{"id":"55601b82-833a-4a74-8e3c-aaca0302da0e","arxiv_id":"2505.06258","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"ABE is a PyTorch framework that integrates attribution algorithms with adversarial robustness modules, but its central axiom-preservation proof rests on an unproven finite-step Taylor equality.","lead":"The paper introduces ABE, a software framework that bundles 17 attribution methods with robustness attacks and evaluation metrics for image, text, multimodal, and object-detection tasks. The authors claim a unifying path-integral formulation guarantees any custom update method satisfies standard attribution axioms, but the proof leaves gaps.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Axiom guarantee rests on exact first-order Taylor equality; finite-step Algorithm 1 does not satisfy Completeness for nonlinear models.","rationale":"The reader correctly identified the load-bearing weakness: the proof equates the discrete Algorithm 1 with the continuous path integral and assumes the first-order Taylor expansion is exact. My stress-test concurs and sharpens the point: Eq. 7 is an equality only for affine-in-the-step models, and no step-size or error condition is supplied. For nonlinear networks, which are the paper's target, Completeness and Sensitivity are not guaranteed by the stated argument. This does not invalidate the engineering contribution — the path-integral idea is sound in the continuous limit and the framework may be useful — but the theoretical guarantee is overstated. The reader's CONDITIONAL verdict remains appropriate: the paper should explicitly state that axioms hold in the continuous limit or provide step-size/error analysis, and correct the proof of Sensitivity/Completeness. No additional objection beyond the reader's was found, so no verdict change is proposed.","tokens_in":18960,"tokens_out":2386,"duration_ms":25084,"concrete_test":"Run Algorithm 1 on f(x) = x^2 with baseline x' = 0, target x = 1, and a linear update (∆x_t = 1/T) for T = 1, 2, 4, 8. Completeness requires the summed attribution to equal f(1) - f(0) = 1. The left Riemann sums give 0, 0.5, 0.75, and 0.875 respectively, demonstrating the finite-step failure. Repeat with a small ReLU MLP and an adversarial update method (e.g., BIM) on a real input, comparing sum_j A_j against f(x) - f(x'); a discrepancy beyond numerical tolerance confirms the axiom guarantee does not hold as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section III-B0a) is that any user can plug arbitrary update methods into the Fundamental Attribution Method and the resulting method is guaranteed to satisfy Completeness and Sensitivity. The proof in Section III-B0b relies on Eq. 7, which treats the first-order Taylor expansion f(x_t) = f(x_{t-1}) + ∇f(x_{t-1})·(x_t - x_{t-1}) as an exact equality. For any nonlinear model and any nonzero finite step, this is false; it holds only when f is affine along each step or in the limit of infinitesimal steps. Consequently, the telescoping argument leading to Eq. 10 does not establish f(x_T) - f(x_0) = sum_j A_j. Algorithm 1 computes a left Riemann sum approximating the path integral in Eq. 6, but no step-size condition, error bound, or convergence requirement is given. Completeness therefore fails for finite T on general neural networks. The Sensitivity argument also collapses: from a non-zero difference f(x) - f(x') one cannot conclude any individual A_j is non-zero when the accumulated sum is only approximately equal to that difference. The continuous path-integral formulation itself is correct by the gradient theorem, but the implemented discrete method is not that integral. The paper's assertion that arbitrary update methods preserve the axioms is unsupported for the actual algorithm.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ABE, a PyTorch-based framework for attribution-based explainability, with four modules (Interpretability, Robustness, Validation, and Data & Model). Its central theoretical contribution is a class of Fundamental Attribution Methods formalized by Eq. (6) as a path integral over arbitrary update trajectories, with the claim that any user-specified update strategy yields an attribution method satisfying Completeness, Sensitivity, and Implementation Invariance. The paper also reports benchmarks of 17 attribution methods on image, text, multimodal, and object-detection tasks, and studies how different adversarial update methods affect AttEXplore's insertion/deletion scores. The authors provide an open-source implementation at https://github.com/LMBTough/ABE-XAI.","tokens_in":19204,"tokens_out":5411,"duration_ms":54746,"significance":"If the theoretical guarantee held, the unifying formulation would be genuinely useful: it would let practitioners plug a wide range of adversarial update methods into an axiom-preserving attribution pipeline, and the framework's modular design and task coverage (image, text, multimodal, object detection) are valuable for the interpretability community. The empirical comparison of many attribution methods across models and the study of different update methods in Figure 3 are useful resources, and the released code supports reproducibility. However, the central axiom-preservation proof is not valid as written, and the experimental sections do not directly validate the framework's claimed advantages over existing libraries such as InterpretDL and OmniXAI.","major_comments":[{"comment":"Eq. (7) treats the first-order Taylor expansion f(x_t) = f(x_{t-1}) + ∇f(x_{t-1})·(x_t - x_{t-1}) as an exact equality for finite update steps. For any nonlinear model and any nonzero step, this is false, so the telescoping argument in Eqs. (8)-(10) does not establish f(x_T) - f(x_0) = Σ_j A_j. Algorithm 1 computes a left Riemann sum approximating the continuous path integral in Eq. (6), but no step-size condition, number-of-steps bound, or convergence requirement is given. Since the guarantee that arbitrary update methods preserve the axioms is the paper's central theoretical claim, the proof must be repaired: either restrict to models that are affine along each update step, or state the result as an approximation with an explicit error bound and a procedure for choosing T accordingly.","section":"Section III-B0a and III-B0b, Eq. (7)"},{"comment":"The proof of the Sensitivity axiom is also invalid. From f(x_T) ≠ f(x_0) and Σ_j A_j = f(x_T) - f(x_0), at most one can conclude that at least one feature has a nonzero attribution; one cannot conclude that every feature j with x_j ≠ x'_j receives a nonzero attribution as Eq. (1) requires. Furthermore, Eq. (1) is stronger than the standard Sensitivity axiom and is violated by simple models, e.g., f(x) = x_1^2 with an irrelevant feature x_2 differing. The authors must state precisely which sensitivity property they prove and align the statement of the axiom with the proof.","section":"Section III-B0b, Sensitivity proof"},{"comment":"The claim that adversarial update methods \"guarantee that interpretability methods will avoid the issues of impossibility theorems\" because they \"inherently do not satisfy the Linear\" property is unsupported. Eq. (11) is a first-order approximation, not an exact identity, and failing one axiom in an impossibility theorem does not by itself show that the other axioms can be simultaneously satisfied. Additionally, the statement that \"any update method can satisfy the attribution axioms\" is too broad: update methods that do not define a continuous path from the baseline to the input, or that leave the differentiability region, are not covered by Eq. (6).","section":"Section III-B0d, Eq. (11) and impossibility discussion"},{"comment":"The proof of Implementation Invariance is only a restatement: the sentence \"by accumulating the gradients at each step, the final attribution results will remain consistent\" does not demonstrate invariance for adversarial update methods whose update rule may depend on model internals beyond the input-output mapping. For path methods based only on gradients of the function, invariance follows from the gradient theorem, but here the update methods (e.g., those using attention maps or intermediate representations) are not proved to be function-only. This needs a statement of the conditions under which the update method itself is implementation invariant.","section":"Section III-B0c, Implementation Invariance proof"}],"minor_comments":[{"comment":"The definition x_t = x_t + Σ_{i=1}^t Δx_i is circular and dimensionally inconsistent; it should read x_t = x' + Σ_{i=1}^t Δx_i. The integral in Eq. (6) also lacks a rigorous definition of the path parameterization.","section":"Section III-B, Eq. (6)"},{"comment":"Algorithm 1 does not take a baseline input x' as an argument and does not specify how x_0 is chosen for methods that do not use baselines, making the correspondence between the algorithm and Eq. (6) ambiguous.","section":"Algorithm 1"},{"comment":"The check and cross symbols used in Tables II and III do not render in the text; please replace them with explicit ✓/✗ or textual labels so the reader can determine which axioms each method satisfies.","section":"Table II and Table III"},{"comment":"Table V contains several concatenated numbers without separators (e.g., \"0.3890.033\"), and Figure 3's legend and axis labels are difficult to read; please clean up the formatting.","section":"Table V and Figure 3"},{"comment":"There are typos: \"will will\" in Section III-B0d and \"thereby satisfying the Sensitivity axiom Complete property\" in Section III-B0b; both should be corrected.","section":"Section III-B0b and III-B0d"},{"comment":"References [8] and [15] are duplicates of the same SHAP paper, and references [29] and [32] are the same ICML paper; please consolidate the duplicates.","section":"References"},{"comment":"The experiments benchmark attribution methods, but they do not compare the ABE framework against InterpretDL, OmniXAI, or Captum on usability, scalability, or integration effort; such a comparison would substantiate the framework-level claims in the introduction and Table I.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The paper is best viewed as a systems and benchmarking contribution whose theoretical wrapper needs repair. The continuous path-integral idea in Eq. (6) is standard and recoverable, so I am recommending major revision rather than rejection. The authors' own prior work constitutes a large fraction of the implemented methods, which is not by itself a problem, but the experiments should more clearly position ABE as a framework rather than only as a collection of existing algorithms. The referee report focuses on the load-bearing axiom-preservation proof, which needs a rigorous treatment of discretization error or a restriction to appropriate model classes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick read of ABE. The honest summary: this is mostly an engineering paper — a PyTorch toolkit bundling 17 attribution methods, a robustness module with adversarial attacks as update paths, and evaluation metrics spanning image, text, multimodal, and object detection. That part is real and likely useful to practitioners. The claimed theoretical unification, however, is the weak link.\n\nWhat is actually new is the framework architecture and the integration of adversarial attacks as pluggable update methods inside a common API. Equation 6 is the standard continuous path-integral formulation of Integrated Gradients; nothing novel there. The theorem that any update method preserves Completeness and Sensitivity is not proven for the implemented algorithm. The proof in Section III-B0b treats the first-order Taylor expansion as exact for finite steps (Eq. 7), which only holds for affine models or in the infinitesimal-step limit. Algorithm 1 computes a discrete Riemann sum; without a step-size condition or error bound, the telescoping argument does not deliver f(x_T) - f(x_0) = sum_j A_j. The Sensitivity argument is also weaker than the axiom: it shows at least one attribution is nonzero, not that the specific changed feature gets nonzero attribution. The impossibility-theorem discussion in Section III-B0d is muddled — failing Linear is not a way to avoid impossibility results.\n\nThat said, the experiments are broad: five image models on ImageNet, text on IMDb/Yelp, CLIP for multimodal, RetinaNet for object detection, plus a robustness study swapping fourteen attack methods as update paths. No error bars, no ablations on T or epsilon, and no reproducibility configuration details, so the empirical claims are suggestive rather than rigorous. The code is available, which is a plus.\n\nBottom line: the framework may be a useful toolkit, but the theoretical guarantee as stated is overstated. A serious revision should restrict the axiom claims to the continuous limit, provide convergence conditions for the discrete algorithm, correct the Sensitivity proof, and clean up the impossibility theorem discussion. I would send it to peer review — the engineering is substantial enough that referees can push for these fixes — but I would not cite it for the theory.","headline":"A practical XAI toolkit whose central theoretical guarantee is overstated: the completeness proof only holds in the continuous limit, not for the discrete algorithm actually implemented.","tokens_in":19718,"tokens_out":2902,"would_cite":false,"duration_ms":28841,"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 claims that any user-supplied update rule inside its Fundamental Attribution Method loop satisfies Completeness and Sensitivity.","keywords":["attribution-based explainability","fundamental attribution methods","integrated gradients","attribution axioms","adversarial robustness","explainability framework","multimodal interpretability","path integral attribution"],"falsifier":"Run Algorithm 1 on the scalar model $f(x)=x^2$ with baseline $x'=0$, target $x=1$, and one update step $\\Delta x=1$. The loop returns $A = \\Delta x \\cdot f'(0) = 0$, while Completeness demands $A = f(1)-f(0) = 1$. Reproducing that mismatch, or sweeping step counts on ImageNet and plotting the residual $|\\sum_j A_j - (f(x)-f(x'))|$, would settle where the claimed universal guarantee actually holds.","tokens_in":1737,"feed_emoji":"🔍","tokens_out":4459,"duration_ms":125809,"temperature":0.7,"pith_summary":"The paper sets out to show that path-based attribution methods share one underlying operation: walk from a reference input to the target input using some update rule, and accumulate the dot product of each step with the model's gradient. It calls this operation the Fundamental Attribution Method and claims that every instantiation, no matter what update rule is used, satisfies the key attribution axioms. The payoff, if the claim holds, is that researchers can design new explainability methods by choosing an update strategy alone, without re-proving axiom compliance, and can test robustness and faithfulness in the same framework. The paper further packages this idea as ABE, a modular toolkit covering image, text, multimodal, and object-detection tasks with 17 attribution methods and validation metrics.","feed_headline":"One generic loop lets any update rule inherit the attribution axioms","feed_subtitle":"Researchers can plug novel update strategies into one generic loop and keep the axiomatic guarantees.","key_machinery":"The central object is the Fundamental Attribution Method, an iterative loop that unifies attribution algorithms. It maintains an attribution accumulator $A_t$ and an input state $x_t$; at each step an update rule produces $\\Delta x_{t-1}$, the gradient $\\nabla f(x_{t-1})$ is dotted into the accumulator, and the state advances. Equation 6 is the continuum form of this loop, and the proof machinery is the first-order Taylor identity in Eq. 7 plus the telescoping cancellation in Eq. 9. A second piece of machinery is the $\\ell^\\infty$ sign-gradient argument, which shows that $\\epsilon\\,\\mathrm{sign}(\\nabla_x f(x))$ is the step of fixed per-feature size that maximizes the first-order output change, justifying adversarial attacks as recommended update rules.","core_discovery":"The central discovery is a reduction: path-based attribution methods can be expressed as one arithmetic identity. Equation 6 defines the sum of feature attributions as $\\sum_{j=1}^{d} A_j(x,x') = \\int \\Delta x_t \\cdot \\frac{\\partial f(x_t)}{\\partial x_t}\\,dt$, where $\\Delta x_t$ comes from any user-supplied update method, and Algorithm 1 computes the same quantity by accumulating $\\Delta x_{t-1}\\cdot \\nabla f(x_{t-1})$. The authors argue that the terms telescope through a first-order Taylor expansion, so the attribution sum equals $f(x)-f(x')$, which is the Completeness property, and any change in output forces at least one nonzero feature attribution, which is Sensitivity. Implementation Invariance follows because only input-output gradients are accumulated, not internal architecture. The framework then treats this loop as the shared skeleton of 17 implemented methods, with update rules ranging from linear baseline interpolation to adversarial attack paths.","pith_inferences":["The universal guarantee is strictly about the continuum integral; for finite step counts the implementation inherits only an approximate Completeness, so the framework's practical promise would be stronger if it exposed a step-size or residual diagnostic.","The path-agnostic formulation suggests treating the update rule as a design object: one could learn or search over update paths that maximize insertion score while checking the completeness residual, effectively turning attack methods into tunable explanation generators.","Because the loop absorbs any update rule, the same machinery could extend to non-gradient update directions by substituting finite-difference gradient estimates, at the cost of approximate axioms.","If adversarial update rules really are axiom-compliant, then robustness evaluation and explanation generation become the same computation, which would tie attack-transferability research to explanation-consistency research more tightly."],"forward_implications":["If the proof is right, a researcher adding a new attribution method to ABE only needs to supply an update rule; Completeness and Sensitivity are inherited from the loop.","The robustness module doubles as an attack library and an attribution generator, so the same adversarial perturbation used to test a model can define the explanation path for that model.","Axiom-compliant methods such as ISA, AttEXplore, MFABA, AGI, BIG, IG, EG, and FIG can be compared under a common validation interface, while faster non-axiomatic methods remain available for resource-constrained use.","The framework's use of the impossibility theorems implies that adversarially updated attributions, because they violate the Linear property, can outperform random feature selection and avoid the triviality that affects linear-complete attribution rules.","Multimodal and object-detection tasks inherit the same loop through task-specific wrappers and metrics such as ICD/ICI, TCD/TCI, INS/DEL, and FPS, making attribution axioms portable across modalities."],"supporting_citations":[{"why":"Defines the Sensitivity and Implementation Invariance axioms and the Integrated Gradients path method that Eq. 6 generalizes.","marker":"[2]"},{"why":"Shows baselines and path selection can be made dynamic; the unified loop absorbs this as one update strategy.","marker":"[3]"},{"why":"Introduces adversarial gradient integration, the first robustness-aware attribution path the framework builds on.","marker":"[5]"},{"why":"Provides the boundary-point computation used by the adversarial attribution methods benchmarked in ABE.","marker":"[6]"},{"why":"Contributes cross-model path exploration and frequency-domain information used as an update strategy and comparison method.","marker":"[7]"},{"why":"Supplies expected gradients over multiple baselines, an example of a custom update and baseline strategy captured by the loop.","marker":"[22]"},{"why":"Provides the fast path-integration approximation that motivates trading step count for speed inside the same loop.","marker":"[23]"},{"why":"Gives impossibility theorems for feature attribution; the paper invokes its Linear property to argue adversarial update rules avoid trivial attributions.","marker":"[26]"}],"fun_headline_variants":["One arithmetic loop makes any attribution method satisfy all axioms","Single telescoping identity turns any update rule into an attribution","ABE reduces 17 attribution methods to one arithmetic loop","Any update rule becomes an attribution method with axioms intact","One identity maps any path-based attribution to a guaranteed sum"],"cache_read_input_tokens":21888,"weakest_assumption_plain":"The proof assumes the first-order Taylor expansion in Eq. 7 is exact at every finite update step, so the discrete accumulation in Algorithm 1 equals the continuous path integral in Eq. 6; for a general nonlinear model with large update steps, leftover second-order terms can break the equality and break Completeness.","fun_headline_variants_meta":{"raw":{"variants":["One arithmetic loop makes any attribution method satisfy all axioms","Single telescoping identity turns any update rule into an attribution","ABE reduces 17 attribution methods to one arithmetic loop","Any update rule becomes an attribution method with axioms intact","One identity maps any path-based attribution to a guaranteed sum"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00099,"raw_usage":{"total_tokens":4172,"prompt_tokens":894,"completion_tokens":3278,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":3199}},"tokens_in":510,"tokens_out":3278,"duration_ms":22565,"temperature":1.0,"reasoning_tokens":3199,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:10:28.210216+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on the scalar model $f(x)=x^2$ with baseline $x'=0$, target $x=1$, and one update step $\\Delta x=1$. The loop returns $A = \\Delta x \\cdot f'(0) = 0$, while Completeness demands $A = f(1)-f(0) = 1$. Reproducing that mismatch, or sweeping step counts on ImageNet and plotting the residual $|\\sum_j A_j - (f(x)-f(x'))|$, would settle where the claimed universal guarantee actually holds.","supporting_citations":[{"cited_title":"Explaining deep neural network models with adversarial gradient integration,","cited_arxiv_id":null,"evidence_quote":"Introduces adversarial gradient integration, the first robustness-aware attribution path the framework builds on."},{"cited_title":"Mfaba: A more faithful and accelerated boundary-based attribution method for deep neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the boundary-point computation used by the adversarial attribution methods benchmarked in ABE."},{"cited_title":"Attexplore: Attribution for explanation with model parameters exploration,","cited_arxiv_id":null,"evidence_quote":"Contributes cross-model path exploration and frequency-domain information used as an update strategy and comparison method."},{"cited_title":"Fast axiomatic attribution for neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the fast path-integration approximation that motivates trading step count for speed inside the same loop."},{"cited_title":"Impossibility theo- rems for feature attribution,","cited_arxiv_id":null,"evidence_quote":"Gives impossibility theorems for feature attribution; the paper invokes its Linear property to argue adversarial update rules avoid trivial attributions."}],"review_version":1}