- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies cross-modal hallucinations in audio-visual large language models. It distinguishes three dimensions that should be evaluated separately: whether the final answer is correct, whether the modality judgments are correct, and whether those judgments remain stable when the other modality changes.
The authors introduce AV-TRACE, which contains 7,344 manually verified test examples. It evaluates correctness and cross-modal stability through explicit modality judgments and audio-visual quartet comparisons. Each quartet crosses audio and video containing or not containing the target, allowing one modality to be held fixed while the other is replaced to detect erroneous judgment flips.
The authors also propose FQFT, which combines a joint reward for answers and modality judgments with quartet regularization within GRPO. The experiments cover five base models and several mitigation methods. FQFT training and ablations use Qwen2.5-Omni-7B, with evaluations on AV-TRACE and an adapted AVHBench* showing higher joint accuracy and lower hallucination rates.
Critical strengths
- The evaluation separates reliability issues that can otherwise be conflated. AV-TRACE examines both final answers and explicit modality judgments. By holding the same clip fixed and replacing the other modality, its controlled comparisons identify erroneous flips caused by cross-modal interference. This separates cases in which an answer is correct but the judgment is wrong from cases in which a judgment is correct on one input but unstable across inputs, providing a useful complement to answer accuracy.
- The training method directly addresses the identified problems and has controlled experimental support. FQFT incorporates judgment correctness and cross-modal stability into training. Starting from the same SFT initialization, it improves joint accuracy from 81.43% to 84.39% and reduces the hallucination rate from 14.23% to 10.32% relative to vanilla GRPO. Results on the adapted AVHBench* also provide some evidence of transfer, supporting the practical value of the method.
Critical weaknesses
- Model coverage remains insufficient. Table 2 evaluates five base models but does not cover the Gemini family or more recent Qwen Omni models. I suggest adding Gemini 3.1 Pro Preview, Gemini 3.8 Flash, Qwen3.5-Omni-Plus, and Qwen3.8-Omni-Flash, restricted to versions publicly available at the time of submission. This would test whether the observed hallucination behavior and main conclusions also hold for stronger models and different model families, strengthening their generality.
- FQFT has not been validated across backbones. FQFT training and ablations are conducted on Qwen2.5-Omni-7B, leaving it unclear whether the gains depend on this backbone. At least one additional trainable backbone should be used to reproduce the comparison among SFT, vanilla GRPO, and FQFT to support the transferability of the method.
- The ablations do not fully explain the roles of the key components. Table 3 separates the contributions of the joint reward and quartet regularization, but target discrimination, non-target modality suppression, interaction suppression, and mean anchoring within the regularizer are still evaluated as a package. Their individual contributions and necessity therefore remain unclear. Moreover, full FQFT achieves higher joint accuracy, whereas quartet-only has a slightly lower Flip score. Targeted ablations, particularly of interaction suppression and mean anchoring, would help explain their different effects on correctness and stability.
- Evidence of external generalization remains limited. External validation is concentrated on AVHBench* after adapting its tasks and output protocol, leaving it unclear how much the gains depend on the current structured evaluation setting. This dataset also lacks quartets with shared clips, so it cannot validate transfer of improvements in Flip. Evaluations under the original task protocol, on other datasets, or on more natural in-the-wild audio-visual data would strengthen the evidence for effectiveness under distribution changes, especially the generalization of cross-modal stability.
Questions for authors
Quality of automated prescreening and annotation cost. The 1.4% error rate reported in the appendix comes from an audit after human verification and adjudication. Is there a comparison between the automated prescreening results and the final human labels, particularly the false acceptance and false rejection rates for audio and vision separately, and the amount of human work actually saved? Given that visual prescreening primarily relies on a single large model judging sampled frames, have the authors evaluated strategies such as identifying supporting frames or time spans, cross-checking with different models, or routing uncertain cases for review? Can these strategies reduce the proportion requiring human review while maintaining annotation quality?
6 comments
We thank the reviewer for this valuable comment.
W2: Validating the transferability of SFT, vanilla GRPO, and FQFT on an additional trainable backbone.
We reproduced the full comparison on MiniCPM-o 2.6, from a distinct model family, using the same data, objectives, and pipeline as Qwen2.5-Omni-7B (Table C).
Table C. Results of MiniCPM-o 2.6 on AV-TRACE (%).
| Model / method | AA ↑ | JA ↑ | JAA ↑ | Hall. ↓ | Flip ↓ |
|---|---|---|---|---|---|
| MiniCPM-o 2.6 | 61.92 | 47.22 | 45.61 | 42.33 | 60.63 |
| +SFT | 77.48 | 75.78 | 75.74 | 17.76 | 11.09 |
| +SFT+Vanilla GRPO | 81.39 | 78.30 | 78.25 | 12.65 | 7.29 |
| +SFT+FQFT (ours) | 83.32 | 80.74 | 80.66 | 10.84 | 7.24 |
As shown in Table C, on MiniCPM-o 2.6, the ordering across variants matches Qwen2.5-Omni-7B: AA, JA, JAA, and Hall. improve consistently from SFT → vanilla GRPO → FQFT (JAA from 45.61% → 75.74% → 78.25% → 80.66%; Hall. from 42.33% → 17.76% → 12.65% → 10.84%), both policy-training variants cut SFT's Flip rate (11.09%) by roughly one-third (to 7.29% and 7.24%), and FQFT raises audio detection JAA from 28.60% in the base model to 69.10% (and visual detection from 63.70% to 90.70%) while cutting Flip in the dominant V→A direction from 72.76% to 9.85% (and A→V Flip from 48.49% to 4.64%). Relative to vanilla GRPO from the same SFT checkpoint, FQFT improves JAA by +2.41 percentage points and lowers Hall. by −1.81 percentage points on MiniCPM-o 2.6, closely matching the gain pattern on Qwen2.5-Omni-7B and confirming transferability across backbones.
We thank the reviewer for this valuable comment.
W3: Component-level ablations inside the quartet regularizer and comparison between quartet-only and full FQFT on accuracy and Flip.
Under the same setup as Table 3 in the paper, we present the module-level comparison of joint reward and quartet regularization (Table D(a)) separately from the ablation of the individual loss terms inside the quartet regularizer (Table D(b)).
Table D(a). Module-level ablation of joint reward and quartet regularization on Qwen2.5-Omni-7B (%).
| Training | Joint reward | Quartet. | AA ↑ | JA ↑ | JAA ↑ | Hall. ↓ | Flip ↓ |
|---|---|---|---|---|---|---|---|
| SFT | ✗ | ✗ | 77.56 | 76.32 | 76.29 | 20.49 | 4.08 |
| Vanilla GRPO | ✗ | ✗ | 83.03 | 81.44 | 81.43 | 14.23 | 2.40 |
| Joint reward | ✓ | ✗ | 82.64 | 81.39 | 81.37 | 14.58 | 2.28 |
| Quartet | ✗ | ✓ | 85.19 | 83.30 | 83.29 | 10.88 | 1.83 |
| FQFT (ours) | ✓ | ✓ | 85.75 | 84.47 | 84.39 | 10.32 | 1.95 |
As shown in Table D(a), adding Joint reward alone only scores single-input responses without quartet regularization to suppress cross-modal interference in representation space. Under cross-modal conflict, completions in a GRPO group still tend to make the same modality-judgment errors (and when answer-correct completions share the same judgment outcome, group normalization cancels the bonus; Appendix E). Thus, Joint reward alone slightly narrows the AA-to-JAA gap (1.60 to 1.27 points) without improving JAA (81.37% vs. 81.43% under Vanilla GRPO) or Hall. (14.58% vs. 14.23%). Conversely, Quartet alone constrains the four-state log-probability margins, cutting Hall. from 14.23% to 10.88% and Flip from 2.40% to 1.83%, but its policy reward still scores only the final answer rather than generated modality judgments, leaving JA and JAA (83.30% and 83.29%) 1.90 points behind AA (85.19%). When combined in full FQFT, the quartet regularizer suppresses cross-modal interference internally while Joint reward directly rewards responses where the final answer and all modality judgments are simultaneously correct. Relative to Quartet, FQFT raises JA by 1.17 points to 84.47% and JAA by 1.10 points to 84.39%, narrows the AA-to-JAA gap to 1.36 points, and lowers Hall. to 10.32%. Because cross-state stability is governed by the quartet regularizer rather than the single-input reward, Quartet and full FQFT achieve similarly low Flip rates (1.83% and 1.95%, well below 2.40% under Vanilla GRPO and 2.28% under Joint reward). Across all 1,207 eligible V→A pairs in the test set, FQFT flips only 47 times (3.89%) and 0 times on A→V, so the 0.12-point difference reflects only a few examples.
Table D(b). Loss-term ablation inside the quartet regularizer on Qwen2.5-Omni-7B (%).
| Training | Target disc. | Mean anchor. | Non-target supp. | Interact. supp. | AA ↑ | JA ↑ | JAA ↑ | Hall. ↓ | Flip ↓ |
|---|---|---|---|---|---|---|---|---|---|
| FQFT w/o cross-state coupling | ✓ | ✓ | ✗ | ✗ | 82.00 | 80.81 | 80.80 | 15.21 | 2.94 |
| FQFT w/o non-target supp. | ✓ | ✓ | ✗ | ✓ | 82.11 | 80.85 | 80.84 | 15.07 | 2.97 |
| FQFT w/o interaction | ✓ | ✓ | ✓ | ✗ | 82.04 | 80.97 | 80.97 | 15.15 | 2.37 |
| FQFT w/o mean anchor. | ✓ | ✗ | ✓ | ✓ | 82.35 | 81.29 | 81.29 | 14.91 | 2.16 |
| FQFT w/o target disc. | ✗ | ✓ | ✓ | ✓ | 81.88 | 80.75 | 80.74 | 15.29 | 2.41 |
| FQFT (ours) | ✓ | ✓ | ✓ | ✓ | 85.75 | 84.47 | 84.39 | 10.32 | 1.95 |
To further examine each loss term inside the quartet regularizer, Table D(b) removes both cross-state terms together (w/o cross-state coupling) or removes each of the four terms individually. Notably, all five ablated variants perform slightly below Vanilla GRPO in Table D(a) (80.74%–81.29% vs. 81.43% JAA; 14.91%–15.29% vs. 14.23% Hall.): when both cross-state terms are removed (w/o cross-state coupling, 80.80% JAA, 15.21% Hall., 2.94% Flip), target discrimination only enlarges the average present-minus-absent gap, which the model can widen via co-occurrence cues on concordant states at the cost of more conflict-state errors. Between the two cross-state terms, non-target modality suppression limits average interference from the other modality across present and absent states (removing it raises Flip to 2.97%, the worst among all variants), while interaction suppression removes opposite interference across the two states (removing it leaves JAA at 80.97% and Hall. at 15.15%). Meanwhile, because the first three terms only constrain relative differences among the four states, removing mean anchoring leaves the four-state mean unconstrained (81.29% JAA, 14.91% Hall.), and removing target discrimination fails to separate positive and negative targets, giving the lowest JAA (80.74%) and highest Hall. (15.29%). Only when all four terms are enabled together (full FQFT) does the model reach 84.39% JAA, 10.32% Hall., and 1.95% Flip, confirming that all four loss terms are indispensable.
We thank the reviewer for the constructive comments and questions. We answer each of them below. Tables A–F are provided in our responses, and we will update all new results and tables into the revised PDF of the paper.
Using the versions publicly available at the submission deadline and the same setup as Paper Table 2, we evaluated Gemini 3.8 Flash, Gemini 3.1 Pro Preview, Qwen3.8-Omni-Flash, and Qwen3.5-Omni-Plus on the full AV-TRACE test set (overall results in Table A; breakdowns by task and interference direction in Table B).
Table A. Results on AV-TRACE (%).
Table B. Results on AV-TRACE by task and interference direction (%).
All main conclusions of the paper hold on these stronger models: although the four closed-source models reach substantially higher JAA than the five open-weight base models, their Hall. remains 10.81%–14.92%, their Flip remains 3.71%–8.16%, and errors remain heavily concentrated in the visual-to-audio (V→A) direction—detection hallucination under audio interference (A→V) is only 0.20%–1.40%, whereas visual interference (V→A) still induces 22.90%–34.20% hallucination and 6.41%–15.82% Flip.
Moreover, high accuracy still does not imply stability: Gemini 3.1 Pro Preview and Qwen3.5-Omni-Plus achieve higher overall JAA than 7B FQFT via near-ceiling visual detection, yet suffer roughly 2× the Flip rate; Qwen3.8-Omni-Flash is within 0.4 JAA points of vanilla GRPO on Qwen2.5-Omni-7B (81.43% JAA, 2.40% Flip) but flips more than 3× as often. In comparison, 7B Qwen2.5-Omni-7B + FQFT achieves the lowest Hall. (10.32%), lowest Flip (1.95%), and highest audio detection JAA (78.20%) across the entire table, outperforming Gemini 3.8 Flash and Qwen3.8-Omni-Flash on all five metrics.