{"id":"b3797ea3-dd52-494a-bdaa-68e8fd57c05c","arxiv_id":"2506.13060","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Unimodal explanation methods should not be used for multimodal AI models; the paper formalizes three desiderata for modality-aware explanations.","lead":"This position paper argues that standard explanation tools, which handle one input type at a time, are not trustworthy for AI models that combine images, text, audio, or graphs. It proposes three properties that multimodal explanations should satisfy: modality influence, joint faithfulness, and stability under small cross-modal perturbations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Formal desiderata are vacuous: Eq. (2) and Eqs. (6)-(7) admit existential constants that any bounded explainer (even the zero explainer) can satisfy, so they cannot ground the paper's evaluation criteria or its ban on unimodal explanations.","rationale":"The reader's weakest assumption identifies the same formal vacuity; I agree. The paper's stated goal is to provide 'key principles' and 'evaluation criteria' for multimodal explainability, and the conclusion is a call to stop using unimodal tools. If the criteria are satisfied by a constant explainer, they do no evidential work. I checked the exact inequalities: Eq. (2) has existentially quantified αm,βm with no coupling across inputs or methods; choosing βm ≥ αm max_x Δfm makes the RHS non-positive for all x, so any nonnegative ΔEm satisfies it. Eq. (6) and Eq. (7) similarly allow zero tolerances/zero Lipschitz constants. This is an internal formal problem, not merely a disagreement with the community's preferred metrics. It is the load-bearing point because the paper explicitly presents these desiderata as the basis for evaluating and designing multimodal explainers; without non-vacuous versions, the formal contribution collapses into restating intuitive properties. The central negative claim might still be true, and the anecdotal examples remain suggestive, so I would keep the reader's CONDITIONAL verdict rather than reject: the required revision is concrete and achievable (fix or eliminate the existential constants, then demonstrate separation on a synthetic ground-truth task). This is why verdict_should_be is UNCHANGED.","tokens_in":14515,"tokens_out":8196,"duration_ms":102465,"concrete_test":"Construct a fixed multimodal model f on a finite input set and run the zero explainer Em(x)=0 through Eqs. (2), (6), (7). For Eq. (2), set αm=1 and βm=max_x |f(x)-f(x_-m)|; the inequality holds for every x. For Eqs. (6)-(7), set ηm=0, ϵ_{m→n}=0, L'_m=Ln,m=0; the bounds hold trivially. If the authors intend α, β, η, ϵ, L to be non-vacuous (e.g., fixed a priori, positive, or shared across explainers), the paper must state this and show that at least one existing method fails the resulting constraints; otherwise the desiderata cannot separate any method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central proposal is that unimodal explanations 'systematically misrepresent' multimodal decisions, and its formal desiderata are offered as the evaluation framework for the alternative. The load-bearing weakness is that these desiderata are not falsifiable as stated. In Eq. (2), ΔEm ≥ αm Δfm - βm with αm>0 and βm≥0 is satisfied by any bounded explainer: choose βm = αm max_x Δfm, then RHS ≤ 0 ≤ ΔEm. In particular the zero explainer Em≡0 passes the Granger Modality Influence test for every f with bounded output change (e.g., finite input set), directly contradicting the text's claim that this inequality 'prevents an explainer from overstating or understating a modality's role.' The stability bounds have the same structure: Eq. (6) is vacuous if ηm can be set to 0 (only x'_m=x_m), and Eq. (7) is vacuous if ϵ_{m→n}=0; with existential Lipschitz constants no bound is ruled out for small perturbations. Thus the formal framework does not distinguish good from bad multimodal explainers, and it cannot provide the 'evaluation criteria' the paper says it provides. This does not by itself disprove the anecdotal case against per-modality attention maps or logit-lens analyses, but it removes the main constructive support for the 'stop relying on unimodal explanations' conclusion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This position paper argues that unimodal explainability methods (saliency maps, attention visualizations, circuit traces, and similar per-modality tools) systematically misrepresent the decisions of multimodal models because they fail to capture cross-modal interactions. The authors propose three desiderata for multimodal explanations—Granger-style modality influence, synergistic faithfulness, and unified stability—formalized in Eqs. (2)-(7) of Section 3. The paper supports its position with illustrative examples (e.g., a clinical chest X-ray plus text report), a discussion of attention's non-causal nature, a critique of mechanistic interpretability tools, and a small original experiment applying logit lens to a graph-language model. It concludes with a call to stop relying on unimodal explanations in multimodal settings.","tokens_in":14879,"tokens_out":3939,"duration_ms":43631,"significance":"If the central claim were rigorously established, the paper would have practical importance for explainable AI, especially in high-stakes domains. The paper usefully compiles many relevant references and highlights real limitations of per-modality attributions. It also ships a small qualitative experiment (logit lens on a graph-language model), which is a concrete attempt to demonstrate the failure of a unimodal tool. However, the paper's constructive contribution depends on the formal desiderata being meaningful constraints, and as stated they are vacuous. The empirical claim of 'systematic' misrepresentation is not supported by a systematic study. Thus, the significance is contingent on substantial revision of the formalism and evidence.","major_comments":[{"comment":"The Granger Modality Influence condition ΔEm(x) ≥ αm Δfm - βm with existential αm > 0 and βm ≥ 0 is vacuous: any explainer with bounded output (including the constant zero explainer) satisfies it by setting βm = αm max_x Δfm, which makes the right-hand side non-positive. The text claims this inequality 'prevents an explainer from overstating or understating a modality's role,' but that is not true. To make the criterion meaningful, the constants must be fixed a priori (e.g., universal across explainers, or determined by the evaluation protocol) or the inequality must be replaced by a scale-invariant condition.","section":"Section 3, Eq. (2)"},{"comment":"The Unified Stability conditions are similarly vacuous. In Eq. (6), the Lipschitz constant L'_m is existential; on a compact input space every Lipschitz continuous explainer satisfies the bound with some L'_m, so the condition rules out no method. In Eq. (7), the tolerance ϵ_{m→n} and Lipschitz constant L_{n,m} are also existential; one can set ϵ_{m→n}=0, making the condition trivially true, or choose L_{n,m} large enough to absorb any bounded change. As a result, the desiderata cannot serve as 'evaluation criteria for comparing explanation methods' as claimed in Section 3.","section":"Section 3, Eqs. (6)-(7)"},{"comment":"The central claim that unimodal explanations 'systematically misrepresent' multimodal decisions is supported only by illustrative examples (Figures 1-2), anecdotal citations (e.g., Neo et al., Casey et al., Wu et al.), and one qualitative logit-lens experiment. No quantitative evidence is provided about how frequently or severely unimodal explanations misrepresent decisions across different models, tasks, or methods. To justify the strong 'stop relying' conclusion, the paper should either report systematic experiments or explicitly frame the claim as a testable hypothesis rather than an established fact.","section":"Abstract and Section 2"},{"comment":"The synergistic faithfulness conditions depend on thresholds ϵ_keep_m and ϵ_remove_m, but the paper gives no procedure for setting these thresholds. Without a principled choice, the sufficiency and necessity tests are not operational: any explainer can pass or fail by adjusting the thresholds. The paper should specify how the thresholds are determined (e.g., fixed a priori, or relative to the model's output distribution) and how the resulting criteria distinguish between explainers.","section":"Section 3, Eqs. (4)-(5)"}],"minor_comments":[{"comment":"There is a typo in 'modelf is trained on'—it should be 'model f is trained on.'","section":"Section 3, Notations"},{"comment":"The caption refers to 'the perturbed x2 image' when the perturbation is applied to x1; this should read 'perturbed x1 image.'","section":"Figure 4 caption"},{"comment":"The phrase 'that satisfies the following bounds' is ungrammatical; it should be 'that satisfy the following bounds.'","section":"Section 3, III) Unified Stability"},{"comment":"The opening definition of faithfulness uses the phrase 'where the irrelevant input features ... are masked or zeroed out,' which is circular unless 'irrelevant' is defined independently of the explainer; the later Keep/Remove construction resolves this, but the initial sentence should be clarified.","section":"Section 3, II) Faithfulness"},{"comment":"The logit-lens experiment is described only qualitatively; it would be clearer if the text explained why tokens like 'bomb' and 'neighbor' are considered evidence of failure, and how the 'static across layers' observation is quantified.","section":"Figure 2 and Appendix A"},{"comment":"The in-text citation 'Till, O'Neill and Bui, 2024' merges two separate references (Demian Till and Charles O'Neill with Thang Bui); this should be corrected to two distinct citations.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is best viewed as a call to action rather than a technical framework. The vacuity of the formal desiderata is a load-bearing issue that must be addressed before publication. If the author reframes the formalism as a checklist with fixed constants or testable predictions, and tempers the empirical claim, the paper could be a useful contribution to the multimodal XAI discussion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's core argument is fair and worth taking seriously: per-modality heatmaps, token highlights, and attention views cannot represent cross-modal interactions, and the field should stop treating them as adequate explanations for multimodal models. The three desiderata (Granger-style modality influence, synergistic faithfulness, unified stability) are a sensible bundle, even if each descends from existing ideas, and the paper does a good job of motivating them with concrete examples, including the clinical shortcut case and the graph-language logit lens experiment. That experiment is small and qualitative, but it is a genuine attempt to show a tool failing on a multimodal model, and it is reproducible in principle. The writing is clear, and the citation pattern is fine; the self-citations support background claims rather than padding.\n\nThe soft spot is not the motivation, it is the formal core. The inequalities in Eqs. (2), (6), and (7) are stated with existential constants and no constraints on their magnitudes. In Eq. (2), for any bounded explainer, including the zero explainer, you can choose beta_m large enough that the right-hand side is non-positive, so the inequality is trivially satisfied. The text claims this condition \"prevents an explainer from overstating or understating a modality's role,\" but as written it prevents nothing. The stability bounds have the same problem: arbitrary Lipschitz constants and tolerances make them vacuous for small perturbations. So the desiderata cannot serve as evaluation criteria in the sense the paper claims. They are names for desirable properties, not testable constraints.\n\nThe empirical case for the strong conclusion that unimodal explanations \"systematically misrepresent\" multimodal decisions is also thinner than the abstract suggests. The evidence is illustrative and the one new experiment is qualitative. That does not mean the conclusion is wrong; it means the paper has not yet established it with the rigor the conclusion demands.\n\nWho is this for? Someone working on multimodal XAI evaluation will find the examples and the call to action useful, but they should not adopt Eqs. (2)-(7) as a benchmark without substantial repair. A serious referee could help fix this: require the constants to be chosen in a principled way, or replace the existential inequalities with comparisons that are scale-invariant, or clearly reframe the desiderata as qualitative guidelines rather than formal evaluation criteria. The paper deserves peer review because the question is important and the argument is clear, but the formalization needs to be made non-vacuous before the evaluation protocol can be taken literally.","headline":"A timely position paper with a real point about unimodal explanations, but the formal desiderata are vacuous as stated and cannot do the evaluative work the paper claims.","tokens_in":15350,"tokens_out":1765,"would_cite":false,"duration_ms":22113,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that per-modality heatmaps, attention views, and circuit traces systematically misrepresent how multimodal models decide, and that the field should stop relying on them.","keywords":["multimodal explainability","unimodal explanations","cross-modal influence","Granger modality influence","synergistic faithfulness","unified stability","attention not explanation","explainable AI"],"falsifier":"Take any multimodal model and a deliberately uninformative explainer that always returns constant attribution, so $\\Delta E_m = 0$. Because the proposed criterion allows choosing $\\beta_m$ large enough to satisfy $\\Delta E_m \\geq \\alpha_m \\Delta f_m - \\beta_m$, the explainer passes the Granger test while contributing nothing; observing this on a real model would falsify the claim that the desiderata separate faithful from misleading explanations.","tokens_in":14336,"feed_emoji":"🧩","tokens_out":10778,"duration_ms":109186,"temperature":0.7,"pith_summary":"This paper argues that existing explainability tools answer the wrong question when applied to multimodal models. Heatmaps, token highlights, attention views, and circuit traces are all built for a single input stream, so they cannot represent the interaction where one modality changes the role of another. The paper's position is that such unimodal explanations can be plausible yet systematically wrong, and that the field should stop treating them as evidence about multimodal decisions. To give the field something to build toward, it proposes three properties a multimodal explanation should satisfy: modality influence, synergistic faithfulness, and unified stability. If the paper is right, the safety case for high-stakes multimodal systems depends on developing and checking explanations against these cross-modal properties.","feed_headline":"Stop using single-modality tools to explain multimodal AI","feed_subtitle":"Unimodal maps hide cross-modal influence; three desiderata define what true multimodal explanations must show.","key_machinery":"The carrying object is a formal definition of a multimodal explanation as a collection of per-modality explanation objects $E = \\{E_1, \\dots, E_M\\}$, paired with three quantitative tests. The first test compares the change $\\Delta E_m$ in the explanation for modality $m$ with the change $\\Delta f_m$ in the model's output when that modality is replaced by a neutral reference, requiring $\\Delta E_m \\geq \\alpha_m \\Delta f_m - \\beta_m$ for all inputs. The second test builds kept and removed versions of each modality from the explainer's own output and requires the model's prediction to stay close under kept inputs and degrade under removed inputs. The third test imposes Lipschitz-style bounds: explanations for one modality must vary slowly under semantic-preserving changes to the same modality and under small perturbations to any other modality. These definitions carry the argument because they turn the complaint that unimodal explainers miss cross-modal influence into concrete criteria that a future multimodal explainer could be checked against.","core_discovery":"The paper's central claim is that a multimodal model's output is a joint function of its inputs, so an explanation that treats each input type separately cannot represent the cross-modal influence behind a decision. The paper states this as a categorical position: the community should stop relying on unimodal explainability techniques for multimodal models. As support, it points to cases where unimodal attributions look plausible while the model actually relies on a different modality, such as a clinical system that predicts pneumonia from the word 'cough' while ignoring a chest X-ray, and to experiments showing that gradient-based and attention-based explanations do not track what vision-language models use. The paper then formalizes what a proper multimodal explanation must do through three desiderata, each stated as a quantitative condition on a model and its per-modality explanation objects.","pith_inferences":["Editorial extension: The three desiderata are declared in a modality-agnostic form, so benchmarking them on synthetic text-image, audio-text, or graph-language tasks with known modality contributions is a direct next step the paper calls for but does not execute.","Editorial extension: The Granger-style influence condition could be used as a screening test on existing deployed models: any explainer showing near-zero importance for a modality whose ablation changes the output would signal a hidden dependence or a broken explainer, without retraining.","Editorial extension: The stability bounds suggest a practical stress test for any candidate multimodal explainer using synonyms, brightness shifts, or other imperceptible perturbations; unimodal explainers would be expected to fail these bounds already at small magnitudes, which would make the paper's position empirically checkable."],"forward_implications":["Per-modality heatmaps, attention maps, and token highlights should no longer be accepted as explanations of a multimodal model's decision.","Evaluation of multimodal explainers should include ablation-based modality-influence tests, sufficiency and necessity perturbation tests, and cross-modal stability bounds.","Shortcut detection becomes possible in principle: a criterion that tracks ablation changes can reveal when a model ignores one modality entirely, even if each unimodal map looks plausible.","Future multimodal model design should incorporate explainability early, for example through explicit alignment modules or disentangled representations that expose modality contributions."],"supporting_citations":[{"why":"Provides the token-ablation result showing gradient-based unimodal attributions underperform in vision-language object identification.","marker":"[Neo et al., 2024]"},{"why":"Shows radiology NLP models can predict disease from text templates without imaging, illustrating the shortcut unimodal explanations cannot expose.","marker":"[Casey et al., 2021]"},{"why":"Demonstrates attention weights can be manipulated without changing predictions, supporting the claim that attention is not causal influence.","marker":"[Jain and Wallace, 2019]"},{"why":"Shows attention distributions can be reweighted without changing model predictions, reinforcing the non-robustness of attention as explanation.","marker":"[Wiegreffe and Pinter, 2019]"},{"why":"Supplies evidence that explanation methods are unstable under small input perturbations, motivating the unified stability desideratum.","marker":"[Alvarez-Melis and Jaakkola, 2018]"},{"why":"Provides the graph language model used in the logit lens case study, demonstrating the failure of a text-centric interpretability method on a multimodal architecture.","marker":"Wang et al. [2024]"},{"why":"Cited as an early approach to measuring cross-modal interactions, providing proof that multimodal explanations are feasible.","marker":"[Wenderoth et al., 2025]"},{"why":"Reports that multimodal clinical models outperform unimodal models and that deriving a meaningful overall explanation across data types is hard.","marker":"[Pahud de Mortanges et al., 2024]"}],"fun_headline_variants":["Unimodal explanations fail to capture cross-modal influence","Multimodal models demand cross-modal explanations","Stop explaining multimodal decisions with one modality","Three principles for true multimodal explainability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proposal assumes that inequalities with adjustable constants can actually separate good from bad explainers; as written, a bounded explainer can satisfy them by choosing a tiny multiplier and a large offset, so the tests bind only if the constants are fixed or constrained.","fun_headline_variants_meta":{"raw":{"variants":["Unimodal explanations fail to capture cross-modal influence","Multimodal models demand cross-modal explanations","Stop explaining multimodal decisions with one modality","Three principles for true multimodal explainability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000348,"raw_usage":{"total_tokens":1883,"prompt_tokens":904,"completion_tokens":979,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":924}},"tokens_in":520,"tokens_out":979,"duration_ms":11975,"temperature":1.0,"reasoning_tokens":924,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:35:59.629577+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any multimodal model and a deliberately uninformative explainer that always returns constant attribution, so $\\Delta E_m = 0$. Because the proposed criterion allows choosing $\\beta_m$ large enough to satisfy $\\Delta E_m \\geq \\alpha_m \\Delta f_m - \\beta_m$, the explainer passes the Granger test while contributing nothing; observing this on a real model would falsify the claim that the desiderata separate faithful from misleading explanations.","supporting_citations":[{"cited_title":"Towards interpreting visual information processing in vision-language models","cited_arxiv_id":null,"evidence_quote":"Provides the token-ablation result showing gradient-based unimodal attributions underperform in vision-language object identification."},{"cited_title":"A systematic review of natural language processing applied to radiology reports","cited_arxiv_id":null,"evidence_quote":"Shows radiology NLP models can predict disease from text templates without imaging, illustrating the shortcut unimodal explanations cannot expose."},{"cited_title":"Attention is not explanation","cited_arxiv_id":null,"evidence_quote":"Demonstrates attention weights can be manipulated without changing predictions, supporting the claim that attention is not causal influence."},{"cited_title":"On the robustness of interpretability methods","cited_arxiv_id":null,"evidence_quote":"Supplies evidence that explanation methods are unstable under small input perturbations, motivating the unified stability desideratum."},{"cited_title":"Llms as zero-shot graph learners: Alignment of gnn representations with llm token embeddings","cited_arxiv_id":null,"evidence_quote":"Provides the graph language model used in the logit lens case study, demonstrating the failure of a text-centric interpretability method on a multimodal architecture."},{"cited_title":"Measuring cross-modal interactions in multimodal models","cited_arxiv_id":null,"evidence_quote":"Cited as an early approach to measuring cross-modal interactions, providing proof that multimodal explanations are feasible."},{"cited_title":"Orchestrating explainable artificial intelligence for multimodal and longitudinal data in medical imaging","cited_arxiv_id":null,"evidence_quote":"Reports that multimodal clinical models outperform unimodal models and that deriving a meaningful overall explanation across data types is hard."}],"review_version":1}