Mitigating Object Hallucination via Concentric Causal Attention

Thank you for your meticulous reading and giving credit to our novelty and analysis of Rotary Position Encoding (RoPE) and LVLM hallucination. We appreciate you pointing out some additional references, which we would include to make our research more complete. Please find our responses as follows.

W1: Clarification on RoPE.

A: Thanks for your kind advice. We will include a short background information on RoPE to indicate that, just as absolute and learnable position encoding, RoPE is also a type of position encoding, which is adopted by existing Large Language Models like LLaMA and inherited by most open-sourced LVLMs. We will also include a detailed guidance in Appendix on definition of RoPE and how it is involved in LLaMA architecture. For now we refer to [55] and lines 127-149 in our manuscript, where we present a mathematical form of RoPE and its long-term decay property for reference. For further clarification of Figure 2 in our manuscript, please find a new illustration in Figure 4 of uploaded pdf. For Figure 3, please refer to Figure 1 (right) of uploaded pdf for a new illustration, where (a) to (f) corresponds to the first eight rows in Figure 3.d of our manuscript.

W2: Alternative position encoding.

A: Thanks for pointing this out. We kindly point out that RoPE is the default position encoding in LLaMA. Simply replacing RoPE with other position encodings is technically viable but deviates from LLaMA training scheme. We train a model with learnable position encoding and a model with relative position encoding [C] on this. Due to resource limitation, we train both models on 20k instruction data (instead of 665k) while train a new CCA model that follows the same setup for fair comparison. According to our results below, LVLM (learnable) performs much worse than LVLM (cca), while training of LVLM (relative) do not converge and not viable.

W3: Comparison with training-based methods.

A: We kindly remind that we compared our method with LLaVA-RLHF in Table 1 (for POPE) and Table 2 (for CHAIR) of our manuscript, a training method that mitigates object hallucination in LVLMs. We should point out that we compare our 7B model results against those from LLaVA-RLHF-13B model and we still stands out, indicating effectiveness of proposed method. As suggested, we compare CCA with a recent training method SeVa [B], which applies DPO training on LLaVA-1.5-7B. Overall our CCA model undergoes less training compute than SeVa, as SeVa applies another DPO training stage beyond LLaVA-1.5 pre-training and fine-tuning stage, while our training strictly follows LLaVA-1.5. POPE results are listed below,

Despite less overall training compute, we highlight that our LLaVA-1.5-7B-CCA still outperforms SeVa-7B on 6 POPE evaluations. For more challenging adversarial evaluations , LLaVA-1.5-7B-CCA surpasses SeVa-7B by large margins consistently. We also compare our results with SeVa on LVLM benchmarks, where results of SeVa are taken directly from their paper. We show that our model surpasses SeVa-7B in most cases.

Q1: Figure 1.

Thank you for your question. We would like to clarify that higher aggregation values at the beginning and end of Figure 1.b (manuscript) is not related to image content as we averaged over 3k COCO images to get visualization results (as detailed in Appendix B.1). This may be attributed to removal of RoPE, which leads to out-of-training-distribution and breaks pre-trained LLaMA position encoding.

Q2: Applicability to Q-formers.

Thanks for sharing this concern. Both InstructBLIP [14] and Qwen-VL [4] adopt RoPE in their language models. Nevertheless, our 2-D positional alignment strategy is designed for spatial-locality-preserved LVLMs [42,41,5], where the full image embedding from vision encoder is kept. Applying CCA to models like InstructBLIP and Qwen-VL is technically viable but not our design intentions. We will include this as a limitation of our method instead.

Q3: Empirical evidence on our concentric design.

Yes. The concentric design is motivated by that more objects are statistically located at the centre. Please find a statistical evidence collected from COCO and GQA in Figure 3 of uploaded pdf. Please refer to our response for Q1 where we clarify results from Figure 1.b (manuscript).

L1: RoPE.

Thanks for pointing this out. We admit that our method cannot apply to models with position encodings other than RoPE. We will include this in Limitation part. However, we think it is not a major drawback as most existing LVLMs use LLaMA as language backbone, where RoPE is applied as position encoding scheme.

Thank you for your valuable insights. Please see our responses to your questions below.

W1: Alternative scanning method.

A:: We justify the design of our method by providing new comparative studies for different position encoding schemes and alternative scanning methods. We first compare CCA with learnable position encoding. Due to resource limitations, we train all positional alignment approaches on 558k pre-training data and only 20k instruction data, including a new CCA model with same setup for fair comparison. We conduct evaluation on POPE and CHAIR benchmarks and the results are shown in tables below. The resulting models with learnable position encoding perform worse than our design.

We also compare alternative scanning designs. We reverse the scanning order of original CCA and start from visual tokens at the center of image and end at periphery, illustrated in Figure 2 (right) of uploaded pdf. As kindly suggested in your question 3, we implemented diagonal scan and provide evaluation results for new scanning designs in a table below. Our original CCA scanning method demonstrated overall show better performance over other design choices. Based on these ablation studies, CCA is chosen as our final method for hallucination mitigation.

For presentation of concentric causal masking, please find new illustrations in Figure 1 of uploaded pdf.

W2: Performance on CHAIR.

A: Thanks for pointing this out. We kindly remind that our trained model achieved the best results (lowest CHAIR scores) when applying CCA alone, as shown in the table below (summarized from manuscript Table 2). We point out that the LLaVA-RLHF model for benchmarking uses stronger Vicuna-13B as its language backbone and involves additional direct preference optimization stage in their training, whereas our model undergoes only two training stages (pre-training and supervised fine-tuning). Despite smaller 7B language backbone and less training compute, our model still outperforms the LLaVA-RLHF-13B model on most metrics.

We also point out that our method benefits POPE as well (manuscript Table 1), indicating good compatibility of our method on both open-ended generation and yes-no tasks. Moreover, we highlight that our method also benefits general perception tasks, where approaches that exclusively address object hallucination cannot always bring performance gain. Please refer to Table 5 in Appendix C.2, where we show that our trained model LLaVA-1.5-7B-CCA surpass LLaVA-1.5-7B over multiple LVLM benchmarks consistently.

Q1: Clarification on concentric causal masking.

A: In Figure 3 (d), we use different colours to highlight visual tokens with different positions. For a 2-D organization of visual tokens with shape of 6x6, our CCA leads to 3 positions in visual tokens. Please refer to our new illustrations in Figure 1 of uploaded pdf, where query tokens, key tokens and masked tokens (tokens not involved in self-attention computation) are highlighted. Our CCA follows the same causal modelling rule in LLaMA Figure 1, where query tokens with larger position values attend to key tokens with smaller or equal position values. The difference is that we use a 2-D positional organization for visual tokens, which is a novel and effective attempt among existing LVLM hallucination studies.

Q2: Aggregated correct responses of concentric causal attention.

A: Thanks for mentioning this. We visualize aggregated correct responses in Figure 5 of uploaded pdf with the proposed concentric causal attention. The resulting distribution is largely different from that from raster scan in Figure 2.a and reverse raster scan in Figure 2.b of our manuscript, showing 2-D and symmetrical distribution, which aligns with our concentric causal design.

Q3: Alternative position encodings and scanning.

A: Please refer to our response in W1.

I appreciate the authors' additional experimental results and clarifications. However, there appears to be inconsistent performance across datasets, such as diagonal-lora outperforming the proposed CCA/CCA-lora in MSCOCO and on the adv part of GQA. The authors have suggested that the performance drop of CCA-r supports the assumption that most image content is concentrated in central regions, aligning with their proposed design. Based on this suggestion, how can we interpret the better performance of the diagonal approach? A more in-depth discussion of these inconsistencies would provide greater insights into the effectiveness and the limitation of the proposed CCA strategy. Nonetheless, I believe the author have provided in-depth answers to many asked questions; I have raised my scores to 5.

Thanks for your detailed and insightful suggestions. Please find our responses as follows.

W1-a: Pretraining setup.

A: Thanks for mentioning this concern. We would share that it is a typo in line 227 where we claim we use CC-595K dataset [42] for pre-training stage. In fact, our pre-training experiments follow LLaVA 1.5 [41] and use a 558K dataset for pre-training. The provided baseline results in Table 1 of our manuscript is for LLaVA 1.5, which are sourced from VCD paper [30]. We will release our model and source code for the community to reproduce our results.

W1-b: Manuscript Table 3 and Table 4.

A: Thanks for pointing this out. Please find new quantitative results of Table 3 and 4 below. We use their official codes and get these results.

Q1: Removal of RoPE from LLaVA.

A: Yes, removing RoPE largely diverges from LLaMA pre-training and leads to nonsense outputs. Our earlier studies showed that LVLMs without RoPE will no longer follow instructions. Take POPE [37] questions as an example. The LLaVA-v1.5-7B pretrained without RoPE fails to answer yes or no as illustrated below.

USER: Is there a scissors in the image?

LLaVA-1.5-7B: No.

LLaVA-1.5-7B w/o RoPE: the bear.\n\n\n\n\n\n\n\n

Since the LLaVA-1.5-7B w/o RoPE will output neither yes nor no, the quantitative accuracy on POPE should be 0.00. Though the model w/o RoPE in Figure 1.b of our manuscript generates non-sense text outputs, the information flows from visual to text tokens are more evenly distributed. This highlights the long-term decay in RoPE in Figure 1.c, the root cause of information aggregating at image tokens that are closer to text tokens.

Q2: Clarifications on concentric causal masking.

Thanks for pointing this out. Please find a new illustration in Figure 1 of uploaded pdf that clarifies the proposed concentric causal masking, where query tokens, key tokens attended by query tokens and key tokens not attended by query tokens (tokens not involved in self-attention computation) are colored. Consistent with our manuscript Figure 3, we take 6x6 visual token organization as an example. The design follows the same causal modeling rule of LLaVA (presented in Figure 1 (left) of uploaded pdf), where query tokens with larger position values attend to key tokens with smaller or equal position values. The first seven rows of manuscript Figure 3 (d) share the same attention masks, corresponding to Figure 1 (right) (a) to (g) of uploaded pdf, where the position indices of key tokens attended by query tokens are exactly the same.

The responses address my concerns and the provided illustration figures are great. I am curious about the attention distribution of model trained with CCA and considering raise my score.

Thanks for your detailed and thorough suggestions. Please find our replies as below.

W1: CCA and OPERA [23].

A: We ground our design on analysis of information flow in LLaVA model. This shares commonalities with OPERA which analyzes information flow in LVLM autoregressive decoding. Thank you for pointing out and we would highlight this in introduction of our manuscript as suggested. Different from OPERA that discovers co-occurrences of aggregation pattern and object hallucination, our CCA explores relations between Rotary Position Encoding and object hallucination. We further refer to quantitative results in Table 2 and 4 of our manuscript, where proposed CCA surpasses OPERA in CHAIR and LLaVA-Bench evaluations.

W2: Figure 2 experiments.

A: It is a valid concern to negate the impact of imbalanced object distribution in different image regions. We kindly let you know that we already addressed this in the paper (lines 167-169) by cropping objects and pasting them on blank images (initialized with ImageNet mean pixel values) at different spatial positions to create synthesized images. Please find a new illustration for this in Figure 4 of uploaded pdf, where 16 (4 by 4) pasting options are given. In our experiments we use 144 (12 by 12) pasting options. Figure 2 of our manuscript is obtained by testing models on these synthesized images which have even distribution of objects at different regions.

W3: Data statistics and our concentric design.

A: Yes, our concentric design assumes that most image contents are located around the centric image region. We validate this assumption from two perspectives.

Firstly, we perform statistical analysis on large amount of natural images (82,081 images and 604,907 annotations from COCO train 2014, and 10,696 images and 174,304 annotations from GQA). Specifically, we count total number of objects in 9 spatial locations (top_left, top_mid, top_right, mid_left, center, mid_right, bottom_left, bottom_mid, bottom_right, respectively). The statistical results in Figure 3 of uploaded pdf show that more objects are located in centre, which aligns with our design.

Secondly, we point out that for our model, positions start from periphery of 2-D visual tokens and ends in centre. For comparison, we implement another LVLM (which we name as CCA-r), where positions start from centre of 2-D visual tokens and ends in periphery. Please find an illustration of CCA-r in Figure 2 of uploaded pdf, where left refers to our CCA method and right refers to CCA-r. Query tokens, key tokens for self-attention calculations are highlighted with colors.

The tables below provides quantitative experiments on POPE and CHAIR, showing that changing from CCA to CCA-r positional alignment causes performance drop in most evaluations. For GQA popular evaluation, accuracy drops from 88.40 to 83.43. For CHAIR evaluation, c_s drops from 43.0 to 50.0. These validate the assumption that most image contents are located around centric image regions, which aligns with our design.

W4: More evaluations on LVLM benchmarks.

A: Please refer to Table 5 in Appendix C.2 where we included more evaluations on LVLM benchmarks. As suggested, we also add TextVQA [A] for your reference. We also compare general perception capabilities on these benchmarks against two hallucination-mitigating methods [30, B], where SeVa [B] is a recent method that explores unsupervised preference alignment in LVLMs. We point out that SeVa trains their models with another Direct-Preference-Optimization stage on top of LLaVA-1.5 [41] whereas our CCA does not involve any new training stages and strictly follows LLaVA-1.5 training scheme. Our model LLaVA-1.5-7B-CCA outperforms SeVa-7B and VCD on most LVLM benchmarks.