{"id":"bf005a24-ce01-4c75-a07f-48cba759fb69","arxiv_id":"2505.11262","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MultiFIX replaces black-box multimodal fusion components with symbolic expressions and Grad-CAM explanations, and on four synthetic problems the interpretable models retain or improve predictive accuracy.","lead":"MultiFIX combines deep learning and genetic programming to build multimodal AI models whose features and fusion rules are expressed as readable symbolic formulas. On synthetic image and tabular datasets, these interpretable models match or exceed the accuracy of the black-box models they replace.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GP surrogates are never checked for fidelity to the DL model; the explainable model may be a different, equally accurate model, so the interpretability claim is not yet established.","rationale":"The reader's weakest_assumption is that the GP symbolic expression is a faithful interpretation of the black-box model, which is exactly the load-bearing point I identify. The paper's experiments demonstrate that the fully replaced explainable model can achieve similar or better accuracy, but accuracy alone cannot establish that the surrogate explains the DL model. Because the GP expressions may realize an equally accurate but different decision function—as seen in the AND and XOR problems, where features are inverted relative to the ground truth—the explanations may say nothing about the internal computations of the DL components they are supposed to replace. This is not merely an external-consensus objection; it is an internal evidential gap: the manuscript contains no fidelity metric, no prediction-agreement analysis, and no feature-level comparison between the DL blocks and their GP replacements. The central claim of interpretability therefore rests on an untested equivalence between the surrogate and the black-box model. The reader's conditional verdict is appropriate: the work is promising and the synthetic experiments are well-structured, but the interpretability claim should be accepted only after the proposed fidelity checks are run. I see no evidence of fraud or methodological misconduct; the issue is a missing line of evidence for the paper's core assertion.","tokens_in":12074,"tokens_out":4263,"duration_ms":46362,"concrete_test":"On held-out folds for all four problems, compute: (1) per-sample agreement (exact match rate or Cohen's kappa) between the DL model's predictions and the explainable model's predictions; (2) for each tabular feature, the R^2 or rank correlation between the DL MLP's extracted feature and the GP expression evaluated on the same inputs; (3) the correlation between the DL fusion logits and the GP fusion output. If agreement or feature-level fidelity is low (e.g., kappa below roughly 0.9 or feature R^2 far from 1), the symbolic expressions are not faithful surrogates. Additionally, rerun GP under two objectives—fitting the DL block outputs versus fitting the ground-truth target—and report both, to separate task-level accuracy from model-level fidelity.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract's central claim is that MultiFIX can generate models that 'accurately explain' the extracted features and their integration. The evaluation, however, only compares the BAcc of the final explainable model with the BAcc of the DL model (e.g., AND: 0.967 vs 0.973; XOR: 0.975 vs 0.940; Multiclass: 0.995 vs 0.920). No fidelity metric is reported between the GP symbolic expressions and the DL blocks they replace. In Section 3.5, GP-GOMEA is used to 'replace' the tabular feature block and the fusion block, but the paper never checks whether the GP expressions reproduce the DL block outputs on the same inputs. The AND/XOR results make the problem concrete: the GP found an 'equivalent' model with inverted intermediate features (Section 4.1.3, 4.2.3). Those inversions may not correspond to what the trained DL model actually computes; they are merely another function that fits the task. Consequently, a high-BAcc explainable model can be explaining the data-generating rule rather than the black-box model, undermining the claim that MultiFIX explains the original DL model's extracted features and integration. The paper itself notes in Section 5 that Grad-CAM's post-hoc nature limits symbolic meaning, but it does not acknowledge or address the analogous fidelity gap for the GP replacements. Without a quantitative check of surrogate fidelity, the central interpretability claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents MultiFIX, a multimodal fusion pipeline that first trains a deep learning model with small bottlenecks for image and tabular features, then replaces the tabular feature-extraction block and the fusion block with symbolic expressions evolved by GP-GOMEA, while explaining image features with Grad-CAM. Experiments on four synthetic problems (AND, XOR, Multifeature, Multiclass) compare balanced accuracy across six training strategies and report interpretable models for the best-performing strategy per problem. The paper claims that MultiFIX can generate multimodal models that accurately explain both the extracted features and their integration without compromising predictive performance.","tokens_in":12376,"tokens_out":5767,"duration_ms":49373,"significance":"If the fidelity of the GP surrogates to the DL blocks were established, the combination of a sparse feature bottleneck with GP symbolic expressions would be a genuinely useful step toward interpretable multimodal modeling, particularly because the synthetic problems have known ground truth and the paper examines several training strategies. The paper is honest about several limitations, including the difficulty of interpreting complex image features and the post-hoc nature of Grad-CAM. However, the central interpretability claim is not yet supported by the reported evidence, since no fidelity metric is provided between the GP expressions and the DL blocks they replace, and the interpretable-model results are reported for only the best training strategy without variance.","major_comments":[{"comment":"The central interpretability claim is unsupported because the GP symbolic expressions are never checked for fidelity to the DL blocks they replace. The evaluation only compares the final explainable model's balanced accuracy with the DL model's balanced accuracy (e.g., AND: 0.967 vs 0.973; XOR: 0.975 vs 0.940; Multifeature: 0.884 vs 0.864; Multiclass: 0.995 vs 0.920). A high-BAcc explainable model could be explaining the data-generating rule rather than the original DL model, especially since Sections 4.1.3 and 4.2.3 report equivalent models with inverted intermediate features. The authors should quantify the agreement between each GP expression and the corresponding DL block output on held-out data (e.g., R^2 or mean absolute error for tabular features, and accuracy or BAcc for the fusion expression), and report how often the final explainable model's predictions match the DL model's predictions.","section":"3.5, 4.1.3, 4.2.3, 4.3.3, 4.4.3"},{"comment":"The interpretable-model results are reported only for the single best-performing training strategy per problem, with point estimates and no variance. For the AND problem, the explainable model's BAcc is reported as 0.967 vs the DL model's 0.973, but the DL results in Table 3 have standard deviations around 0.02-0.04, and no standard deviation is given for the explainable model. This selection and the lack of variance make it impossible to assess the robustness of the 'no performance compromise' claim. The authors should report the explainable model's performance across all folds and, ideally, for all training strategies, not only the best one.","section":"4.1-4.4"},{"comment":"The manuscript itself acknowledges that in the Multifeature problem the intermediate image features are 'not easy to interpret' and that the fusion expression with depth three is 'not easy to interpret' (Section 4.3.3), and that Grad-CAM's post-hoc nature limits symbolic meaning (Section 5). This directly undercuts the conclusion that MultiFIX produces interpretable models for that problem. The authors should provide a concrete interpretability metric or a clear protocol for deciding when an obtained expression counts as interpretable, and discuss how the Multifeature result is compatible with the abstract's claim.","section":"4.3.3 and 5"},{"comment":"The statistical comparison of training strategies relies on paired t-tests over 5 folds. With n=5, the tests have very low power, and the paper interprets 'not significant' as evidence of similar performance, which is not justified. The authors should either report effect sizes with confidence intervals, use a resampling-based procedure such as the bootstrap or the corrected repeated k-fold CV test, or clearly state the power limitations of the comparison.","section":"3.6, Tables 3-7"}],"minor_comments":[{"comment":"The operator list 'numeric[+,−,∗,/,.2,.3,]]' is malformed; the intended numeric operators and constants (e.g., 0.2 and 0.3) are unclear and should be stated without stray brackets.","section":"Table 1"},{"comment":"Clarify whether the GP symbolic expressions for tabular features are fitted to the DL-extracted feature values using the raw tabular inputs as independent variables, and whether the fusion GP uses the DL-extracted or the GP-reconstructed features.","section":"3.5"},{"comment":"The text states that HPO chooses the learning rate, weight decay, and 'number of extracted features (up to three) for the image and tabular inputs,' but Table 2 lists Image Bottleneck and Tabular Bottleneck; specify whether these are the same quantities.","section":"3.6"},{"comment":"The word 'MutiFIX' appears as a typo in the first paragraph of Section 5; it should be 'MultiFIX.'","section":"5"},{"comment":"The symbolic expressions in the figures are too small to read in the manuscript; consider displaying them in a larger font or presenting them in the text as well.","section":"Figures 3 and 4"},{"comment":"Clarify how the 1,000 samples are reused across the four problems and whether the same image/tabular pairs are used in each problem, since this affects the comparability of the reported results.","section":"3.7"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about surrogate fidelity is well-founded and is the main obstacle to the paper's central claim. The missing fidelity check is, however, obtainable within the scope of the paper, so I see no fundamental flaw that would justify rejection. The authors should be asked to add quantitative fidelity checks and to report explainable-model performance with variance across folds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead MultiFIX. Bottom line: the training-strategy comparison is solid, but the central interpretability claim is under-supported by the absence of a fidelity check between the GP surrogates and the DL blocks they replace.\n\nWhat's new: the pipeline itself is from the authors' earlier preprint; this paper's contribution is a systematic comparison of six training strategies on four synthetic multimodal problems with known ground truth. That comparison is a legitimate, useful contribution. The experimental protocol is reasonable: 5-fold CV, paired t-tests with Bonferroni correction, and the authors are candid about several limitations—notably the Multifeature problem where the evolved expression is no longer easy to interpret. Forcing a narrow bottleneck of at most three features per modality is a genuinely sensible design choice for interpretability. The related work is fairly cited, and the authors explicitly flag their own preliminary work [14] as the source of the architecture.\n\nWhere it gets soft: the abstract's claim that MultiFIX can 'accurately explain both the extracted features and their integration' outruns the evidence. The GP expressions are said to replace the tabular feature block and the fusion block, yet the paper never reports whether those expressions reproduce the DL block outputs on the same inputs. The AND and XOR results make the problem concrete: the GP finds an 'equivalent' model with inverted intermediate features. That may be a faithful reproduction of what the DL network learned, or it may be a different function that achieves the same accuracy. Without a fidelity metric, the reader cannot tell which, so the interpretability claim is not established. The paper also reports the interpretable-model results only for the best-performing strategy per problem, with point estimates and no standard deviations, and no code or data are provided. Finally, Section 4 says 'five synthetic problems' but only four are described—a minor editorial inconsistency.\n\nNone of this is fatal. The training-strategy findings stand on their own, and the authors' self-awareness about the method's limitations is a point in their favor. But the headline claim needs to be either tightened or backed by a quantitative comparison between the GP surrogates and the DL blocks.\n\nMy recommendation: send it for peer review, with a request that reviewers push for a fidelity analysis and full experimental reporting. This is a paper that deserves referee time, but it needs revision before it should be accepted.","headline":"The training-strategy comparison is solid, but the central claim that MultiFIX explains the DL model's features is under-supported by the missing fidelity checks.","tokens_in":12879,"tokens_out":4122,"would_cite":true,"duration_ms":39223,"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":"MultiFIX replaces the opaque parts of a multimodal deep-learning model with readable symbolic expressions, and on synthetic problems the resulting explainable model matches or beats the black box's balanced accuracy.","keywords":["genetic programming","interpretability","multimodality","multimodal fusion","symbolic regression","explainable AI","deep learning","Grad-CAM"],"falsifier":"Run MultiFIX on a synthetic multimodal problem with a known hidden rule, give the user only the recovered expressions, and check whether they can correctly answer a question the intended rule answers. The AND and XOR results show the model can be right while describing a different, inverted logic; if a non-expert reader cannot tell whether the intended rule was actually learned, the claim that MultiFIX accurately explains the model's integration of modalities is not established.","tokens_in":11906,"feed_emoji":"🧩","tokens_out":6320,"duration_ms":53783,"temperature":0.7,"pith_summary":"MultiFIX is a pipeline for building multimodal AI models that can explain themselves. It trains deep-learning blocks to extract a deliberately narrow set of features from each modality, then replaces the tabular feature extractor and the fusion layer with symbolic expressions evolved by genetic programming, and explains the image block with Grad-CAM heatmaps. The paper's central claim is that these explainable replacements do not cost predictive performance: on synthetic problems including AND, XOR, Multifeature, and Multiclass, the explainable model's balanced accuracy stayed within 0.006 lower to 0.075 higher than the original black-box model. If true, this matters in high-stakes domains such as healthcare, where practitioners need to know how each modality contributes to a decision.","feed_headline":"MultiFIX explains multimodal AI with no accuracy loss","feed_subtitle":"Symbolic expressions replace black-box blocks; balanced accuracy lands within 0.006 low to 0.075 high.","key_machinery":"The load-bearing mechanism is sparse feature engineering plus symbolic replacement. Each modality is reduced to a bottleneck of at most three real-valued features; the tabular features and the fusion block are then approximated by tiny symbolic expressions (trees of arithmetic, Boolean, and if-then-else operators) evolved by GP-GOMEA, while Grad-CAM visualizes which image regions drive each image feature. The narrow bottleneck is what makes fusion-level symbolic regression feasible and readable.","core_discovery":"The paper claims that multimodal predictions can be made interpretable by design without sacrificing accuracy, provided each modality is forced through a sparse bottleneck of at most three engineered features. On problems built from synthetic images of shapes and ten tabular variables, the DL model's tabular features and its fusion were replaced with compact symbolic expressions (using GP-GOMEA), while image features were explained post-hoc with Grad-CAM. For the AND, XOR, Multifeature, and Multiclass problems, the resulting interpretable model matched or improved on the DL model's balanced accuracy, by up to 0.075. The authors also show that the recovered expressions may encode equivalent-but-inverted logic: e.g., a learned tabular feature was the complement of the intended $x_1 > x_2$, yet the model remained correct. They interpret this as evidence that component-wise explanation lets a user see what the model actually learned, not just that it predicts well.","pith_inferences":["The same replacement strategy could be applied to other modality blocks—text or time series encoders—as long as the bottleneck stays small enough for symbolic regression to fit; on current evidence there is nothing image- or table-specific about the mechanism.","A controlled ablation that keeps the DL fusion but uses symbolic features would separate two effects the paper conflates: the gains on XOR and Multiclass (0.035 and 0.075 BAcc) may come from GP's regularization rather than from interpretability itself.","The observed equivalence phenomenon suggests a concrete diagnostic test for trust: if the symbolic expressions and a known ground-truth rule are equivalent-but-different, the model's reasoning is correct but non-unique, and a deployment decision should be reviewed by a domain expert."],"forward_implications":["Interpretable models can be obtained from DL-trained multimodal pipelines without a performance penalty, at least on problems whose true feature structure is simple and known.","Component-wise inspection can reveal that the model uses a different but equally valid encoding than the intended one, turning equivalent-model discovery from a surprise into a feature of the audit trail.","Training strategy matters less than problem structure; end-to-end, hybrid, and sequential variants were mostly not significantly different, so practitioners can pick based on available pretrained blocks.","The hard case is when a modality needs more than one complex feature: the Multifeature problem produced expressions that are transparent but not easy to interpret, motivating complexity penalties in future work."],"supporting_citations":[{"why":"the preliminary MultiFIX work that introduced the feature-inducing pipeline and end-to-end training","marker":"[14]"},{"why":"GP-GOMEA, the model-based genetic programming algorithm that evolves the small symbolic expressions","marker":"[25]"},{"why":"Grad-CAM, the gradient-based method used to produce visual explanations of the image block","marker":"[20]"},{"why":"Evans et al.'s demonstration that GP can approximate ML estimators as a surrogate without large accuracy loss, the premise for replacing black-box blocks","marker":"[5]"},{"why":"ResNet, the pretrained convolutional backbone used as the image feature engineering block","marker":"[8]"},{"why":"the higher-arity GP-GOMEA adaptation that allows Boolean and if-then-else operators in the symbolic expressions","marker":"[18]"},{"why":"Multimodn, the inherently interpretable multimodal baseline that MultiFIX positions itself against","marker":"[23]"}],"fun_headline_variants":["MultiFIX: interpretable multimodal AI without accuracy loss","Symbolic expressions replace black-box fusion, keep accuracy","Sparse features yield explainable multimodal models","See what multimodal AI learns, accuracy is preserved","Inverted logic still correct: MultiFIX explains and predicts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on the assumption that a handful of symbolic expressions can faithfully stand in for the deep-learning feature extractors; if the extracted features are complex, the paper itself shows the recovering expressions can become hard to interpret, and the explanation then no longer explains much.","fun_headline_variants_meta":{"raw":{"variants":["MultiFIX: interpretable multimodal AI without accuracy loss","Symbolic expressions replace black-box fusion, keep accuracy","Sparse features yield explainable multimodal models","See what multimodal AI learns, accuracy is preserved","Inverted logic still correct: MultiFIX explains and predicts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001231,"raw_usage":{"total_tokens":5042,"prompt_tokens":918,"completion_tokens":4124,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":4048}},"tokens_in":534,"tokens_out":4124,"duration_ms":30464,"temperature":1.0,"reasoning_tokens":4048,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:54:15.338366+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MultiFIX on a synthetic multimodal problem with a known hidden rule, give the user only the recovered expressions, and check whether they can correctly answer a question the intended rule answers. The AND and XOR results show the model can be right while describing a different, inverted logic; if a non-expert reader cannot tell whether the intended rule was actually learned, the claim that MultiFIX accurately explains the model's integration of modalities is not established.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"the preliminary MultiFIX work that introduced the feature-inducing pipeline and end-to-end training"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Evans et al.'s demonstration that GP can approximate ML estimators as a surrogate without large accuracy loss, the premise for replacing black-box blocks"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"the higher-arity GP-GOMEA adaptation that allows Boolean and if-then-else operators in the symbolic expressions"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Multimodn, the inherently interpretable multimodal baseline that MultiFIX positions itself against"}],"review_version":1}