acceptodds
Desk rejected at ICLR 2027

ACHER: Availability-Conditioned Hierarchical Evidence Reasoning for Multimodal Fake News Detection

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal fake news detection (MFND) identifies misinformation by jointly exploiting text, images, and their cross-modal relations, a task made increasingly hard as generative AI produces more complex cross-modal fakes. Existing methods extract veracity-discriminative signals via cross-modal alignment, consistency modeling, or feature fusion, yet typically assume sufficient corresponding evidence. In practice, low consistency either reflects conflict among existing evidence or insufficient evidence that leaves the relation unverifiable; confusing the two invites unfounded judgments. To address this, we propose the Availability-Conditioned Hierarchical Evidence Reasoner (ACHER), which conditions evidence reasoning on evidence availability: bidirectional sparse localization with null-hypothesis competition estimates it, relation reasoning proceeds only when evidence is available, and each direction yields compatible, conflicting, and unverified states. From these states, ACHER derives four complementary evidence mechanisms with state gating for joint veracity prediction. Experiments on Weibo, FineFake, and AMG show that ACHER consistently outperforms representative methods, confirming its effectiveness and generalization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Learning, Fake News Detection, Cross-Modal Reasoning

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

This paper proposes ACHER for multimodal fake news detection, motivated by the distinction between conflicting cross-modal evidence and insufficient counterpart evidence. It estimates evidence availability through bidirectional sparse localization with null-hypothesis competition, and combines availability with conditional conflict estimates to form compatible, conflicting, and unverified state probabilities in each direction. These states are further organized into four state-gated evidence mechanisms for final prediction. Experiments on Weibo, FineFake, and AMG show competitive or improved performance over the reported baselines, while ablations and interventions provide evidence consistent with the intended distinction between evidence availability and conditional relations.

Critical strengths
  1. Multi-faceted mechanistic validation: The paper goes beyond standard classification metrics by probing module behavior through targeted evidence deletion, availability-controlled counterpart replacement, and mechanism knockouts across different regimes. While these interventions do not fully prove the semantic meaning of the internal states, they provide much stronger empirical grounding for the model's design than mere accuracy gains or t-SNE visualizations.
  2. Consistent ablation support for the proposed structure: Removing availability modeling, bidirectional reasoning, state gating, or any of the four mechanisms consistently degrades performance on both Weibo and FineFake. This provides solid empirical backing that the proposed components have practical predictive value, rather than being purely conceptual additions.
Critical weaknesses
  1. Lack of independent validation for core latent variables: The concepts of "availability" and "conditional conflict" are learned primarily through the final classification loss and the model's own structural objectives, lacking independent annotations to verify they truly represent "evidence sufficiency" and "semantic conflict." While targeted deletion shows the model is sensitive to its own localized candidates, it does not rule out that these states are merely task-specific gating signals. Since the paper's main contribution hinges on the semantic interpretation of these states, this validation gap weakens the core claim.
  2. Insufficient isolation of the symmetric-antisymmetric decomposition's contribution: The performance drop when replacing this decomposition with direct concatenation is significant. However, the paper does not clarify if this ablation variant maintains the exact same parameter count, mechanism branches, state gating, and auxiliary training objectives. Without this, it is difficult to attribute the performance gap solely to the decomposition method itself, rather than confounding structural or optimization changes.
Questions for authors
  1. Do you have any independent annotations or strictly controlled examples (decoupled from the model's own outputs) that can verify "availability" and "conditional conflict" truly correspond to evidence sufficiency and semantic conflict, rather than just acting as generic classification gating signals?
  2. For the "direct concatenation" ablation, does the variant maintain the exact same parameter scale, mechanism branches, state gating, and auxiliary training objectives as the full model? Please specify exactly what was changed in this variant so we can determine if the performance drop is genuinely attributable to the decomposition method itself.
1 comment
AuthorsOpenReview ↗

Semantic Validation and Controlled Ablation of Structured Reasoning

We thank the reviewer for the constructive feedback. We provide complementary semantic validation and controlled-ablation evidence to clarify the roles of ACHER's evidence states and the contribution of structured bidirectional reasoning.

Q1. Semantic validation of Availability and Conditional Conflict

Answer: We thank the reviewer for requesting independent semantic validation. We provide complementary evidence from human annotations, controlled interventions, and external visual-entailment evaluation.

  1. Human-annotated validation. On 342 AMG samples with five annotated misinformation attribution categories, five-fold out-of-fold probes using frozen ACHER representations show that adding Conditional Conflict reduces attribution log loss by 0.0422, with adjusted p≈0.045p\approx0.045. This demonstrates attribution-relevant information beyond binary veracity prediction.
  2. External semantic validation. Using SNLI-VE's human-annotated entailment, neutral, and contradiction labels as additional semantic supervision, the learned Availability and Conditional Conflict scores achieve AUROCs of 0.7916 (Entailment/Contradiction vs. Neutral) and 0.8376 (Contradiction vs. Entailment), respectively, versus 0.4976 and 0.5009 under shuffled-label supervision. Within-image macro ordering reaches 80.42% and 84.39%, with image-level bootstrap 95% confidence intervals of 79.68%–81.16% and 83.58%–85.19%. Lexically matched within-image comparisons further achieve 82.98% and 80.49%, providing complementary evidence of learned semantic discrimination under reduced lexical variation.
  3. Availability intervention. In Table 3, at 20% deletion, targeted counterpart removal produces Availability reductions 27.8 and 7.3 times larger than matched random removal on Weibo, and 15.8 and 5.4 times larger on FineFake across the two directions. This supports its intended role in estimating counterpart evidence sufficiency.
  4. Conditional Conflict intervention. In Table 4, replacing counterpart content while fixing the source and constraining ∣ΔA∣<0.01|\Delta A|<0.01 increases discrepancy DD and Conditional Conflict cc across all four settings. Mean ∣ΔA∣|\Delta A| remains at most 0.0054, with positive paired-bootstrap 95% confidence intervals in Appendix A.9. This demonstrates sensitivity to cross-modal disagreement while estimated Availability remains stable.
  5. State-dependent prediction. ACHER defines S=A(1−c)S=A(1-c), K=AcK=Ac, and U=1−AU=1-A. Removing Unverified reduces Macro-F1 by 2.35 points in the low-availability Weibo cohort versus 0.45 points in the high-availability cohort, supporting its functional relevance when counterpart evidence is scarce. Together, independent annotations, targeted interventions, and external semantic evaluations provide converging evidence that Availability and Conditional Conflict capture distinct semantic and functional aspects of cross-modal evidence beyond generic classification gating.

Q2. Controlled ablation of symmetric–antisymmetric decomposition

Answer: The original direct-concatenation ablation tests the overall structured mechanism design, not decomposition alone. We added a more controlled variant that retains the four experts, state gating, and training objectives to better isolate feature organization.

  1. Original ablation. The direct-concatenation variant in Table 2 retains the pretrained encoders, bidirectional grounding, Availability and Conditional Conflict estimation, and state-conditioned feature weighting. It replaces the structured mechanism representations with direct fusion of bidirectional evidence features and adjusts mechanism-specific supervision accordingly. The parameter count differs by approximately 1.2%. ACC decreases from 94.9% to 93.1% on Weibo and from 80.9% to 78.5% on FineFake. This establishes the contribution of the overall structured mechanism design.
  2. Additional controlled ablation. To assess bidirectional feature organization more directly, we further evaluate direct concatenation while retaining all four mechanism experts, evidence-state gating, auxiliary classification, and Structural Semantic Loss. On FineFake, ACC decreases from 80.90% to 79.70%, while Macro-F1 decreases from 80.57% to 79.29%. This result shows that structured bidirectional organization remains beneficial when the mechanism branches and training objectives are preserved.
  3. Interpretation. The original ablation evaluates the complete structured mechanism design, while the additional control more specifically examines feature organization. Their consistent performance reductions provide complementary evidence for the value of symmetric–antisymmetric reasoning rather than attributing the improvement solely to additional mechanism branches. Overall, the additional controlled ablation further strengthens the evidence for structured bidirectional reasoning.

We will clarify the experimental configurations and incorporate these results into the revised manuscript.

Reviewer_HzoXOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ACHER to distinguish conflict among existing evidence from insufficient evidence through availability-conditioned reasoning, demonstrating its effectiveness and generalization.

Critical strengths

The problem formulation is reasonable and practically meaningful.

Experiments include both performance comparisons and mechanism analyses.

Critical weaknesses

MMRGV reports higher performance on Weibo, but the paper cites it without comparing against it. Please explain why it was excluded.

BERT and CLIP features are not explicitly aligned. Why should the distance between them represent semantic mismatch? I also do not see any such processing in the provided code. More evidence is needed.

The method is complex. Although its design appears reasonable, I am not fully confident in my understanding of how the components contribute to its effectiveness. I therefore assign a confidence score of 2.

Questions for authors

Please address the questions above.

4 comments
AuthorsOpenReview ↗

Clarifications on MMRGV, Cross-Modal Alignment, and Component Contributions

We thank the reviewer for the constructive feedback. We clarify the MMRGV comparison, learned cross-modal discrepancy, and the contributions of ACHER's components.

Q1. Comparison with MMRGV

Answer: We did not include MMRGV in Table 1 because its published Weibo results use a different split and LLM-based resources. We acknowledge its higher reported accuracy without claiming a direct comparison.

  1. Evaluation protocol. MMRGV reports 96.63% ACC on 7,853 Weibo items using a 6:2:2 stratified split, while ACHER follows the official 9,528-item partition. Their reported accuracies thus come from different test sets.
  2. Methodological setting. MMRGV uses Qwen2-VL-72B-generated rationales and verification modules, while ACHER uses the given image-text pair with BERT and CLIP. A matched evaluation is needed for a fair comparison.

Q2. Cross-modal discrepancy and alignment

Answer: ACHER does not compare raw BERT and CLIP distances or assume their pretrained spaces are aligned. It learns a task-conditioned discrepancy through trainable projections, counterpart localization, and relation modeling.

  1. Feature projection and localization. Eq. 1 maps BERT and CLIP features through trainable PT,PVP_T,P_V into a shared 256-dimensional evidence space. Eq. 2 uses learned Wq,WkW_q,W_k in a 128-dimensional matching space. Sparsemax localizes counterparts, and Eq. 3 estimates availability AA through null competition.
  2. Relation computation. Eq. 4 uses direction-specific RdR_d to transform and normalize source and localized counterpart features, combining cosine and weighted absolute discrepancies with a shared calibrator. Joint training with classification and mechanism-level objectives makes DD a task-conditioned discrepancy, not a raw pretrained-feature distance. No separate contrastive-alignment module is used.
  3. Implementation clarification. The previously provided code was a simplified illustration that omitted intermediate operations in Eqs. 1–4. We will provide the complete implementation, including projections, matching, relation computation, and tensor shapes.
  4. Intervention evidence. Table 4 shows that counterpart replacement increases DD and cc across all four dataset-direction settings while mean ∣ΔA∣|\Delta A| remains at most 0.0054. Appendix A.9 reports positive paired-bootstrap 95% confidence intervals for these changes. These results support the model's sensitivity to cross-modal content changes at nearly fixed estimated availability.
  5. External semantic probe. We separately retrained ACHER on SNLI-VE with human-annotated entailment, neutral, and contradiction labels as additional semantic supervision. Availability achieves AUROC 0.7916 for Entailment/Contradiction versus Neutral, and conditional conflict achieves 0.8376 for Contradiction versus Entailment. Shuffled-label supervision yields 0.4976 and 0.5009. These image-level proxy results support semantic learnability under explicit supervision.

Q3. Contributions of individual components

Answer: Ablations quantify the predictive contribution of ACHER's stages, while interventions and evidence-regime analysis clarify their distinct functions.

  1. Structured prediction. Replacing the structured mechanism organization with direct concatenation lowers ACC by 1.8 points on Weibo and 2.4 on FineFake. An additional FineFake control retaining all four experts, gating, and training objectives still lowers ACC from 80.9% to 79.7%. Removing state gating lowers ACC by 0.7 and 1.0 points. The states S=A(1−c)S=A(1-c), K=AcK=Ac, and U=1−AU=1-A organize four complementary evidence mechanisms.
  2. Localization and availability. Removing sparse localization lowers ACC by 0.8 points on both datasets, and removing null competition lowers ACC by 0.6 and 0.4 points on Weibo and FineFake. Targeted deletion affects availability more than matched random deletion.
  3. Conditional relation and bidirectionality. Sharing relation descriptors lowers ACC by 0.6 and 0.4 points. Using only one direction lowers ACC by 0.5–1.1 points.
  4. Evidence-dependent utility. Removing Unverified changes Macro-F1 by −2.35 versus −0.45 points in Weibo's both-low versus both-high availability cohorts. FineFake shows −0.52 versus +0.06 points, indicating greater dependence when evidence is scarce.
  5. Training objectives. Removing auxiliary supervision lowers ACC by 0.6 and 0.8 points, while removing structural semantic loss lowers ACC by 0.6 and 0.4 points on Weibo and FineFake.

Overall, ACHER follows a conditional hierarchy from evidence availability to relation estimation and state-constrained prediction. The ablations and interventions support the complementary roles of these stages.

Thank you for the detailed response. I am satisfied with the clarifications regarding Q1 and Q2.

However, Q3 mainly demonstrates that the components improve performance, without sufficiently explaining why they are effective. Given the complexity of the method, I am still not fully confident in my understanding, which explains my confidence score of 2.

Regarding Reviewer b9hv's comments, I believe using AI to create methodological figures is not inherently problematic. What matters is whether the figure meets academic standards in clarity, visual quality, and technical precision.

If the authors can provide a more convincing explanation of why the components work, or significantly improve the methodological figure to meet ICLR standards, I am willing to raise my rating. I will revisit the discussion in approximately 12 hours.

AuthorsOpenReview ↗

Clarification of ACHER's Design Rationale and Figure 1

We thank the reviewer for accepting our Q1 and Q2 clarifications. We agree that our earlier Q3 response emphasized performance gains rather than underlying mechanisms.

Q3. Why the components work

Answer: ACHER's components cannot be simply merged without weakening the distinctions they preserve. Matching is not evidence sufficiency, discrepancy is not necessarily conflict, directional evidence is not interchangeable, and representing evidence states does not guarantee their use in prediction. The hierarchy separates these tasks while connecting them through explicit evidence conditions.

  1. Why matching and availability must be distinguished. A conventional matching mechanism identifies the most relevant counterpart, but relevance is relative to available candidates. Even when none provides meaningful evidence, it may still select a best match. Using that match for relation reasoning risks treating missing evidence as disagreement. ACHER therefore combines sparse localization with null competition. Localization identifies plausible counterparts, while null competition allows the model to reject insufficient correspondence: Aid=σ(rid−nid),nid=bd+γdH(αid).A_i^d=\sigma(r_i^d-n_i^d),\quad n_i^d=b_d+\gamma_d H(\alpha_i^d).

The entropy-dependent null score accounts for localization uncertainty. Relation reasoning requires not merely a selected counterpart, but sufficient evidence to justify comparison. Table 3 supports this role, as targeted deletion changes estimated availability far more than matched random deletion.

  1. Why availability and conflict cannot be collapsed into one consistency score. Even after identifying a counterpart, the model must distinguish whether it provides enough evidence and whether it contradicts the source. A single consistency score conflates these questions, since low consistency may arise from contradiction or missing correspondence. ACHER instead defines: S=A(1−c),K=Ac,U=1−A.S=A(1-c),\quad K=Ac,\quad U=1-A.

The key is enforcing their dependency. Conflict evidence KK requires availability AA, so a high conditional-conflict score cannot establish contradiction when evidence is insufficient. Such cases contribute primarily to Unverified rather than being forced into Compatible or Conflicting. Table 4 supports this separation. Counterpart replacement increases discrepancy DD and cc across all four settings while mean ∣ΔA∣|\Delta A| remains at most 0.0054. Deletion and replacement probe evidence presence and relational content separately.

  1. Why bidirectional states require structured organization. Separating availability from conflict resolves ambiguity within each direction, but not how the two directions should be interpreted together. Text-to-visual and visual-to-text reasoning ask different questions. Missing visual details and unmentioned image details may therefore yield asymmetric evidence states. Averaging the directions would obscure this imbalance. Concatenation preserves the values but leaves shared evidence and directional differences entangled. ACHER explicitly separates them: K+=(KT+KV)/2,K−=(KT−KV)/2.K^+=(K_T+K_V)/2,\quad K^-=(K_T-K_V)/2.

The symmetric term captures overall conflict intensity, while the antisymmetric term preserves directional imbalance. Analogous unverified terms retain overall insufficiency and directional differences. Broadly shared conflict and conflict concentrated in one direction should remain distinguishable. In an additional FineFake control retaining all four experts, gating, and training objectives, replacing structured organization with direct concatenation reduces ACC from 80.90% to 79.70%.

  1. Why structured evidence still requires state gating. Separating evidence patterns does not ensure that the classifier uses them appropriately. If unrestricted global features remain accessible, the model can bypass the evidence states and rely on content even when its supporting counterpart is insufficient. ACHER therefore makes evidence conditions govern information flow rather than merely providing additional features: zm=gmΦm([lm;g~m;ηm]).z_m=g_m\Phi_m([l_m;\tilde g_m;\eta_m]).

State strengths constrain mechanism inputs and outputs, allowing compatibility, conflict magnitude, directional conflict, and unverified evidence to contribute according to their respective conditions. The Unverified mechanism retains source-side content without constructing unsupported cross-modal discrepancies. Table 5 supports this behavior. Removing Unverified changes Macro-F1 by -2.35 points in Weibo's both-low-availability cohort versus -0.45 in both-high. FineFake shows -0.52 versus +0.06.

Regarding Figure 1

We appreciate the reviewer's view that figures should be judged by clarity, visual quality, and technical precision rather than the tools used. We will refine Figure 1 to make this availability → conditional relation → directional states → state-gated prediction hierarchy clearer.

I have adjusted the score and hope you can finish redrawing the figures before the rebuttal deadline.

Reviewer_b9hvOpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper targets multimodal fake‑news detection. It points out that existing models tend to confuse two scenarios of low cross‑modal consistency: modal evidence conflict and insufficient evidence (unverifiable). To address this issue, the authors propose the ACHER framework. The framework first estimates bidirectional evidence availability and divides observations into three states: compatible, conflicting, and unverified. It then derives four evidence mechanisms via symmetric‑antisymmetric decomposition, and applies state gating to constrain information flow towards the final classifier.

Experiments are conducted on three datasets: Weibo, FineFake, and AMG, with extensive baseline comparisons. Beyond standard ablation studies, the paper designs intervention‑based mechanistic analyses including evidence deletion and evidence replacement to validate module functionality from a causal‑intervention perspective. The appendix is comprehensive, with rich supporting materials covering multi‑seed stability, bootstrap confidence intervals, and t‑SNE visualizations.

Critical strengths

Novel problem insight that captures a key existing limitation in the field. Prior multimodal fake‑news detection work directly equates low cross‑modal consistency with conflict, while overlooking the unverified case caused by insufficient evidence. This paper clearly distinguishes two situations:

  • Corresponding evidence exists but contains contradictory content;
  • There is inadequate corresponding evidence to form a judgment.
Critical weaknesses
  1. The baseline comparisons omit several strong recent baselines, especially LLM / VLM‑based reasoning‑style multimodal fake‑news detection methods.

  2. The workflow diagram appears to be AI‑generated.

  3. Many experimental result cells for compared methods are left blank.

  4. Performance gains brought by several key components are modest; some ablations only produce accuracy drops of a few tenths of one percentage point.

  5. The null‑hypothesis formula contains learnable parameters bdb_d and γd\gamma_d, which control the strength of entropy‑driven penalty applied to availability scores. The paper does not provide hyper‑parameter sensitivity analysis for these parameters: how different values of bdb_d and γd\gamma_d affect availability estimation and downstream performance. A brief sensitivity sweep could be added in the appendix.

Questions for authors
  1. The baseline comparisons omit several strong recent baselines, especially LLM / VLM‑based reasoning‑style multimodal fake‑news detection methods.

  2. The workflow diagram appears to be AI‑generated.

  3. Many experimental result cells for compared methods are left blank.

  4. Performance gains brought by several key components are modest; some ablations only produce accuracy drops of a few tenths of one percentage point.

  5. The null‑hypothesis formula contains learnable parameters bdb_d and γd\gamma_d, which control the strength of entropy‑driven penalty applied to availability scores. The paper does not provide hyper‑parameter sensitivity analysis for these parameters: how different values of bdb_d and γd\gamma_d affect availability estimation and downstream performance. A brief sensitivity sweep could be added in the appendix.

2 comments
AuthorsOpenReview ↗

Clarifications on Baselines, Ablations, and Parameter Sensitivity

We thank the reviewer for the constructive feedback. We address the five concerns below with clarifications and experimental evidence.

Q1. Recent LLM and VLM baselines

Answer: These methods are relevant, but differences in evaluation protocols and external resources prevent direct comparison. We do not claim superiority over them.

  1. Protocols. FKA-Owl mainly uses DGM4. GLPN-LLM uses 4,141 training and 1,125 test Weibo items. MMRGV uses 7,853 Weibo items with a 6:2:2 split, versus ACHER's official 9,528-item partition. T2Agent addresses six-class forgery attribution on AMG rather than binary veracity detection.
  2. Resources. FKA-Owl uses forgery-specific knowledge, GLPN-LLM uses LLM pseudo-labels, MMRGV uses Qwen2-VL-72B rationales, and T2Agent uses external tools. ACHER uses the given image-text pair with BERT and CLIP features.
  3. Interpretation. We acknowledge MMRGV's higher reported Weibo accuracy under its own protocol. A matched evaluation is required for direct comparison.

Q2. Provenance of Figure 1

Answer: Figure 1 was manually created by the authors, not generated by an AI image-generation model. We can provide the original editable source file, including its graphical elements, labels, and connections, for verification.

Q3. Blank entries in Table 1

Answer: Blank cells arise because Table 1 combines primary Weibo/FineFake and extended AMG evaluations with different baseline sets.

  1. Evaluation design. Weibo and FineFake are our primary benchmarks. We additionally introduce AMG to evaluate detection across diverse misinformation scenarios. CAFE and BMR serve as common reference methods across all three datasets.
  2. Table organization. We merged the primary and extended evaluations into Table 1 for compact presentation. Their baseline sets differ, so some method–dataset combinations have no reported results, producing blank entries.
  3. Revision. We will revise the layout to distinguish the two evaluation settings and prevent confusion.

Q4. Modest ablation gains

Answer: Although some ACC drops are modest, ablations and interventions jointly support the complementary and evidence-dependent roles of ACHER's components.

  1. Structured design. Replacing symmetric–antisymmetric mechanism organization with directional-state concatenation reduces ACC from 94.9% to 93.1% on Weibo and from 80.9% to 78.5% on FineFake. These 1.8 and 2.4-point drops demonstrate the contribution of the overall structured mechanism design.
  2. Functional interventions. At 20% deletion, targeted removal reduces estimated availability 7.3–27.8 times more than random removal on Weibo and 5.4–15.8 times more on FineFake. In Table 4, counterpart replacement increases discrepancy and conditional conflict while mean absolute availability change remains at most 0.0054.
  3. External semantic evaluation. We separately retrained ACHER on SNLI-VE using its human-annotated entailment, neutral, and contradiction labels as additional semantic supervision. The resulting availability and conditional-conflict scores achieve AUROCs of 0.7916 and 0.8376, respectively, versus 0.4976 and 0.5009 under shuffled-label supervision. These image-level proxy results complement accuracy-based ablations by supporting the learnability of human-defined semantic distinctions.
  4. Evidence-specific utility. In Table 5, removing Unverified changes Macro-F1 by -2.35 points in the both-low-availability Weibo cohort versus -0.45 in both-high. The corresponding FineFake changes are -0.52 and +0.06. Its contribution is therefore strongest when model-estimated counterpart evidence is scarce.
  5. Component contributions. Removing individual mechanisms lowers ACC by 0.4–1.5 points. Removing evidence availability, null competition, sparse localization, or state gating also reduces ACC on both datasets, supporting their complementary roles.

Q5. Null-score parameter sensitivity

Answer: We conducted two one-factor-at-a-time FineFake sweeps of learnable bdb_d and γd\gamma_d initializations, fixing the other parameter in each sweep.

  1. Initialization selection. Preliminary experiments considering validation performance and availability estimates selected bd=−0.40b_d=-0.40 and γd=0.15\gamma_d=0.15 as the reference.
  2. Accuracy. Across both sensitivity sweeps, ACC ranges from 79.3% to 80.9%. The adopted initialization achieves 80.9%. The trends agree with nid=bd+γdH(αid)n_i^d=b_d+\gamma_d H(\alpha_i^d). The sweep shows a stronger effect from bdb_d over the tested ranges and characterizes initialization sensitivity.
  3. Bias sensitivity. With γd=0.15\gamma_d=0.15, increasing initial bdb_d from -0.80 to -0.20 changes mean availability from 0.7565 to 0.6322 for text-to-visual and from 0.7098 to 0.5859 for visual-to-text.
  4. Entropy sensitivity. With bd=−0.40b_d=-0.40, increasing initial γd\gamma_d from 0.05 to 0.25 changes the respective means from 0.6886 to 0.6646 and from 0.6364 to 0.6137.

We will clarify the evaluation scope, table layout, and sensitivity results in the revision.

Reviewer_b9hvOpenReview ↗

response

Thank you for the clarification. However, several concerns remain unresolved.

  1. Completeness and provenance of the experimental results. The tables 1 contain many missing entries. Please clarify whether these correspond to experiments that were not conducted, results unavailable in the original publications, or settings that are not applicable. Please also distinguish results reproduced under your evaluation protocol from those quoted from prior papers. The current presentation makes it difficult to assess the completeness and fairness of the comparisons.

  2. Comparisons with VLM and agent-based methods. I acknowledge the differences in datasets, evaluation protocols, and external resources. However, listing these differences does not resolve the need for empirical comparisons. I recommend evaluating representative, task-relevant VLM and agent-based baselines under a matched protocol, with any adaptations and additional resources clearly documented. I am asking for additional experimental evidence; acknowledging that the published numbers are not directly comparable does not adequately address this concern.

  3. Presentation of Figure 1. Thank you for clarifying that the figure was manually created. My remaining concern is its presentation quality: the current figure appears to be rasterized, and its visual style resembles generic AI-generated illustrations. This is an observation about its appearance, not a claim about its provenance. I recommend reconstructing it as a vector graphic with consistent typography, clearly defined modules, and precise connections, following the conventions of technical diagrams in relevant top-tier conference papers. Providing the editable source would clarify provenance, but the figure itself still needs improvement.

The current response does not sufficiently address these concerns, particularly those regarding the experimental evaluation. I therefore maintain my original rating.

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

Related papers

Discussion (0)

Sign in to comment.