acceptodds
Withdrawn from ICLR 2027

EGAA: Evidence-Grounded Answer Alignment for Mitigating Cross-Modal Hallucination in Omni-Modal LLMs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-modal hallucination remains a major challenge for omni-modal large language models. Existing benchmarks such as AVHBench and CMM measure these failures at the task level, but their standard scores provide limited information about the specific error within a generated response. We therefore examine generated answers at two complementary levels. Cross-Modal Source Misattribution (CSM), Absence Overgeneralization (AO), and Unsupported Certainty (UC) are identified at the statement level, while Wrong-Modality Answer (WMA) is assessed at the response level. To support both analysis and training, we augment AVHBench and CMM separately with visual and audio evidence annotations, supported statements, grounded answers, and error-targeted negative answers for these four error types. Based on these annotations, we propose Evidence-Grounded Answer Alignment (EGAA), a parameter-efficient post-training method that jointly learns from modality-specific evidence, supported statements, grounded answers, benchmark-format targets, and constructed preference pairs. We evaluate EGAA through bidirectional transfer between AVHBench and CMM. With Qwen2.5-Omni-7B, training on AVHBench improves CMM accuracy from 81.1% to 88.8%, while training on CMM improves AVHBench overall accuracy from 76.1% to 82.2%. These results show that explicitly connecting generated statements to multimodal evidence can reduce cross-modal hallucination across benchmarks. The code and configuration files used to reproduce our results are available at https://anonymous.4open.science/r/EGAA-DF71/, and the corresponding evidence annotations and dataset records are available at https://osf.io/j7x89/overview?view_only=2ff822a2fee8471f9b6384c25d00ecf6.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Hallucination, Multimodal Large Language Models, Audio-Visual Understanding, Preference Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_sTrJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EGAA studies hallucination in omni-modal LLMs. It identifies four error types and augments AVHBench and CMM with modality-specific evidence (CSM, AO, UC, and WMA), supported statements, grounded answers, and error-targeted negative answers. The proposed method jointly optimizes three formats and a DPO training through LoRA. Experiments across 3 backbones show improved accuracy in bidirectional transfer between the two benchmarks.

Critical strengths

This paper establishes a link between 4 types of cross-modal errors and specific preference pairs, thereby constructing an effective framework for converting diagnostic results into supervisory signals for training. Bidirectional transfer experiments conducted on AVHBench and CMM demonstrate that EGAA possesses generalization capabilities surpassing the original benchmarks.

Critical weaknesses

Existing evidence has does not fully establish the additional value and broader applicability of EGAA. Comparative analyses have employed mismatched settings, evaluations have focused on binary judgments and brief explanations, and the characterization of error coverage across the four categories remains insufficient, particularly regarding different models, datasets, and more complex cross-modal scenarios.

Questions for authors
  1. Appendix B.1 provides comparisons under different experimental settings. Could the authors compare EGAA with a representative method such as OmniDPO or MoD-DPO using the same backbone, source data, and evaluation, with comparable training budgets while retaining each method’s supervision design? This would help clarify EGAA’s value over existing approaches.
  2. Although caption generation are included during training, the evaluation focused primarily on binary judgments (Yes/No) and brief descriptions. Can EGAA also effectively mitigate hallucinations in open-ended audiovisual generation tasks, such as dense caption generation, non-binary question answering, speaker attribution and so on?
  3. These four error categories are identified based on the responses of Qwen2.5-Omni-7B on AVHBench. How are the coverage of these error categories verified for other models and datasets? Are there any cross-modal errors that do not fall into the aforementioned categories or are difficult to classify? (For example, a cross-modal temporal relationship error, that a door closes at 2-second mark of the video while a bell rings at 5-second mark of the audio, the model correctly indentifies both the door closing and the ringing bell but answer “the bell rang before the door closed”) How might such errors impact the evaluation process and the construction of preference data?
2 comments
AuthorsOpenReview ↗

Response to Reviewer sTrJ

We thank the reviewer for recognizing that EGAA turns diagnostic categories into training signals. We have uploaded a revised paper that contains all new results.

1. Controlled comparison with OmniDPO and MoD-DPO (Q1)

Neither method has released official code, and the anonymous repository announced by OmniDPO is no longer accessible. This is why the paper originally listed their published numbers separately. We implemented both from their papers, kept each method's own objective, and trained them on Qwen2.5-Omni-7B with the same grounded answers and negatives as EGAA, the same LoRA configuration, and the same schedule (Sec. 4.2, Table 2, Appendix C). EGAA outperforms both on every metric in both directions. CMM accuracy is 88.8% against 86.3% for OmniDPO and 87.5% for MoD-DPO, and AVHBench overall accuracy is 82.2% against 79.0% and 80.8%.

The largest gap is on AVHBench matching (81.0% against 74.6% and 78.1%), which asks whether the two streams belong together, a question that no CMM probe asks. Label-only fine-tuning drops matching to 52.7%. The two preference methods keep most of it because they learn from our error-targeted pairs, and EGAA keeps the most because its Evidence Trace and Grounded Answer make the model describe each stream and state what follows from their combination. The comparison therefore shows both the value of the pairs and what the evidence targets add on top of them.

2. Open-ended generation and non-binary questions (Q2)

The paper originally reported only the judgment tasks of AVHBench, the same tasks that OmniDPO and MoD-DPO report, to keep the results comparable. We now evaluate all three backbones on captioning and on multiple-choice questions, using models that saw neither task format in training (Sec. 4.4).

On AVHBench captioning, scored with its official metrics on the CMM-trained models, EGAA gives the highest METEOR and CIDEr for all three backbones. GAVIE-A, which penalizes content absent from the reference, rises for MiniCPM-o-4.5 and VideoLLaMA2.1-7B-AV and falls for Qwen2.5-Omni-7B, whose EGAA captions are 24.8 characters longer on average. Among the 398 captions scored below Base, the judge cites content absent from the reference in 73.9%, and a manual check of 100 of them found that almost all add correct detail. For a beehive video, for example, EGAA adds that the man speaks Russian and mentions a queen bee, which the reference omits (Table 4, Appendix D.3).

For non-binary question answering and speaker attribution, AV-SpeakerBench poses 3,212 four-way questions on who speaks, what is said, and when. EGAA raises overall accuracy by 1.7 to 2.3 points for every backbone and training direction, and speaker detection, which links a voice to a visible person, by 1.4 to 4.5 points for every model (Table 5, Appendix D.2). SVHalluc adds three-way questions, discussed under Q3.

3. Coverage and the temporal example (Q3)

The four types were defined by the authors from the relation among the requested modality, the per-modality evidence, and the claim, not from the outputs of one model on one benchmark (Sec. 3.1). CMM played no part in this process and serves as a check. Miss occurs in only 0.33% of CMM questions, against 21.79% with any error (Table 3).

In your example, both events are perceived and assigned to the correct modality, and the error lies in their temporal relation. In evaluation, such a statement is counted as Miss and as unsupported, so it lowers the scores as it should. For preference data, the evidence text already records when each event occurs, so negatives that swap the order of events across the two streams can be built in the same way as the four current types.

SVHalluc (CVPR 2026) tests such errors on narrated video, and we evaluated all three backbones on it (Sec. 4.4, Appendix D.2). On its semantic tasks, which ask whether narrated speech is visible and thus test the counterpart of CSM, EGAA improves every backbone by 2.4 to 9.1 points. On its temporal tasks, which ask whether a narrated event happens before, during, or after the speech, EGAA also improves Qwen2.5-Omni-7B and VideoLLaMA2.1-7B-AV in both training directions and MiniCPM-o-4.5 when trained on CMM. The evidence text states when each event occurs, and the Evidence Trace asks the model to reproduce the evidence of each stream, so training already exposes the model to the timing of both streams. EGAA targets the errors that dominate the explanations of current models, while Miss and statement-level support keep the remaining errors visible.

We would be grateful to know whether any point remains unclear or whether other analyses would be helpful. We hope these responses lead you to view the paper more favorably.

Thank you for the supplementary comparative experiments and broader evaluations by the author. However, the preference data uniformly using EGAA still makes it difficult to fully assess the advantages of its data structure over the existing supervised design. The hallucination indicators in open-ended generation also show inconsistent performance. Therefore, I keep the original score.

Reviewer_Uqj8OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies cross-modal hallucination in omni-modal LLMs through four error categories and augments AVHBench and CMM with modality-specific evidence, supported statements, grounded answers, and targeted negative answers. It proposes EGAA, a LoRA-based post-training method combining evidence-trace and answer supervision with DPO on the constructed preference pairs. Bidirectional transfer experiments between the two benchmarks show improved overall task accuracy across three backbones. Fine-grained evaluation of Qwen2.5-Omni-7B additionally reports reduced error incidence in explanatory answers.

Critical strengths
  1. The four error categories provide a useful framework for characterizing cross-modal hallucination. The paper defines CSM, AO, UC, and WMA to distinguish errors in evidence-source attribution, the scope of absence claims, evidential certainty, and the modality addressed by an answer. This categorization makes distinct failure modes explicit, helping clarify how cross-modal hallucination manifests and providing a basis for designing targeted supervision. It constitutes a useful contribution to the formulation of the problem.

  2. Bidirectional cross-dataset experiments demonstrate the method’s practical effectiveness. EGAA improves overall task accuracy across all three backbones in both AVHBench-to-CMM and CMM-to-AVHBench transfer. For Qwen2.5-Omni-7B, accuracy increases from 81.1% to 88.8% on CMM and from 76.1% to 82.2% on AVHBench. These gains across backbones and transfer directions provide empirical support for the effectiveness and cross-dataset transferability of the overall training recipe.

Critical weaknesses
  1. The effects of preference optimization on the four error categories lack fine-grained ablation analysis. The w/o Pref. ablation in Table 2 reports only task-level metrics, while Table 4 does not show how removing the preference component affects the incidence of CSM, AO, UC, and WMA. Given that prior methods already discussed in the paper, such as MoD-DPO [1], also use preference optimization to mitigate cross-modal hallucinations, EGAA should clarify which errors its preference component reduces and by how much. The results support the overall training recipe, but leave this central component’s contribution to category-level error reduction insufficiently explained.

  2. The incremental contribution of evidence-grounded annotation is insufficiently distinguished from prior work. The paper presents evidence-grounded annotation as a major contribution, but using structured audio-visual information to construct question–answer or reasoning supervision has relevant precedents. For example, OmniVideo-100K [2] organizes audio-visual content into structured scripts and extracts cross-modal clue chains as the basis for constructing question–answer data. AVDC [3] generates visual, auditory, and joint descriptions, which are subsequently used to construct questions, answers, and reasoning traces. In these pipelines, the scripts, clue chains, and modality descriptions serve as intermediate evidence supporting the generated answers or reasoning. These works therefore provide relevant precedents for extracting and organizing modality-specific evidence to construct answer or reasoning supervision, directly relating to EGAA’s evidence-grounded annotation pipeline for generating supported statements and grounded answers. However, the paper does not discuss these works or compare its annotation pipeline with theirs, leaving its incremental contribution over existing approaches unclear. The authors should clarify which designs follow established ideas and which constitute substantive advances. Since evidence-grounded annotation is presented as a major contribution, the insufficient positioning relative to prior work makes its novelty and incremental value difficult to assess.

[1] Ashutosh Chaubey, Jiacheng Pang, and Mohammad Soleymani. MoD-DPO: Towards Mitigating Cross-modal Hallucinations in Omni LLMs using Modality Decoupled Preference Optimization. arXiv preprint arXiv:2603.03192, 2026.

[2] Xinyue Cai, Chaoyou Fu, Yi-Fan Zhang, Ran He, and Caifeng Shan. OmniVideo-100K: A Dataset for Audio-Visual Reasoning through Structured Scripts and Evidence Chains. arXiv preprint arXiv:2606.14702, 2026.

[3] Kaiying Yan, Luoyi Sun, Xiao Zhou, and Weidi Xie. Empowering Long-form Omni-modal Understanding with Robust Audio Perception. arXiv preprint arXiv:2607.10299, 2026.

Questions for authors
  1. What specific design advances and contributions does EGAA’s evidence-grounded annotation offer over OmniVideo-100K [2] and AVDC [3]?

  2. Could the authors compare the four error rates between full EGAA and w/o Pref. under Table 4’s evaluation protocol to clarify the preference component’s effects on each error category?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Uqj8

We thank the reviewer for viewing the four categories as a contribution to the formulation of the problem. We have uploaded a revised paper that contains all new results.

1. Effect of the preference component on each error type (W1, Q2)

We evaluated EGAA w/o Pref. under the protocol of the original Table 4, which is now part of Table 3 (Sec. 4.3). The supervised evidence targets alone already remove most AO errors (CMM: 7.21% to 1.17%, AVHBench: 6.21% to 3.25%), and adding the preference pairs changes AO and CSM very little. The pairs act mainly on UC (CMM: 10.24% to 9.21%) and WMA (AVHBench: 0.49% to 0.21%), and Miss also drops on AVHBench (5.20% to 4.04%). Questions with any error fall from 13.42% to 12.17% on CMM and from 22.21% to 20.24% on AVHBench. The evidence targets alone also raise the share of supported statements from 82.1% to 85.3% on CMM and from 65.9% to 72.5% on AVHBench, and the pairs raise it further to 85.6% and 73.8%.

The pattern follows from how the pairs are built. In UC and WMA, a negative differs from the grounded answer by only a few words, such as a dropped hedge or a switch to the other stream. A positive target shows what a well-scoped answer looks like, while a contrastive objective tells the model which small change makes it unsupported. The pairs also act on the final Yes/No, which the statement counts do not capture. Without them, the model answers No too readily, and w/o Pref. has the lowest perception accuracy and the highest hallucination resistance among the EGAA variants on every backbone (Table 2). Adding the pairs raises perception accuracy on CMM from 82.4% to 88.2% for Qwen2.5-Omni-7B, and overall accuracy rises with it.

This breakdown also shows how EGAA differs from input-side methods such as MoD-DPO, which asks the answer to stay the same when the irrelevant stream is corrupted and to change when the relevant one is. This targets which stream an answer depends on, closest to WMA and CSM. Scope and certainty errors (AO and UC) occur even when the model relies on the right stream, and these are what our evidence targets and pairs reduce. Trained on the same pairs, EGAA also outperforms MoD-DPO on every metric (Sec. 4.2, Table 2).

2. Difference from OmniVideo-100K and AVDC (W2, Q1)

Thank you for these references, which Sec. 2.2 now discusses. Describing each modality separately is shared with both works, and atomic statements follow FActScore and FaithScore. The differences are as follows.

The intermediate text of the two pipelines is descriptive. OmniVideo-100K builds an entity-anchored script with a summary, an entity list, and time-stamped visual and audio descriptions for each segment. AVDC writes visual, audio, and joint captions for each segment and then summarizes them into global captions. Neither records whether a concept was directly observed, only possibly present, or simply not observed in a given stream, and observation and cross-modal inference are mixed in the same narrative. Without this distinction, AO and UC cannot even be defined, since “not heard” cannot be told apart from “absent” and “possible” cannot be told apart from “certain”. A sound wrongly attributed to a visible object in a joint caption also becomes part of the training data without being marked. EGAA assigns each concept a status in each modality from an annotator that has access to that modality only, distinguishes an unavailable modality from a concept that is not observed in it (as needed for CMM), and writes cross-modal conclusions as supported statements that cannot overwrite the observations.

Checking every target against human labels. Both pipelines generate new questions and answers, so their targets have no independent reference, and only about 38% of the OmniVideo-Test drafts passed manual checks of factual correctness, cross-modal dependency, and answer uniqueness. AVDC filters its captions with a model-based quality score. EGAA annotates existing AVHBench and CMM items, keeps their media, questions, labels, and official metrics, and compares every grounded answer with the human-verified label. The two agree on 88.2% of items, all 1,041 disagreements were reviewed by humans, and only 2.0% of items were rewritten (Sec. 3.2, Appendix A.3).

Both pipelines turn their descriptions into correct answers for training, and AVDC perturbs the input with mismatched audio. EGAA instead uses its statuses to write wrong answers that each misstate one aspect of the evidence (Sec. 3.2). A bark that is heard while no dog is seen yields a CSM negative claiming that a dog is visible, and a dog that is seen but silent yields an AO negative claiming that there is no dog.

We would be grateful to know whether any point remains unclear or whether other analyses would be helpful. We hope these responses lead you to view the paper more favorably.

Reviewer_s7fhOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines the audio-visual QA answers based on 4 cross-modal error types: Cross-Modal Source Misattribution (CSM), Absence Overgeneralization (AO), Unsupported Certainty (UC), and Wrong-Modality Answer (WMA). Based on these, it augments AVHBench and CMM with this evidence-ground annotations including modality-specific evidence, a description, supported statements, a grounded answer, and error-targeted negative answers. Then, it proposes Evidence-Grounded Answer Alignment (EGAA), which trains AVLM to learn modality-specific evidence trace, grounded answer, B-F answers. It newly proposes the supervised loss term for each term, weighted summed with the original direct preference optimiztaion objective. The experiment results shows that proposed EGAA outperforms the base model at diverse audio-visual LLMs.

Critical strengths
  • This paper systemically divides the error types that can occur at audio-visual hallucinations into 4 types: Cross-Modal Source Misattribution (CSM), Absence Overgeneralization (AO), Unsupported Certainty (UC), and Wrong-Modality Answer (WMA). Those error types shows complementary relations, and becomes a good criteria to analyze the current audio-visual scene and model's responses, and to check which evidence is required in the audio and video.
  • Based on 4 cross-modal error types, it proposes a Evidence-Grounded Answer Alignment (EGAA) method, which make the model to learn the 4 evidence signals: evidence trace, a grounded answer, and benchmark-format answer, and a preference pair. For each term, they propose new supervised loss terms as a negative log likelihood, and use those terms with weighted sum with original direct preference optimization objective.
Critical weaknesses
  • Coverage of Four Error types. It seems that proposed 4 error types may not fully cover some other cases of audio-visual answering. For example, the models may miss the objects in some modalities at all, but such type are not considered as error types. Additionally, the criteria between "observed" and "possible" evidence in Unsupported Certainty (UC) is somewhat subjective, making its classification criteria unclear. Validating that such four error types are sufficient, and providing a criteria for UC's observed and possible would mprove the clarity of the paper.
  • Reliance over original AV-LLMs for error type and dataset construction. The authors said that the initial error type analysis conducted by Qwen3.7-Plus (line 170). And the question parsing is conducted as VL-LLMs, and visual evidence and audio evidence is extracted from another audio-visual models, and those may lead to additional error accumulation for final answers. It seems correcting such errors from the model's output should be done for each step that models have produced.
Questions for authors
  • I assume whether unsupported certainty (UC) may be vague to say it is wrong. For example, the cat closing its eye, and the question may be 'did you see cat sleeping in the video?'. In this case, assuming the cat is sleeping seems reasonable, but at the other side it can be one example of unsupported certainty. Please provide some example for UC if there is any criteria for deciding it.
  • I also think current form of supervision can be also achieved in GRPO form also, rather than the DPO. It seems like giving a supervision loss + GRPO loss may further enhance its evidence-ground ability through RL method.
1 comment
AuthorsOpenReview ↗

Response to Reviewer s7fh

We thank the reviewer for recognizing the complementary relations among the four error types. We have uploaded a revised paper with all new results.

1. Coverage of the four types

No classification can cover every error a model makes, and we do not claim that the four types do. In our evaluation, erroneous statements outside the four types are counted as Miss, so the Miss rate is an upper bound on what the four types leave out (Sec. 3.1). For base Qwen2.5-Omni-7B, 21.79% of CMM questions contain an error but only 0.33% contain a Miss error, and on AVHBench the figures are 24.31% and 4.60% (Sec. 4.3, Table 3). The four types therefore cover most of the errors we observe, and the remaining errors stay visible in the evaluation.

A missed object is classified by what the answer says. Stating that a visible object is absent keeps the right modality, scope, and certainty but misdescribes observable content, so it is Miss. Concluding from a missing sound that the object is absent from the whole scene ("there is no dog" because no bark is heard) is AO.

2. Criteria for observed, possible, and UC (Q1)

Observed means identifiable evidence in the stream, and possible means a plausible interpretation that the stream alone cannot resolve. A statement is UC only when an interpretation is possible within one modality, fusion finds nothing in the other modality that resolves it, and the answer still states it as certain. UC concerns how strongly a claim is stated, so a reasonable inference is acceptable when it is presented as one (Sec. 3.1).

In your example, the video shows a cat lying still with its eyes closed and the audio has nothing relevant. "Yes, the cat appears to be sleeping" is supported, while "Yes, the cat is sleeping" is UC although its Yes/No is the same. If snoring were audible, the two streams would jointly support the conclusion and the second answer would also be correct.

Appendix A.5 adds two real UC negatives. In the human audit of the judge, 56 of the 69 statements labeled UC (81.2%) match the annotators, and the other 13 are errors that they labeled Miss.

3. Error accumulation in the pipeline

Qwen3.7-Plus only grouped erroneous responses into recurring patterns, and the authors defined the four types from the relation among the requested modality, the per-modality evidence, and the claim (Sec. 3.1). A single omni-modal annotator would avoid passing errors between stages but would itself suffer from the hallucinations we study. We therefore isolate the visual and audio annotators (Appendix A.2), and fusion cannot change their statuses. Errors that remain are caught by comparing every grounded answer with the human-verified benchmark label. The two agree on 88.2% of items, all 1,041 disagreements were reviewed by two annotators, and only 173 items (2.0%) needed their evidence and answer rewritten. Official scores use the original labels (Sec. 3.2, Appendix A.3).

4. GRPO (Q2)

We ran your suggestion on Qwen2.5-Omni-7B in both directions, replacing DPO with on-policy GRPO while keeping the three supervised losses. Each question gets 8 samples, and the reward is r=rans+λ rstatusr = r_{\text{ans}} + \lambda\ r_{\text{status}}, where rstatusr_{\text{status}} is the F1 over (modality, concept, status) triples (Sec. 4.2, Table 2, Appendix C). It reaches 87.9% on CMM and 80.9% on AVHBench, well above Base (81.1%, 76.1%) and close to EGAA (88.8%, 82.2%), but nearly identical to EGAA w/o Pref.. We see three reasons why GRPO adds little.

First, GRPO can only score what it samples, and its responses contain the status lines and a Yes/No but no explanation (Appendix C). The errors the pairs target, UC in particular, arise in the explanation, such as stating "the cat is sleeping" as a fact when the frames only make it possible. GRPO gets no signal on them, while a UC pair contrasts such a claim with its hedged version. The gap is largest on AVHBench matching, which asks whether the two streams are related, so asserting an unestablished relation directly yields a wrong answer (74.5% vs. 81.0% for EGAA).

Second, advantages are computed within each group of 8 samples. When all samples give the same Yes/No, as on questions the model answers confidently, ransr_{\text{ans}} yields no advantage, and rstatusr_{\text{status}} rewards statuses the Trace loss already teaches. Targeted errors are also rare in samples (WMA: 0.38% of base answers on CMM), so few groups contain the contrast each DPO pair supplies by construction.

Third, for a controlled comparison we kept the one-epoch budget of the other methods and did not tune λ\lambda or the group size, so this result reflects one configuration of your idea.

We would be grateful to know whether any point remains unclear or whether other analyses would be helpful. We hope these responses lead you to view the paper more favorably.

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

Related papers

Discussion (0)

Sign in to comment.