acceptodds
Under review as a conference paper at ICLR 2027

Mitigating Multimodal LLMs Hallucinations via Relevance Propagation at Inference Time

Abstract

Multimodal large language models (MLLMs) achieve strong performance on vision- and audio-language tasks, yet can generate responses that conflict with the given visual or auditory inputs, a problem known as multimodal hallucinations. Prior work suggests that this occurs when models rely more on textual cues and learned language patterns than on evidence from the perceptual input. To obtain a more direct account of this imbalance, we apply Layer-wise Relevance Propagation (LRP), which attributes predictions to individual input tokens, and use the resulting relevance scores to analyze and mitigate hallucinations. First, we examine whether this imbalance leads to multimodal hallucinations. We find that hallucinations often arise when the model relies less on perceptual inputs, and that changing this reliance affects its predictions. We further leverage LRP and propose a training-free framework that shifts relevance toward perceptual tokens by optimizing key-value representations during decoding, without modifying model parameters or requiring training data. We call this method Learning Inference-time Modality Enhancement (LIME). Despite using no spatial or temporal supervision, LIME concentrates relevance on query-relevant regions. We evaluate LIME across multiple multimodal benchmarks in both vision and audio domains, demonstrating consistent reductions in hallucinations and enhanced grounding while preserving generation quality.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multimodal Learning, Interpretability, Hallucinations

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_LCoKOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes hallucination reduction method in multimodal LLMs including LVLMs and LALMs. The paper argues that the hallucination occurs when the model do not attend enough to the perceptual inputs. This motivates training-free LIME method, in which they optimize key-value representations so that the models rely more on the perceptual inputs rather than textual inputs.

Critical strengths

The method is generalizable across various LLMs including LVLMs and LALMs.

Critical weaknesses

Major Concerns

  1. The claims in lines 48-79 are weakly supported. I cannot find any quantitative results in the main body that link perceptual relevance to hallucination. Moreover, even if supporting results are available, the findings themselves do not appear to offer substantial novelty compared to prior papers. Also, line 204 states that this relation is analyzed in Section 4.3, but Section 4.3 contains no such analysis. I can only find the table about the LIME method itself.

  2. Limited practicability as training-free method: The overhead is too heavy for the gains. The main results use 7 steps, which causes a 9.43x slowdown (Table A.9). Compared to prior methods, the improvement is small to justify this cost. In addition, no video or other long-input benchmark is evaluated. I think the cost in both time and memory should grow quickly with the number of perceptual tokens.

  3. The explanation for why localization emerges without supervision is not convincing. The objective alone should favor a uniform spread over perceptual tokens. Potential explanation is needed.

Minor Concerns

  1. Sections 4.1 and 4.2 claim evaluation on five vision and three audio models, but results for several of them appear only in the appendix. This makes the main results hard to follow.
  2. Numeric precision is inconsistent across tables (xx, xx.x, xx.xx). Please use a uniform format.
  3. The paper states that some baseline methods are "not applicable across all architectures" (lines 302-303), but contrastive decoding methods such as VCD should be applicable to Qwen2.5-VL and Qwen2.5-Omni, and others. Such inaccurate statements raise doubts about whether the comparison is fair.
Questions for authors
  1. The baseline comparisons are mainly on older models (LLaVA-1.5-7B, Qwen-VL-Chat, SALMONN-7B). Does LIME generalize to more recent, stronger models such as Qwen3-VL?
  2. Is there an ablation without the KL term?
  3. How do the slowdown and peak memory scale with sequence length (e.g., long video inputs)?
  4. Is there any quantitative result supporting the claim in line 53 that both patterns are "most pronounced on samples the model gets wrong"?
5 comments
AuthorsOpenReview ↗

Response to Major Concerns by Authors

Thank you for your comments. We appreciate your response and would like to address the points you raised. We also uploaded a revised version of the paper.

Major Concerns:

  • We would like to clarify the evidence linking relevance distribution to hallucination behavior. Section 4.3 (Table 4) examines perceptual utilization and relevance localization, while Appendix C.1 provides complementary behavioral analyses, including per-sample comparisons and controlled interventions on both positive and negative samples (Tables A.9 and A.10).

    First, Table A.9 shows that corrected predictions exhibit larger increases in perceptual utilization (+0.121 vs. +0.031) and relevance localization (+0.147 vs. +0.026) than predictions that remain hallucinated. Second, Table A.10 presents controlled interventions showing that increasing perceptual relevance improves accuracy (83.24% to 90.44%), while suppressing it reduces accuracy (82.65%), particularly when targeting query-relevant regions (79.82%). Random interventions have negligible effect (83.35%). Together, these results provide observational and interventional evidence linking relevance distribution to hallucination behavior. We have clarified this distinction from the modality utilization analysis in Section 4.3 in the revised manuscript.

  • The novelty of our approach lies in using prediction-specific relevance to both characterize and mitigate multimodal hallucinations. While existing methods primarily rely on indirect signals, such as attention weights or output changes under perturbed inputs [1, 2, 3], LIME uses LRP to quantify the contributions of perceptual and textual tokens to model predictions. To the best of our knowledge, we are the first to extend LRP beyond post-hoc explainability by using relevance directly as an inference-time optimization signal to mitigate multimodal hallucinations.

 

  1. We agree that LIME introduces additional computational overhead due to its iterative relevance-guided optimization. To better characterize the accuracy-latency trade-off, we have expanded the efficiency evaluation to include three-step configurations, reporting measured throughput, slowdown, peak memory, and accuracy (Table 5 in the revised manuscript). On LLaVA, three steps achieve 88.65% accuracy at a 3.33× slowdown, compared with 90.27% at 9.43× using seven steps. Similarly, on Qwen2-Audio, three steps achieve 61.00% accuracy at 3.21×, compared with 63.36% at 9.20× using seven steps. We use higher learning rates for the lower-budget configurations, as we believe this enables more effective optimization with fewer steps, while acknowledging the need for hyperparameter selection. While methods such as MemVR incur lower computational costs, LIME allows users to balance accuracy and latency through its optimization budget and demonstrates effectiveness across both vision and audio models.

    Regarding long-input scalability, LIME has not yet been evaluated on long-form inputs such as video, where the computational overhead may increase with sequence length. Improving the efficiency of relevance-guided optimization therefore remains an important direction for future work. We have moved the efficiency analysis and limitations discussion into the main paper, to provide a clearer account of the measured computational costs, configurable trade-offs, and the remaining scalability challenges.

 

  1. We appreciate the reviewer's observation that the relevance objective in Eq. 3, considered independently, favors a uniform distribution over perceptual tokens. This property is explicitly derived in Appendix C.2 (Eq. 10). However, LIME does not optimize relevance scores directly. Instead, it optimizes key-value perturbations through the frozen model, under KL regularization. As shown by our local analysis in Appendix C.2 (Eq. 13), the resulting relevance changes are shaped by the model's sensitivity structure and the KL constraint, rather than directly following the uniform distribution favored by the relevance objective. This provides an explanation for how relevance can become concentrated on query-relevant regions without spatial or temporal supervision. Consistent with this analysis, Section 4.3 demonstrates improved relevance localization across both vision and audio models, approaching the performance of the region-supervised variant (LIME-Sup).

[1] "OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation", Huang el al., CVPR 2024

[2] "Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding", Leng et al., CVPR 2024.

[3] "Reducing Object Hallucination in Large Audio-Language Models via Audio-Aware Decoding", Hsu et al, 2025

AuthorsOpenReview ↗

Response to Minor Concern and Questions by Authors

Minor Concerns:

  1. We agree that the presentation of our experimental results could be clearer. Due to space limitations, we included representative results in the main paper and placed additional evaluations in Appendix B.3. To improve clarity, we have revised the references to these additional results throughout Sections 4.1 and 4.2.

 

  1. We thank the reviewer for pointing out this inconsistency. We have fixed the numerical formatting across all tables in the revised manuscript

 

  1. We thank the reviewer for pointing out this distinction. Our goal was to provide a comprehensive comparison against available training-free methods, including results where compatible and reproducible implementations were available. The limitations regarding baseline applicability primarily concern implementation availability rather than fundamental architectural incompatibility. For example, VCD is conceptually applicable to Qwen2.5-VL and Qwen2.5-Omni, but its publicly released implementation does not directly support these architectures and would require further adaptation. We have clarified this distinction in the revised manuscript and believe that the existing comparisons remain representative of established training-free approaches, providing a meaningful basis for evaluating LIME across different models and benchmarks.

Questions:

  1. To further evaluate the generality of LIME, we have added experiments with Qwen3-VL on the POPE benchmark using both MSCOCO and A-OKVQA (Table A.5). On MSCOCO, LIME improves accuracy from 91.47% to 93.31% and F1 from 91.28% to 93.25%. On A-OKVQA, accuracy improves from 93.88% to 94.78% and F1 from 93.71% to 94.59%. Our original evaluation also included MiniCPM-o 4.5, released in 2026, for which LIME consistently improves accuracy and F1 on both datasets. Together, we believe that these results provide further evidence of LIME's generalization to newer multimodal architectures

 

  1. We have added an ablation (lines 1026-1033) removing the KL term entirely (λ=0\lambda=0), while keeping all other optimization settings unchanged. As reported in the new Table A.7, LIME without KL achieves 88.70% accuracy on LLaVA and 62.10% on Qwen2-Audio, compared with baseline accuracies of 83.49% and 56.19%, respectively. Including KL regularization further improves accuracy to 90.27% and 63.36%. These results show that the relevance objective alone provides substantial improvements, while KL regularization yields additional gains by limiting deviations from the original output distribution.

 

  1. As discussed in our response to Major Concern 2, LIME introduces computational overhead that may increase with input sequence length. While our efficiency analysis in Section 4.5 and Appendix D examines the effects of optimization budget and intervention scope, we have not directly evaluated scalability to long-form inputs such as video. We acknowledge this a in the revised manuscript.

 

  1. As discussed in our response to Major Concern 1, Appendix C.1 provides quantitative evidence linking relevance distribution to hallucination behavior. Table A.9 presents a per-sample analysis showing that corrected predictions exhibit substantially larger increases in perceptual utilization and relevance localization than predictions that remain hallucinated. Complementing this analysis, Table A.10 presents controlled interventions examining how the direction and target of relevance optimization affect prediction accuracy.
AuthorsOpenReview ↗

Official Comment by Authors

Thank you again for your review. We believe that our rebuttal addresses each of the questions and concerns raised in your review. If any points remain unclear or if further clarification would be helpful, we would be happy to provide additional details during the remaining discussion period.

If you feel that your concerns have been adequately addressed, we would be grateful if you would consider updating your score accordingly.

Reviewer_z4L8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LIME mitigates multimodal hallucinations by optimizing key-value perturbations during decoding. This method leverages least-reactive perturbations (LRPs) to enhance the relevance of perceptual inputs and constrains output variations through key-value regularization (KL). Experiments cover visual and audio tasks across multiple model families.

Critical strengths

The method is clearly designed, requires no training, and is applicable to both visual and audio input.

Critical weaknesses
  1. Validation of improvements in input dependency mechanisms remains insufficient. LIME uses the same attribution framework to optimize and evaluate relevance. The intervention experiments in Table A8 only contain positive samples, therefore it is impossible to distinguish whether the performance improvement comes from more accurate utilization of input or from a stronger inclination to answer "yes".

  2. The trade-off between quality and computational cost limits the demonstrated advantages. In Table 2, LLaVA shows an improvement in the illusion metric, but recall decreases from 75.2 to 72.0.

Questions for authors

Do the conclusions of the intervention experiment hold true when both positive and negative samples are included, especially regarding whether the false positive rate decreases?

2 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for their thoughtful feedback and constructive suggestions. We would like to address the concerns raised and provide additional clarifications and experimental evidence below. We also uploaded a revised version of the paper.

Critical Weaknesses + Questions:

1+Q. We would like to clarify the scope of our intervention analysis and provide further evidence addressing potential positive-answer bias. The original analysis in Appendix C.1 (Table A.10 in the revised manuscript) is restricted to positive samples because the supervised variants (LIME-Sup and Reverse-LIME-Sup) require ground-truth object regions, which are undefined for negative samples.

To further investigate this concern, we have extended the analysis to both positive and negative samples, comparing the applicable variants (Baseline, Random, LIME, and Reverse-LIME) under identical evaluation settings. As shown in the new Table A.11, LIME increases the average true-positive rate (TPR) from 76.85% to 87.13%, while reducing the false-positive rate (FPR) from 17.18% to 11.34%. These findings suggest that the improvements are not simply driven by an increased tendency to predict "yes", but reflect better discrimination between present and absent objects. Importantly, although LRP guides the optimization, the intervention outcomes are evaluated using ground-truth accuracy, TPR, and FPR, independently of the attribution scores. This provides additional behavioral evidence for the effectiveness of relevance-guided interventions beyond improvements in relevance metrics alone.

 

  1. We recognize the computational overhead introduced by LIME's iterative optimization. Following Reviewer 1F5D's suggestion, we have moved the efficiency analysis and limitations discussion to the main paper and expanded the evaluation to include three-step configurations. On LLaVA, three steps achieve 88.65% accuracy at a 3.33× slowdown, compared with 90.27% at 9.43× using seven steps, with a similar trend observed for Qwen2-Audio. We use higher learning rates for the reduced-step configurations, as we believe this enables more effective optimization with fewer steps, while acknowledging the need for hyperparameter selection. Although LIME incurs higher computational costs than some competing methods, its optimization budget can be adjusted to balance accuracy and latency according to application requirements.

    Regarding CHAIR recall, we acknowledge the decrease for LLaVA. However, recall improves for Qwen-VL (64.3% to 68.7%) and Qwen2.5-VL (55.1% to 56.5%), while hallucination rates decrease across all three models (Table 2). To further investigate whether these reductions result from shorter captions, we have added a caption-length analysis across the full CHAIR evaluation set (Table A.8, lines 1035-1042). Across all three models, caption lengths remain largely unchanged under LIME, with slightly longer captions for LLaVA (80.05 vs. 78.42) and slightly shorter captions for Qwen-VL (89.26 vs. 91.73) and Qwen2.5-VL (92.49 vs. 94.84). These findings suggest that the hallucination reductions are not simply attributable to shorter captions. Nevertheless, the decreased recall for LLaVA indicates a remaining trade-off in object coverage, which we have clarified in the revised manuscript.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you again for your review. We believe that our rebuttal addresses each of the questions and concerns raised in your review. If any points remain unclear or if further clarification would be helpful, we would be happy to provide additional details during the remaining discussion period.

If you feel that your concerns have been adequately addressed, we would be grateful if you would consider updating your score accordingly.

Reviewer_1F5DOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes LIME, a training-free method for mitigating multimodal hallucinations. At each decoding step, LIME uses Layer-wise Relevance Propagation to optimize additive key–value updates, encouraging greater relevance for perceptual tokens while constraining changes to the output distribution through KL regularization. The model parameters remain frozen, and no spatial or temporal annotations are required. Experiments across vision and audio models demonstrate improvements on several hallucination and understanding benchmarks. Additional analyses investigate perceptual utilization, relevance localization, and the effects of interventions on the relevance objective.

Critical strengths
  • A compelling idea addressing an important problem. The paper is strongly motivated by the gap between fluent multimodal generation and grounding in perceptual evidence. I found the use of relevance attribution as an active inference-time guidance signal particularly interesting and inspiring. The connection between analyzing modality utilization and directly optimizing it gives the approach a coherent rationale. This is the main reason for my positive assessment: the paper offers a promising research direction beyond the individual benchmark improvements.

  • Broad empirical support and useful investigation of the method's behavior. The evaluation spans vision and audio, multiple model families, and complementary tasks. Beyond aggregate performance, the paper examines key/value updates, regularization, localization without region annotations, efficiency ,and interventions on the target and direction of relevance. These experiments strengthen the contribution by investigating when and how the method helps, rather than relying exclusively on benchmark gains.

Critical weaknesses
  • The substantial computational overhead is central to interpreting the results and needs to appear in the main paper. Table A.9 reports slowdowns of 9.43x for LLaVA and 9.2x for Qwen2-Audio at seven optimization steps, the default configuration in Table A.1. The comparison with competitive methods is especially relevant: in the reported LLaVA setting, LIME achieves 90.3% accuracy versus MemVR's 88.5%, but throughput falls from 2.90 to 0.32 tokens/s. At five steps, LIME achieves 88.8% at 0.45 tokens/s. I appreciate that the authors quantify and discuss this limitation, including mentioning it in the conclusion. However, relegating the quantitative comparison to the final appendix table gives insufficient prominence to a major and big practical limitation. I consider it essential to include the core efficiency analysis in the main paper and discuss the performance gains alongside their cost. This does not negate the scientific contribution, but it materially affects its practical significance.

  • The claim of preserving generation quality needs a more explicit account of the trade-off between faithfulness and coverage. In Table 2, LIME improves LLaVA's CHAIR scores but lowers recall from 75.2 to 72.0, compared with 80.8 for MemVR. This raises the question of how much of the hallucination reduction accompanies more conservative descriptions. Recall improves for the other two models in the table, so the concern is not a universal loss of coverage. Nevertheless, the broad claim of preserving generation quality needs qualification or additional evidence concerning informativeness and response length. This matters because reducing unsupported content while retaining useful information is central to the method's intended benefit.

Questions for authors
  • Efficiency and practical significance. Could the authors bring the core accuracy-cost comparison into the main paper? Are measured throughput and memory results available for the three-step and restricted-layer variants proposed as lower-cost alternatives? These results would help assess whether the method offers a practically attractive operating point relative to competitive baselines.
  • Faithfulness versus coverage. For the LLaVA CHAIR results, how do response length and the number of correctly mentioned objects change under LIME? Is there evidence that the hallucination reduction persists when comparing outputs with similar coverage or length?
  • Clarification of relevance conservation. Section 4.3 states that LIME can only redistribute relevance rather than change its total amount. Conservation across layers for a fixed prediction does not itself establish conservation across different KV perturbations, and KL regularization does not fix the absolute value of the explained logit. Is total input relevance empirically approximately constant during optimization, or is an additional normalization or assumption used?

My current rating is 3, reflecting an overall positive assessment of the paper and the concerns outlined above. I would consider raising my rating to 4 if the authors convincingly address these concerns and commit to incorporating an expanded efficiency analysis into the main paper, including measured costs for lower-budget configurations and a clearer comparison of performance gains against computational overhead.

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for their positive assessment and constructive suggestions. Below, we address each critical weakness together with its corresponding question, providing additional experimental evidence and clarifications where relevant. We also uploaded a revised version of the paper.

Critical Weaknesses and Questions:

1+Q1. We acknowledge the reviewer's concern regarding computational efficiency and agree that LIME introduces substantial overhead due to its iterative relevance-guided optimization. To better characterize the accuracy-latency trade-off, we have expanded the efficiency evaluation to include three-step configurations, reporting measured throughput, slowdown, peak memory, and accuracy. We have also moved the core efficiency analysis and limitations discussion into the main paper (Section 4.5). On LLaVA, three optimization steps achieve 88.65% accuracy at a 3.33× slowdown, compared with 90.27% at 9.43× using seven steps. Similarly, on Qwen2-Audio, three steps achieve 61.00% accuracy at 3.21×, compared with 63.36% at 9.20× using seven steps. Peak memory remains approximately unchanged across optimization budgets (Table 5). We use higher learning rates for the lower-budget configurations, as we believe this enables more effective optimization with fewer steps, while acknowledging the need for hyperparameter selection.

We recognize that methods such as MemVR remain considerably more computationally efficient. Nevertheless, LIME provides a configurable accuracy–latency trade-off through its optimization budget and demonstrates effectiveness across both vision and audio models. Additional layer-scope ablations are presented in Appendix D, showing that restricted interventions can retain meaningful accuracy improvements, although their computational costs have not been measured separately. We have clarified these trade-offs and remaining limitations in the revised manuscript.

 

2+Q2. We appreciate the reviewer's concern that the reduction in CHAIR hallucination rates might result from shorter or more conservative captions, particularly given the decrease in recall for LLaVA. To investigate this, we have added a caption-length analysis across the full CHAIR evaluation set (lines 1035-1042 Table A.8).

For LLaVA, LIME produces slightly longer captions than standard decoding (80.05 ± 16.30 vs. 78.42 ± 16.10), while reducing CHAIR$_S$ from 52.0% to 42.7% and CHAIR$_I$ from 15.8% to 13.0%. Caption lengths also remain comparable for Qwen-VL-Chat and Qwen2.5-VL-7B-Instruct, where recall improves under LIME. These findings suggest that the hallucination reductions are not simply attributable to shorter captions. While comparable caption lengths do not guarantee identical object coverage, the results provide additional evidence that LIME reduces hallucinations without substantially shortening generated captions. We have clarified this distinction in the revised manuscript.

 

Q3. We thank the reviewer for highlighting this distinction. As discussed in the AttnLRP [1] work, exact relevance conservation is not guaranteed for certain transformer operations, such as softmax, where bias-related terms may absorb part of the relevance. Importantly, LIME does not rely on strict conservation. At each optimization iteration, the key-value updates are applied before computing the output logit and its corresponding relevance scores. The relevance objective encourages shifting the relative allocation of relevance toward perceptual tokens, while the KL term constrains changes to the output distribution. Our method therefore requires neither exact conservation during backward relevance propagation through the model nor a constant total relevance across optimization iterations. We have clarified this distinction in the revised manuscript.

 

[1] "AttnLRP: Attention-Aware Layer-Wise Relevance Propagation for Transformers". Achtibat el at., ICML 2024.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you again for your review. We believe that our rebuttal addresses each of the questions and concerns raised in your review. If any points remain unclear or if further clarification would be helpful, we would be happy to provide additional details during the remaining discussion period.

If you feel that your concerns have been adequately addressed, we would be grateful if you would consider updating your score accordingly.

I thank the authors for the detailed response and for addressing my main concerns. I will keep my current positive score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Remarks by Authors

Dear Reviewers and Area Chair,

As the discussion period draws to a close, we would like to thank all reviewers again for their thoughtful feedback and engagement. We have uploaded an updated manuscript incorporating additional experiments, analyses, and clarifications, and we believe we have addressed the questions and concerns raised. We summarize the key updates below:

  • Relevance–hallucination relationship: We extended our LRP-based analyses to Qwen3-VL, including perceptual utilization, relevance localization, per-sample analysis, and controlled interventions (Tables 4, A.9, and A.10). We also examined both positive and negative samples to address potential answer bias (Table A.11).

  • Evaluation on recent models: We added VCD and ICD comparisons on Qwen3-VL and MiniCPM-o 4.5 (Table A.5), and on Qwen2.5-VL for CHAIR and HallusionBench (Tables 2 and A.6). LIME consistently outperforms these methods in the new comparisons.

  • Methodological analysis: We expanded the KL-regularization ablation to three models (Table A.7) and clarified the theoretical explanation for relevance localization and the role of relevance conservation.

  • Computational efficiency: We expanded the evaluation to include different optimization budgets, reporting measured throughput, slowdown, memory usage, and accuracy. We also moved the efficiency analysis and practical limitations into the main paper (Section 4.5).

  • Caption coverage: We added a caption-length analysis across 3 MLLMs (Table A.8) to investigate whether the reductions in hallucination rates could be attributed to shorter generated captions, and clarified the observed recall trade-off.

We hope these additions and our detailed responses have addressed the reviewers' concerns and provided a clearer picture of LIME's contributions. We appreciate your time and consideration.

Sincerely,
The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.