PaperlayerView exact revision on arXiv

Source-linked AI summary

Mitigating Object Hallucination via Concentric Causal Attention

Yun Xing, Yiheng Li, Ivan Laptev, Shijian Lu

arXiv:2410.15926v1cs.CVcs.CL

TL;DR

Object hallucination leaves LVLM responses factually misaligned with images, and the paper associates it with RoPE’s long-term decay across distant visual and instruction tokens. It proposes Concentric Causal Attention (CCA), which reorganizes visual tokens and rectifies causal masking; CCA outperforms existing mitigation methods on object-hallucination benchmarks, including on POPE.

  • Problem

    LVLMs can generate responses that are not factually aligned with image inputs, challenging their faithfulness and reliability.

    Source: §1 · p. 1

  • Method

    CCA reorganizes visual tokens from image peripheries toward centers and rectifies causal masking to reduce visual–instruction distances and model 2-D spatial locality.

    Source: §1 · p. 2, §4 · pp. 6–7

  • Results

    CCA surpasses existing hallucination mitigation methods on multiple benchmarks; on POPE, it improves Accuracy by +4.24% and F1 score by +2.73% over [34].

    Source: §1 · p. 2

  • Takeaways & Limitations

    The results support positional alignment as a way to mitigate RoPE’s long-term decay and object hallucination while improving general perception across six multimodal benchmarks.

    Source: §1 · p. 2, §5 · p. 7

    The study focuses on image-text inputs; positional alignment for audio and video is left to future work.

    Source: §6 · p. 9

Abstract

from arXiv · show

Recent Large Vision Language Models (LVLMs) present remarkable zero-shot conversational and reasoning capabilities given multimodal queries. Nevertheless, they suffer from object hallucination, a phenomenon where LVLMs are prone to generate textual responses not factually aligned with image inputs. Our pilot study reveals that object hallucination is closely tied with Rotary Position Encoding (RoPE), a widely adopted positional dependency modeling design in existing LVLMs. Due to the long-term decay in RoPE, LVLMs tend to hallucinate more when relevant visual cues are distant from instruction tokens in the multimodal input sequence. Additionally, we observe a similar effect when reversing the sequential order of visual tokens during multimodal alignment. Our tests indicate that long-term decay in RoPE poses challenges to LVLMs while capturing visual-instruction interactions across long distances. We propose Concentric Causal Attention (CCA), a simple yet effective positional alignment strategy that mitigates the impact of RoPE long-term decay in LVLMs by naturally reducing relative distance between visual and instruction tokens. With CCA, visual tokens can better interact with instruction tokens, thereby enhancing model's perception capability and alleviating object hallucination. Without bells and whistles, our positional alignment method surpasses existing hallucination mitigation strategies by large margins on multiple object hallucination benchmarks.

1 Introduction

LVLMs can produce responses that do not match image content, undermining their reliability. The paper links this problem to RoPE’s long-term decay and proposes CCA to reduce visual–instruction distances.

1 Introduction

  • Object hallucination makes LVLM responses factually misaligned with images, challenging their faithfulness and practical reliability.Existing mitigation approaches include post-hoc correction, instruction-tuning changes, and training-free decoding adjustments, each with implementation or inference-efficiency costs.Source: §1 · p. 1
  • RoPE attenuates information flow from visual tokens farthest from instruction tokens, a pattern associated with weaker multimodal alignment and object hallucination.The illustrated comparison aggregates visual-to-instruction information flow; the reported effect is especially apparent for leftmost visual tokens.Source: §1 · p. 2
  • On POPE, CCA improves Accuracy by +4.24% and F1 score by +2.73% over the state-of-the-art method [34].The authors also report consistent gains over the baseline across six multimodal benchmarks.Source: §1 · p. 2
  • CCA reorganizes visual tokens from image peripheries toward centers and rectifies causal attention to reflect 2-D spatial locality.The strategy is intended to reduce relative distances between visual and instruction tokens.Source: §1 · p. 2
  • The paper analyzes the correlation between RoPE and hallucination, then evaluates CCA against existing mitigation methods on multiple benchmarks.It also reports improved general perception over the baseline across six multimodal benchmarks.Source: §1 · pp. 2–3

2 Related Works

Related work covers the development of conversational LVLMs, mitigation strategies for object hallucination, and transformer positional encodings. The paper situates its analysis at the intersection of RoPE and hallucination.

Large Vision Language Models

  • LVLMs connect vision encoders with language models, enabling multimodal interaction and capabilities such as referential dialogue and interleaved image-text handling.Examples include LLaVA, InstructBLIP [15], and MiniGPT4 [84].Source: §2 · p. 3

Object Hallucination

  • Hallucination mitigation includes post-hoc correction [73] [83] [33], instruction-tuning changes [43] [76] [62] [29], and training-free decoding methods [26] [34] [10].The approaches may require labor-intensive annotations or slower decoding through candidate comparisons.Source: §2 · p. 3

Position Encoding in Transformers

  • RoPE encodes position by rotating input embeddings and provides relative-position information in self-attention.The paper investigates how this positional design relates to object hallucination in LVLMs.Source: §2 · p. 3

3 Motivation

The section examines how RoPE’s long-term decay may weaken visual-instruction alignment and contribute to object hallucination. In a pilot experiment, correctness varied with object position in opposite ways for raster-scan and reverse-raster-scan models.

RoPE Long-term Decay

  • RoPE long-term decay reduces attention as relative token distance grows, weakening visual-to-instruction information flow for visual tokens farther from instruction tokens.The section notes this contrasts with language modeling, where weaker connections between distant tokens are favorable.Source: §3 · p. 4

Pilot Experiment

  • The pilot compares two LVLMs trained with raster-scan and reverse raster-scan visual-token orders, then tests object recognition across pasted image positions.For each cropped object, the experiment asks whether it is present in a template at different spatial positions.Source: §3 · pp. 4–5
  • Reversing visual-token order changes relative distances to instruction tokens; the section uses this contrast to examine the relationship between RoPE decay and hallucination.For example, the reverse order changes the relative distance between the final visual token and the first instruction token from 1 to V.Source: §3 · p. 5
  • Correct responses were more likely for lower-positioned objects with the raster-scan model, but for upper-positioned objects with the reverse raster-scan model.The results are shown in Figure 2(a) for Fb and Figure 2(b) for Fr.Source: §3 · p. 5, Figure 2 · p. 5

4 Concentric Causal Attention

Concentric Causal Attention (CCA) reorganizes visual tokens and rectifies causal masking to reduce visual-instruction distance while preserving 2-D spatial structure. Its masking and ordering are designed to model continuous positional dependencies in images.

Concentric Causal Attention

  • CCA combines concentric visual-token reorganization with a rectified causal mask to reduce RoPE long-term decay’s negative effect on object hallucination.The design aims to minimize visual-instruction distance and mitigate performance discrepancies between raster-scan and reverse raster-scan models.Source: §4 · pp. 6–7

Concentric Positions

  • Concentric ordering arranges visual tokens from the image periphery toward the center, reducing their relative distance to instruction tokens while retaining 2-D spatial locality.Unlike raster-scan ordering, the sequence increases across concentric square positions.Source: §4 · p. 6, Figure 3 · p. 6
  • Under concentric ordering, nearby visual tokens are causally correlated as position increases, while tokens sharing a position are correlated in visual self-attention.The ordering is intended to better maintain the spatial locality of 2-D image features.Source: §4 · p. 6

Concentric Causal Masking

  • CCA’s causal mask models 2-D continuity by allowing center visual tokens to attend to peripheral ones, rather than only preceding tokens in a 1-D sequence.The illustrated mask uses 36 visual tokens arranged in a 6-by-6 layout.Source: Figure 3 · p. 6, §4 · p. 7

5 Experiments

The experiments evaluate the position-alignment approach on object-hallucination benchmarks and broader perception tasks. Results report gains on POPE, CHAIR, and MME, as well as better GPT-4 ratings than OPERA on LLaVA-Bench.

5 Experiments

  • Experiments report consistent gains over the baseline across six multimodal benchmarks, with Vicuna-7B used by default unless stated otherwise.These broader perception results are reported in Appendix D.1.Source: §5 · p. 7

5.1 Training Details

  • Training uses CLIP ViT-L/14 at 336x336 with Vicuna-7B and a 2-layer MLP, followed by pre-training on CC-558K and instruction tuning on 665k conversations.The pre-training and instruction-tuning global batch sizes are 256 and 128, respectively.Source: §5.1 · p. 7

5.2 POPE

  • Across POPE datasets and negative-sampling setups, the method achieves the highest accuracy and F1, improving accuracy over baseline by 5.48%, 7.86%, and 6.70% and F1 by 5.89%, 7.71%, and 6.19%.It also surpasses VCD by 1.02%, 4.51%, and 2.65% across three datasets.Source: §5.2 · p. 7, Table 1 · p. 8

5.3 CHAIR

  • CHAIR scores depend on response length and prompt phrasing, so evaluation compares short (64-token) and long (512-token) outputs using a fixed setup.The evaluation samples 500 images from COCO VAL 2014.Source: §5.3 · p. 8, Table 2 · p. 9
  • CHAIR evaluates caption hallucination at instance and sentence levels; for greedy long responses, the method improves over baseline by 3.2% while object recall is 80.3% v.s. 80.4%.For short responses, it improves over baseline by 2.8% on sentence-level and 0.8% on instance-level while maintaining high object recall.Source: §5.3 · pp. 7–8, Table 2 · p. 9
  • With beam search, the method beats baseline on long and short responses and also improves over baseline using OPERA decoding.Against baseline with OPERA, gains are 1.8% and 1.1% for CS and CI on long responses, and 1.6% and 0.9% on short responses.Source: §5.3 · p. 8

5.4 MME

  • On MME perception subtasks, the method surpasses baseline by 76.33 and shows gains over VCD across subtasks, including an example of 24.00 improvement versus VCD's 2.00.The subtasks examine object-level and attribute-level hallucinations.Source: §5.4 · pp. 8–9, Table 3 · p. 9

5.5 GPT4V-Aided Evaluation

  • On LLaVA-Bench, GPT-4 rates the method above OPERA in detailness and correctness for open-ended responses.The evaluation covers conversation, detailed description, and complex reasoning questions.Source: §5.5 · p. 9, Table 4 · p. 9

6 Conclusion and Limitations

The paper proposes Concentric Causal Attention (CCA) to reduce the impact of RoPE long-term decay on object hallucination, with experimental support across multiple evaluation benchmarks. Its scope is image-text inputs, and the discussion notes potential inherited risks in models built from pretrained components.

6 Conclusion and Limitations

  • CCA reduces relative distances between visual and instruction tokens to mitigate RoPE long-term decay and object hallucination, with experiments supporting the positional-alignment strategy across multiple benchmarks.The authors identify CCA as a positional alignment strategy for LVLMs.Source: §6 · p. 9
  • The study addresses image-text inputs only; positional alignment for other modalities, including audio and video, is left for future work.Source: §6 · p. 9

A Broader Impact

  • CCA-trained models may inherit risks from their pretrained visual encoders and language models, including malicious-input handling, hallucination, and potential biases.The authors say they mitigate these issues following other LVLMs.Source: A Broader Impact · p. 15

B RoPE in LLaMA

  • Figure 4 illustrates position-dependent rotations of query or key features; RoPE operates in all LLaMA self-attention layers, unlike ViT’s one-time absolute positional encoding after patch embedding.Source: B RoPE in LLaMA · p. 15, Figure 4 · p. 15

C Implementation Details

The appendix points to implementation details for the main paper’s Figure 1 and Figure 2 results, covering data, model architecture, and training.

C Implementation Details

C.1 Pilot Experiment

The pilot experiment compares raster-scan and reversed raster-scan models trained with a shared schedule, and tests object-existence questions using cropped objects pasted at varied image locations.

C.1 Pilot Experiment

  • The raster-scan baseline Fb and reversal-raster-scan model Fr use two-stage training, with 20K LoRA steps in stage two due to resource limits and the same hyperparameters as the 665K-step full schedule.The two models follow the training procedure in.Source: §C.1 · p. 15
  • For motivation experiments, 3,000 COCO VAL 2014 annotations are sampled and each annotated object is cropped and pasted onto a gray template at 144 positions arranged as a 12-by-12 grid.Source: §C.1 · p. 15
  • The synthetic testing workflow places each cropped object on every spatial position of a template, producing object-existence questions across diverse locations.Source: Figure 5 · p. 16

C.2 Information Flow

The information-flow analysis aggregates attention from visual tokens to instruction tokens using a sample of POPE adversarial image-query pairs and LLaVA-1.5-7B.

C.2 Information Flow

  • Using 3,000 POPE adversarial image-query pairs and LLaVA-1.5-7B, the analysis averages first-decoder-layer self-attentions across heads and images to obtain the quantitative results.Figure 6 provides pseudocode for computing visual-to-instruction information flow.Source: §C.2 · p. 16, Figure 6 · p. 16

D.1 Comparison over Multiple-Choice Benchmarks

The positional alignment approach achieves gains over the baseline across six multimodal perception benchmarks. The authors report these results as evidence of its potential to enhance general visual perception in LVLMs.

D.1 Comparison over Multiple-Choice Benchmarks

  • Evaluation covers SEED-Bench, ScienceQA, GQA, Vizwiz, MMBench, and MMStar.MMStar is described as requiring visual grounding; SEED-Bench includes image and video questions across 12 dimensions.Source: §D.1 · p. 16
  • The positional alignment approach achieves non-negligible gains over LLaVA across all six perception benchmarks.The comparison is presented in Table 5.Source: §D.1 · p. 16, Table 5 · p. 17
  • The reported performance gains require no additional computation during training and suggest potential to enhance general visual perception in LVLMs.Source: §D.1 · p. 17

D.2 Qualitative Comparison

The supplied material covers qualitative comparisons of baseline and trained-model errors in object hallucination, OCR, and numerical prediction, alongside a checklist statement that the paper involved no crowdsourcing or human-subjects research.

D.2 Qualitative Comparison

  • The baseline more often hallucinates objects such as bathtubs, sinks, knives, and cups, while the trained model reduces these errors.These examples appear in open-ended generation comparisons.Source: §D.2 · p. 17, Figure 7 · p. 17, Figure 8 · p. 17
  • In LLaVA-Bench cases, CCA avoids a hat hallucination and outperforms LLaVA on optical character recognition and numerical prediction.The paper illustrates these outcomes in individual case studies.Source: §D.2 · p. 17, Figure 9 · p. 18, Figure 10 · p. 18

E Computation Resources

NeurIPS Paper Checklist

  • The checklist says results are reproducible, but open access to code and data is not yet provided; code and trained models are planned for release upon acceptance.The reproducibility checklist points to Sec. 5.1 and Appendix C.1 for training details.Source: NeurIPS Paper Checklist · p. 20
  • The checklist reports that experimental error bars are omitted because LVLM training is computationally expensive.Source: NeurIPS Paper Checklist · p. 21
  • The checklist marks the item NA, indicating no crowdsourcing or human-subjects research.It notes that local IRB approval may be required for human-subjects research, and asks authors to follow institutional and NeurIPS ethics guidance and preserve anonymity in initial submissions.Source: NeurIPS Paper Checklist · p. 24

Related papers

Loading 2410.15926v1…