive decoding, while discussion notes a distinct image-biased failure mode.

3. Method

  • IBD contrasts ordinary and image-biased model logits to score candidate tokens, boosting information associated with image content against text-biased hallucination.The proposed image-biased model shares the original model’s parameters.Source: §3 · p. 3, §3.1 · p. 3

3.1. Image-biased Model

  • Adding a coefficient to attention calculations for image-token keys increases visual focus while leaving the vanilla model’s parameter weights unchanged.The modified attention structure can introduce noise; prompt tuning is later used to adapt the image-biased model.Source: §3.1 · p. 3

3.2. Hypothesis Testing

  • Across four LVLMs, content-word ground truths frequently have the highest CD score, whereas function-word ground truths frequently score below hallucinatory candidates.The figure reports the proportions of ground tokens with and without the highest CD score; the prose attributes the difference to visual association for content words and linguistic priors for function words.Source: Figure 2 · p. 4, §3.2 · p. 4
  • When predictions from the two models are too similar, lower JSD correlates with a greater chance that the highest-CD candidate is not the ground truth.The paper attributes reduced identification accuracy in this condition to low-valued CD scores being more susceptible to noise.Source: Figure 3 · p. 4, §3.2 · p. 4

3.3. Dynamic Adjustment For Decoding

  • The adaptive decoder balances maximum-likelihood decoding and image-biased contrastive decoding using prediction divergence and an indicator of content-word probability.The content-word indicator is estimated from the distance between final-layer and middle-layer predictions, avoiding direct POS tagging that may mislabel subword tokens.Source: §3.3 · p. 5

3.4. Full Method

  • The full method adds prompt tuning for the image-biased model and an adaptive plausibility constraint that selects only sufficiently probable candidate tokens.Prompt tuning adapts the modified attention architecture without LoRA’s stated additional-parameter burden; the plausibility constraint is intended to prevent false positives and false negatives in contrastive decoding.Source: §3.4 · pp. 5–6

3.5. Discussion: Are there Image-biased Hallucinations?

  • Image-biased hallucinations can occur when visual cues conflict with world knowledge, but the paper does not propose targeted solutions for them.The authors instead identify comprehensive assessment and corresponding solutions as future research; they note such cases are rarer in mainstream evaluation settings.Source: §3.5 · p. 6

4. Experiments

The experiments compare IBD with other methods using CHAIR and assisted evaluations, then examine component and hyperparameter settings. The reported results favor IBD across baselines, while CHAIR performance depends on the choice of ϵ.

4.1. Settings

  • Experiments use InstructBLIP, MiniGPT-4, LLaVA-1.5, and Shikra, each with a 7B LLM; settings include ϵ=2, α=1.5 × 10^4, and β=0.1.The selected hyperparameters are examined in the ablation study and appendix.Source: §4.1 · p. 6

4.2. Main Results

  • On CHAIR, IBD consistently outperforms the compared method types across baselines, with 74K additional prompt parameters.The comparison covers traditional decoding, hallucination-focused decoding, and assistance-based methods.Source: Table 1 · p. 7, Table 2 · p. 7, §4.2 · p. 7
  • GPT-4 evaluation on VG assesses hallucinations beyond object existence, including attributes and relations, using four hallucination-related metrics.Figure 5 names the metrics HSPI, HWPI, HSR, and HWR; lower values indicate fewer hallucinations.Source: §4.2 · p. 7, Figure 5 · p. 7
  • 11.5% higher Shikra Detailedness than the runner-up; GPT-4V evaluation also finds significant enhancement over other methods.GPT-4V scores description Accuracy and Detailedness from 0 to 10.Source: Table 3 · p. 8, §4.2 · p. 8

4.3. Ablation Study

  • Table 4 evaluates IBD components on CHAIR, including dynamic adjustment (DA) and the adaptive plausibility constraint (APC).The paper reports that its component evaluations demonstrate the designs’ effectiveness.Source: Table 4 · p. 8, §4.3 · p. 8
  • CHAIR performance is relatively stable for 1.5 < ϵ < 3, while values that are too small or too large perform suboptimally.The paper attributes the extremes to insufficient image bias and excess noise, respectively.Source: Figure 6 · p. 8, §4.3 · p. 8

5. Conclusion

The conclusion presents IBD as a decoding approach for alleviating LVLM hallucinations and reports experimental effectiveness.

5. Conclusion

  • IBD contrasts predictions from the original and image-biased models to amplify image-related information and improve generated-text factuality.The method also uses dynamic adjustment to handle different vocabulary types.Source: §5 · p. 8
  • The authors report that experiments demonstrate IBD’s effectiveness in addressing LVLM hallucinations.They present their statistical analysis as a potential source of insights for future research.Source: §5 · p. 8
  • The paper frames its statistical analysis and method as contributions that may provide insights for future research.This is stated as the authors’ hope.Source: §5 · p. 8

A.1. Early Exit’s Patterns for Different Words

The section validates early-exit prediction patterns in LVLMs: content words and ambiguous content subwords continue changing in deeper layers, unlike most function words. These patterns support the indicator used by the method.

A.1. Early Exit’s Patterns for Different Words

  • Prior LLM observations found that function-word predictions stabilize by middle layers while content-word predictions keep changing; the paper uses these patterns to compute Icon.The cited observations are from Chuang et al. (2023).Source: §A.1 · p. 11
  • 74.14% of content-word predictions had JSD >1.4, while 88.47% of function-word predictions had JSD <0.4 between LLaVA-1.5’s 24th and final layers.The results align with the previously observed distinction between content and function words in LLMs (Chuang et al., 2023).Source: §A.1 · p. 11
  • 70.45% of ambiguous content-subword predictions had JSD >1.4, indicating that these predictions also change in deeper LVLM layers.The paper interprets this as ambiguous content subwords following the pattern observed for content words (Chuang et al., 2023).Source: §A.1 · p. 11

A.2. Image-biased Hallucinations in Mainstream Evaluation Frameworks

In a sample of LLaVA-1.5 errors from MME, GPT-4V attributed only 4% to image-biased hallucinations. The analysis treats such hallucinations as errors driven by misleading image regions or information.

A.2. Image-biased Hallucinations in Mainstream Evaluation Frameworks

  • GPT-4V assessed 500 randomly selected LLaVA-1.5 errors from MME, a dataset with 14 subtasks.The assessment asked whether each erroneous response could be attributed to image-biased hallucinations.Source: §A.2 · p. 11
  • Image-biased hallucinations are incorrect responses influenced by particular image regions or information, such as mistaking a nearby table-tennis racket as evidence that a person plays table tennis.Source: §A.2 · p. 12
  • Only 4% of the 500 sampled incorrect LLaVA-1.5 responses on MME were attributed by GPT-4V to image-biased hallucinations.The authors take this as evidence that image-biased hallucinations are rare in mainstream LVLM evaluation frameworks.Source: §A.2 · p. 12

B. More Discussions: Why not directly train an image-biased LVLM?

The discussion contrasts direct use of an image-biased model’s predictions with IBD’s focus on changes in predictions. It argues that very small or large structural changes can limit the direct model’s performance, while contrastive decoding can identify correct tokens with small epsilon.

B. More Discussions: Why not directly train an image-biased LVLM?

C. More Ablation Study for Hyper-parameters

The ablation varies IBD’s α, β, and early-exit layer settings on LLaVA-1.5 with CHAIR and reports consistent performance. The accompanying figure captions identify the CHAIR metrics and the qualitative comparison’s color key.

C. More Ablation Study for Hyper-parameters

  • The hyperparameter assessment uses the LLaVA-1.5 baseline and CHAIR metric, varying α, β, and the layer used for the early-exiting prediction.Source: C. More Ablation Study for… · p. 12
  • LLaVA-1.5 showed consistent CHAIR performance across α, β, and early-exit layer settings, demonstrating robustness to these hyperparameters.Figure 8 evaluates the three settings; its first-row plots show CHAIR CS and its second-row plots show CHAIR CI, with smaller values indicating fewer hallucinations.Source: C. More Ablation Study for… · pp. 12–13, Figure 8 · p. 13
  • Figure 9 marks regular-decoding hallucinations in red and information omitted by regular decoding but generated by IBD in green.Source: Figure 9 · p. 13
Loading 2402.18476v1…