Mitigating Hallucinations in Large Vision-Language Models with Instruction Contrastive Decoding

Xintong Wang; Jingheng Pan; Liang Ding; Chris Biemann

Source-linked AI summary

Mitigating Hallucinations in Large Vision-Language Models with Instruction Contrastive Decoding

Xintong Wang, Jingheng Pan, Liang Ding, Chris Biemann

arXiv:2403.18715v2Mar 27, 2024cs.CVcs.AIcs.CLcs.MM

TL;DR

LVLMs can generate fluent responses that inaccurately reflect visual content, motivating methods that reduce hallucinations during inference. ICD contrasts distributions from standard and disturbance instructions to detach hallucinated concepts, and experiments report reduced object- and attribute-level hallucinations alongside improved general perception and recognition.

Problem
LVLMs generate fluent, coherent text that can inaccurately represent visual content, including hallucinated objects, attributes, and relationships.
Method
ICD is a training-free, model-agnostic inference method that contrasts distributions from original and disturbance instructions to suppress hallucinated concepts.
Results
Experiments across POPE, MME, and LLaVa-Bench report effective mitigation of object- and attribute-level hallucinations and improved general perception and recognition.
Takeaways & Limitations
ICD provides a training-free approach for reducing hallucinations while enhancing LVLM perception and recognition performance.
Takeaways & Limitations
Generative benchmarks lack established metrics for thoroughly analyzing hallucinations, limiting evaluation of open-ended generation.

Abstract

from arXiv · show

Large Vision-Language Models (LVLMs) are increasingly adept at generating contextually detailed and coherent responses from visual inputs. However, their application in multimodal decision-making and open-ended generation is hindered by a notable rate of hallucinations, where generated text inaccurately represents the visual contents. To address this issue, this paper introduces the Instruction Contrastive Decoding (ICD) method, a novel approach designed to reduce hallucinations during LVLM inference. Our method is inspired by our observation that what we call disturbance instructions significantly exacerbate hallucinations in multimodal fusion modules. ICD contrasts distributions from standard and instruction disturbance, thereby increasing alignment uncertainty and effectively subtracting hallucinated concepts from the original distribution. Through comprehensive experiments on discriminative benchmarks (POPE and MME) and a generative benchmark (LLaVa-Bench), we demonstrate that ICD significantly mitigates both object-level and attribute-level hallucinations. Moreover, our method not only addresses hallucinations but also significantly enhances the general perception and recognition capabilities of LVLMs.

1 Introduction

LVLMs produce fluent but visually inaccurate content, with hallucinations spanning objects, attributes, and relationships. ICD addresses this by contrasting standard and disturbance-instruction distributions to suppress hallucinated concepts during inference.

LVLM hallucinations are fluent, coherent responses that fail to reflect factual visual content.
Hallucinations extend beyond object existence to include erroneous attributes and relationships.
Known contributors include training-time statistical bias, language-prior dependence, and multimodal misalignment.
Existing mitigation strategies can require substantial human involvement and significant additional computational costs.
ICD contrasts original and disturbance-instruction distributions to suppress hallucinations, while remaining training-free and agnostic to the underlying LVLM.
Experiments across discrimination and generation benchmarks validate ICD's effectiveness and robustness in enhancing LVLM performance.

2 Related Work

Related work frames VLM hallucinations as fluent but visually inaccurate outputs involving objects, attributes, and relations. Prior mitigation strategies include annotation enrichment and methods addressing multimodal grounding, while ICD suppresses hallucinations through disturbance instructions.

LVLMs build on advances in large language models and multimodal alignment modules, using architectures such as linear projections and Q-Formers.
VLM hallucinations manifest as detailed, fluent, and coherent responses containing erroneous objects, attributes, and relations.
Prior mitigation approaches include annotation enrichment and methods targeting hallucination reduction through multimodal alignment.

3 Method

The method analyzes how instruction disturbances increase hallucinations through multimodal alignment uncertainty and uses Instruction Contrastive Decoding to suppress disturbance-associated concepts during inference.

3.1 Inference in LVLMs: LVLM inference combines a visual encoder, a multimodal fusion module, and a language model to generate responses from visual features and text queries.InstructBLIP uses an instruction-aware Q-Former, while the language model transforms fused visual features into responses.
3.1 Inference in LVLMs: Hallucinations emerge when probable tokens lack grounding in visual context, with multimodal misalignment identified as a critical contributing factor.The method therefore focuses on instructions inside the multimodal alignment module.
3.2 Instruction Can Amplify Hallucination: Instruction disturbance appends role prefixes to original instructions to alter multimodal alignment confidence, either encouraging definitiveness or introducing ambiguity.The disturbance is represented by combining a role prefix Xd with the original instruction Xq.
3.2 Instruction Can Amplify Hallucination: Instruction disturbances amplify hallucinated objects and biased co-occurrence concepts, including “person and dining table” and “fork and person.”The analysis compares baseline, positive-disturbance, and negative-disturbance settings using hallucination ratios on the MSCOCO validation set.
3.3.1 Contrastive Decoding with Disturbance: ICD contrasts token distributions from original and disturbed instructions, favoring tokens supported by the original output while penalizing disturbance-influenced tokens.The hyperparameter λ controls the intensity of the contrastive penalty, with larger values applying a more decisive penalty.
3.3.1 Contrastive Decoding with Disturbance: The decoding objective can also penalize accurate predictions, motivating probability-distribution truncation to reduce the influence of implausible tokens.The truncation parameter α modulates this adjustment according to the LVLM’s confidence level.

4 Experiment

The experiments evaluate ICD on object- and attribute-level hallucination discrimination, as well as open-ended generation, using POPE, MME, and LLaVa-Bench. Across these evaluations, ICD improves hallucination-related performance over foundation LVLMs and VCD, while also improving broader MME task scores.

Datasets and Evaluation Metrics: POPE evaluates object-level hallucinations across MSCOCO, A-OKVQA, and GQA using Accuracy, Precision, Recall, and F1.The benchmark uses binary questions in random, popular, and adversarial settings.
Results on POPE: ICD improves POPE performance over miniGPT4 and InstructBLIP across all four metrics and three subsets, with reported gains of 10.5% and 6.0%, respectively.The reported metric improvements are 7.0% in accuracy, 8.5% in precision, 8.7% in recall, and 7.9% in F1 for both models.
Results on POPE: Performance declines from random to popular to adversarial POPE settings, while ICD improves across all settings by subtracting errors associated with statistical bias and language priors.The paper links these settings to increasing effects of statistical biases and language priors on hallucinations.
Results on MME: On the full 14-task MME benchmark, ICD improves scores with both miniGPT4 and InstructBLIP, and it outperforms miniGPT4 and matched-backbone VCD across all miniGPT4 subtasks.Under InstructBLIP, VCD is weaker on several tasks, while ICD is reported to preserve broader reasoning capability.
Discussions on ICD and VCD: The VCD-enhanced ICD variant contrasts VCD outputs under standard and disturbance instructions, combining visual uncertainty with multimodal-alignment uncertainty.The combined method is evaluated on selected MME subtasks including color, posters, landmarks, OCR, commonsense reasoning, and text translation.

5 Conclusion

The paper introduces instruction contrastive decoding, which detaches hallucinatory concepts by contrasting distributions from standard and disturbance instructions. Experiments across benchmarks and LVLMs show reduced hallucinations and improved general perception and recognition.

Instruction contrastive decoding detaches hallucinatory concepts by contrasting distributions from standard and disturbance instructions with appended role prefixes.The method uses a highlight-then-detach strategy during inference.
Comprehensive experiments across various benchmarks and LVLMs demonstrate that ICD mitigates hallucinations.
ICD substantially improves the general perception and recognition performance of LVLMs.

Limitations

The paper evaluates ICD on hallucination discrimination benchmarks and qualitatively assesses generative hallucination, but robust automatic metrics for open-ended generation remain unavailable.

The paper rigorously evaluates ICD on various hallucination discrimination benchmarks and qualitatively assesses it on generative benchmarks.
Generative benchmarks lack established metrics for thoroughly analyzing hallucinations in open-ended generation.The paper identifies robust automatic metrics for this evaluation as an area for future research.
Improving automatic evaluation metrics remains important for assessing open-ended generation performance.

Ethics Statement

The paper proposes ICD to address LVLM hallucinations and reports publicly accessible datasets and code to support transparency, reproducibility, and implementation. Its experiments use specified contrastive-decoding and sampling configurations, alongside positive and negative role-prefix disturbances.

Ethics Statement: The paper proposes Instruction Contrastive Decoding to address hallucination issues in LVLMs and enhance their safety and reliability.
Ethics Statement: The datasets used for inferring and evaluating ICD are publicly accessible, promoting transparency and reproducibility.
Ethics Statement: The authors make their code publicly available so researchers and practitioners can access and implement the method.
Ethics Statement: Experiments set the disturbance penalty λ = 1 and truncation hyperparameter α = 0.1, using sampling with top p = 1 and repetition penalty = 1.
Ethics Statement: The experiments use positive and negative role-prefix instructions as instructional disturbances.

B Qualitative Evaluation on LLaVa-Bench

On LLaVa-Bench qualitative cases, baseline LVLMs introduce hallucinated concepts linked to statistical co-occurrence and language-prior biases, whereas ICD produces text more closely grounded in visual context.

Baseline LVLMs identify spoons, tables, and cups in an Asian-dishes image but also introduce the unrelated concept of a person.The person error is attributed to frequent person–table co-occurrence in training data.
In a Titanic scene, baseline LVLMs incorrectly perceive Jack and Rose as two women, producing inaccurate text about same-sex relationships.The passage attributes this error to language-prior biases.
ICD produces fluent, coherent text that is closely grounded in visual context and mitigates hallucinations caused by statistical biases and language priors.

C General QA Benchmarks Performance

The paper extends ICD evaluation beyond hallucination-focused benchmarks to general QA and captioning datasets, using CHAIR, CIDEr, and BLEU. These evaluations report improvements across MSCOCO, OK-VQA, and TextVQA.

Evaluation scope: The evaluation broadens beyond POPE, MME, and LLaVa-Bench to MSCOCO, OK-VQA, and TextVQA.The additional datasets assess general question answering and captioning.
MSCOCO: MSCOCO results measure instance-level and sentence-level hallucinations with CHAIR_I and CHAIR_S.Table 3 uses the MSCOCO validation set, while Table 4 uses 500 sampled training and validation instances.
OK-VQA: OK-VQA performance is evaluated with CIDEr and BLEU metrics.The reported results are presented in Table 5.
TextVQA: TextVQA performance is evaluated with CIDEr and BLEU 1, 2, 3, and 4.The reported results are presented in Table 6.
Overall findings: Improvements in CHAIR, CIDEr, and BLEU across the three benchmarks support ICD’s versatility and efficacy.The paper describes these gains as reaffirming the method’s significant impact.

D Further Analysis on VCD-Enhanced ICD

The paper examines combining ICD with VCD on the full MME benchmark using InstructBLIP. The combination improves performance across multiple task types, while the authors identify the need for more refined integration strategies.

Combined-method evaluation: The combined ICD and VCD method is evaluated on the full MME benchmark with InstructBLIP as the backbone LVLM.The analysis is illustrated in Figure 6.
Combined-method evaluation: Integrating ICD significantly enhances VCD performance across existence, count, color, celebrity, scene, landmark, and artwork tasks.The reported task-level improvements concern the full MME evaluation.
Combined-method evaluation: Incorporating VCD into ICD also improves performance on color-related evaluation.The supplied passage reports this improvement while indicating that the detailed list is truncated.
Future direction: The authors identify refined combination strategies as necessary to fully harness the potential of integrating ICD and VCD.Developing and refining contrastive decoding methods for seamless integration is proposed as future work.

E Optimal Position to Apply Contrastive Decoding

The paper compares three insertion points for ICD—Q-Former instructions, LLM instructions, and both—using POPE GQA Random with InstructBLIP and LLaVa. Results indicate that Q-Former placement is generally most effective, especially for InstructBLIP.

Experimental design: ICD is evaluated at the Q-Former instruction, LLM instruction, and combined instruction positions.The comparison uses POPE in the GQA random setting with InstructBLIP and LLaVa backbones.
Position comparison: Under InstructBLIP, ICD improves performance at all tested implementation sites, with Q-Former-only placement producing the largest improvement.Figure 7 reports the comparison on POPE GQA Random.
Position comparison: LLaVa also benefits when ICD is applied within the LLM, although exclusive LLM application yields less pronounced improvements.The latter pattern mirrors the observations under InstructBLIP.
Position comparison: The findings identify deployment within the Q-Former architecture as the most effective ICD strategy.This conclusion follows the cross-model position comparison described in the supplied passages.