{"id":"1b0d99b0-8119-4751-99eb-51d602001cd6","arxiv_id":"2411.12643","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"DLBacktrace is a backward relevance-propagation technique that resembles existing layer-wise relevance propagation (LRP) but is presented as novel, with benchmarks that omit the closest competitor.","lead":"This paper introduces DLBacktrace, an explainability method that propagates importance scores backward from a model's output to its inputs. It claims to outperform SHAP, LIME, and GradCAM on finance, image, and sentiment tasks, but the evidence is weakened by missing baselines and unverified code.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The model-agnostic claim breaks at the attention extension: §3.4's propagation formulas are mathematically undefined for the actual attention computation, so the BERT/LLM support on which the universal claim rests is not established.","rationale":"The reader's weakest assumption — that the relevance redistribution rule is a heuristic that lacks a valid extension to nonlinear, normalized, and attention-based layers — is exactly where the central claim fails. I have sharpened it to the specific, checkable claim that Algorithm 3 is not a well-defined propagation rule for the attention computation softmax(QK^T/√d)V. If the shapes in Eqs. (10)–(13) do not match or the relevance is not conserved, then the paper's advertised support for BERT, Llama, and 'any deep learning models' is unsupported by the stated mathematics. This is a soundness issue, not merely a disagreement with the field's consensus: the paper itself asserts that the method is deterministic, model-agnostic, and that all relevance loss is due to bias, and none of those assertions is defensible without a valid attention rule. The empirical results also contain an internal contradiction with the abstract's blanket 'significantly outperforms' claim — in Table 3, IG achieves MoRF AUC 6.723 versus DLBacktrace's 15.431 and Delta AUC 53.479 versus 15.259 — which independently weakens the superiority claim, but the missing attention derivation is the more load-bearing obstruction because it prevents the method from being evaluated at all on transformer models. The open-sourced library is a positive step, but it does not repair the missing mathematical definition; the paper needs to state the actual propagation rules used for softmax, multi-head concatenation, residual connections, and layer normalization before the benchmark numbers can be trusted. For these reasons the reader's reject verdict stands.","tokens_in":13303,"tokens_out":7694,"duration_ms":74735,"concrete_test":"Implement a single-head attention block with d_model=8, seq_len=4, random fixed weights. Run Algorithm 3 with r_y = 1 at the output. Check (a) whether every operation in Eqs. (10)–(13) has matching tensor shapes; (b) whether the sum of relevance assigned to the input token x equals 1 exactly (up to machine precision) through the full block, including softmax. If the shapes fail or conservation fails, the attention extension is not a defined propagation rule and the BERT/LLM claims in §5.3.3 rest on an unspecified or approximate computation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DLBacktrace's defining claim is that relevance can be propagated 'from the output back to the input' for any deep learning model. The validity condition is that the redistribution rule Eq. (3) — and its counterparts for other layers — exactly account for the forward computation. Even for dense layers this is already fragile: Eq. (3) distributes relevance in proportion to signed pre-activations W_ij x_ij, which is exact only for a linear map; for ReLU, softmax, layer-norm, or attention, the activation discards or renormalizes parts of the pre-activation, and the saturation branch of Eq. (3) responds by setting some terms to 0 rather than attributing the lost relevance to the bias or saturated unit. The paper's statement that 'any loss of relevance during propagation is due to network bias' is therefore false. The gap becomes decisive for the attention rule in Algorithm 3. Eq. (10) defines rQK = (rO · xV) · xQK, but rO, xV, and xQK have incompatible shapes under standard conventions (rO, xV ∈ R^{B×S×D}, xQK ∈ R^{B×H×S×S}); the formulas (10)–(13) do not recover the softmax weights, do not specify the contraction, and are never derived. Since the advertised support for BERT and LLMs depends entirely on this rule, the central 'model-agnostic, any deep learning model' claim lacks a defined algorithm for the very architectures used to demonstrate it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DLBacktrace, a deterministic, backpropagation-style relevance attribution method for deep networks. It presents propagation rules for dense layers in default and contrastive modes, a proposed extension to attention layers, and benchmarks on tabular (Lending Club MLP), image (CIFAR-10 ResNet-34), and text (SST-2 BERT) tasks against LIME, SHAP, Grad-CAM, Vanilla Gradient, SmoothGrad, and Integrated Gradients. The authors claim that DLBacktrace is model-agnostic and 'significantly outperforms traditional methods ... across all key metrics,' while also acknowledging in Section 5.3.3 that Integrated Gradients is the strongest method on the text benchmarks.","tokens_in":13612,"tokens_out":6010,"duration_ms":60113,"significance":"If the method worked as claimed, it would offer a useful deterministic, baseline-free alternative to gradient- and perturbation-based explanation, with attractive properties for live deployment. Strengths include the absence of auxiliary models or baselines, the deterministic nature, open-source availability, and the breadth of attempted modalities. However, the current manuscript does not establish the central claims: the attention propagation algorithm is mathematically underspecified, relevance conservation is asserted without proof, and the reported benchmarks either use nonstandard metrics without error bars or directly contradict the headline claim. The contribution is therefore not yet ready for publication.","major_comments":[{"comment":"The attention-layer rule is not well defined. With standard shapes rO, xV ∈ R^{B×S×D} and xQK ∈ R^{B×H×S×S}, the dot products in Eqs. (10)–(13) are undefined; the softmax nonlinearity is never handled; and the claim after Algorithm 3 that 'any loss of relevance during propagation is due to network bias' is false for softmax and layer-norm. Because the BERT/LLM demonstrations rest entirely on this rule, the paper's central 'any deep learning models' claim is not established by a defined algorithm. The authors must either supply a complete, dimensionally consistent attention rule with a derivation or drop the transformer/LLM claims.","section":"§3.4, Algorithm 3, Eqs. (10)–(14)"},{"comment":"The dense-layer redistribution rule also lacks a conservation argument. For a ReLU or saturating activation, the branch that sets terms to zero discards pre-activation relevance without transferring it to the bias or to saturated units; the statement that all lost relevance is due to network bias is therefore not generally true. Since the method's explanatory value depends on exactness of redistribution, the authors need a proof of conservation (or a precise statement of when it holds) before claiming the attributions are faithful.","section":"§3.3.1, Eq. (3)"},{"comment":"The reported text results contradict the paper's headline. In Table 3, Integrated Gradients has the best MoRF AUC (lower), LeRF AUC (higher), and Delta AUC, while DLBacktrace is mid-ranked; Section 5.3.3 explicitly acknowledges IG's superiority. The abstract and Section 5.3.2's claim of 'significantly outperforms ... across all key metrics' is therefore not supported by the paper's own data. The claims need to be restricted to the metrics and modalities where they hold.","section":"§5.3.3, Table 3"},{"comment":"The evaluation is not statistically grounded. MPRT and Complexity are introduced in this paper without validation as established metrics; no error bars or significance tests are given for any table; and Table 2 reports Max-Sensitivity values computed from a 'non-infinite mean' after discarding infinite values, a nonstandard procedure that can arbitrarily favor a method with finite outputs. The comparison should be repeated with standard XAI metrics (e.g., from Quantus) or with a principled justification and confidence intervals for the new metrics.","section":"§5.2.1, Table 1 and §5.2.2, Table 2"},{"comment":"Sample sizes and experimental detail are insufficient for the strength of the claims: only 1,024 tabular, 500 image, and the SST-2 evaluation set are used; no seeds, hyperparameters, or perturbation protocols are reported; and the Lending Club/ResNet/BERT checkpoints are not linked. This makes the claimed superiority non-reproducible from the manuscript alone.","section":"§5.1"}],"minor_comments":[{"comment":"The reference 'Figure ??' is unresolved; it should be replaced with an actual figure reference or removed.","section":"§4"},{"comment":"The cases in Eq. (3) are not mutually exclusive: the first and second branches both apply when W_ij x_ij > 0, and the third and fourth branches both apply when W_ij x_ij < 0. The intended precedence should be stated explicitly.","section":"Eq. (3)"},{"comment":"The input/output notation is confusing: the algorithm is declared with Input x and Output ry, but Step 3 computes rO from the linear projection layer without defining how rO relates to x and y. Clarify the variable mappings and the contraction operations in Eqs. (10)–(13).","section":"Algorithm 3"},{"comment":"There are typographical errors in possessive forms: 'it’s usage' should be 'its usage' in both sections.","section":"§8 and §9"},{"comment":"The cross-reference 'as illustrated in Fig. 13' points to an appendix figure; use the main-text figure number or fix the cross-reference.","section":"§5.3.3"}],"recommendation":"reject","confidential_remarks":"The manuscript is closer to a software release note than a validated research contribution. The central model-agnostic claim depends on an attention rule that is not mathematically defined, and the evaluation tables contradict the abstract's superiority claim. These are load-bearing issues that cannot be fixed by local edits. I would encourage the authors to derive a proper attention propagation rule, validate it on established metrics with error bars, and resubmit a substantially revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read DLBacktrace. Bottom line: it's Layer-wise Relevance Propagation re-derived without attribution, plus an attention extension that doesn't parse. The universal claim collapses because §3.4's formulas (Eqs. 10–13) use tensors with incompatible shapes—rO and xV are (batch, seq, d_model) while xQK is (batch, heads, seq, seq); there's no defined contraction, and the softmax is never recovered. The note that any loss of relevance is due to network bias is false. So for BERT/LLM the method isn't even defined on paper.\n\nThat said, the library is real, open source, and the authors clearly put effort into a clean workflow across tabular, image, and text. The contrastive mode is a small but useful addition. They also honestly report in Table 3 that Integrated Gradients beats DLBacktrace on all text metrics—which makes the abstract's blanket superiority claim self-contradictory.\n\nThe core propagation rule (Eq. 3) is the LRP alpha-beta decomposition for signed pre-activations. The paper never cites LRP (e.g., Bach et al. 2015) and never benchmarks against it. That's a novelty problem, not just a citation lapse. For dense networks, the method reduces to a known algorithm with a couple of tweaks. That might be fine for a library paper, but this is positioned as a new technique.\n\nMetrics: MPRT, Complexity, and ToPEQ are self-defined and appear without error bars. The image metrics are standard, though some entries (e.g., GradCAM sensitivity 1070) look off. No statistical support for the claimed outperformance.\n\nMy take: the paper is not ready as a research contribution. It needs a rewrite that acknowledges LRP, fixes or removes the attention section, and replaces the self-defined metrics with established ones. As is, I would not send it to referees; the load-bearing claim is unsupported. If the authors resubmit after addressing these, it could be a modest practitioner-oriented toolkit paper. For now, desk reject with an invitation to revise.","headline":"DLBacktrace is LRP under a new name, and its attention extension doesn't type-check; the empirical claims conflict with the paper's own table.","tokens_in":14137,"tokens_out":4241,"would_cite":false,"duration_ms":40674,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DLBacktrace claims a single, baseline-free backtrace gives faithful feature attribution for MLPs, CNNs, and transformer LLMs.","keywords":["explainable AI","feature attribution","relevance propagation","model-agnostic interpretability","deep learning","transformer attention","deterministic explanations"],"falsifier":"For a two-layer ReLU network with known weights and bias, compute the sum of input relevances from DLBacktrace and compare it to the output relevance after subtracting bias shares; if the two diverge for a range of inputs, the rule is lossy. For the attention formulas, scale $Q$ and $K$ by a constant and check whether $r_{QK}$ remains invariant; a change would show the stated equations are dimensionally unstable.","tokens_in":13119,"feed_emoji":"🔍","tokens_out":9209,"duration_ms":83950,"temperature":0.7,"pith_summary":"DLBacktrace is a proposed explainability method that traces a relevance score from a model's output back to its input, splitting each neuron's contribution into positive and negative parts according to the signed pre-activation products $W_{ij}x_{ij}$. The authors claim this gives deterministic, baseline-free local explanations that work across MLPs, CNNs, and transformer-based LLMs, with global importance obtained by averaging normalized local scores. They benchmark the method against LIME, SHAP, Grad-CAM, Vanilla Gradient, SmoothGrad, Integrated Gradients, and attention-rollout on tabular, image, and text tasks. If the central claim holds, a single method would provide faithful feature attribution for high-stakes applications without needing auxiliary white-box models or baseline samples.","feed_headline":"Backtrace method scores every feature in MLPs, CNNs, and BERT","feed_subtitle":"A deterministic, baseline-free way to trace a prediction to its inputs, benchmarked on tabular, image, and text.","key_machinery":"The carrying mechanism is the proportional relevance-distribution rule of Section 3.3: relevance arriving at a neuron is split into positive, negative, and bias shares in proportion to the positive pre-activation sum $X_{pj}$, the negative pre-activation sum $X_{nj}$, and the bias $b_j$, then distributed to the inputs in proportion to each signed product $W_{ij}x_{ij}$ within the same sign class. This rule is what makes the method deterministic and independent of sampling or auxiliary models, since it depends only on the model's weights, activations, and biases. For attention, the paper adds a separate recipe that routes relevance through the $QK^T$, $V$, $Q$, and $K$ outputs using elementwise products, treating any loss as bias.","core_discovery":"On its own terms, the paper claims that DLBacktrace is a model-agnostic explainability technique that reads only the trained network and the single test sample, requiring no auxiliary model, no background dataset, and no baseline input. The algorithm builds a graph from the network weights and propagates output relevance backward in a breadth-first manner. For each neuron, the total pre-activation is $T_j = X_{pj}+X_{nj}+b_j$, and the neuron's relevance is divided into positive, negative, and bias shares $R_{pj}=X_{pj}/T_j r_{yj}$, $R_{nj}=X_{nj}/T_j r_{yj}$, $R_{bj}=b_j/T_j r_{yj}$; these shares are then passed to each input unit in proportion to the signed contribution $W_{ij}x_{ij}$, with zeroing under saturation. A Contrastive Mode tracks separate positive and negative relevance streams, and Algorithm 3 extends the same idea to attention layers by distributing relevance through elementwise products of the attention outputs. The reported benchmarks show DLBacktrace with the best numbers on tabular stability (MPRT) and on image faithfulness, sensitivity, and pixel-flipping, while on the SST-2 text benchmark Integrated Gradients achieves the best scores and DLBacktrace is competitive rather than best.","pith_inferences":["A natural next test the paper does not run is to check relevance conservation on deeper nonlinear networks; the stated claim that any loss is due to bias is a testable identity that would settle whether the split rule is exact.","The attention formulas are presented without derivation, so a reader cannot yet tell whether they are correct for multi-head attention; a toy single-head attention model with known token contributions would resolve this.","The claim of universal model-agnosticism extends beyond the tested set of MLP, ResNet, U-Net, BERT, and one Llama demo; recurrent and multimodal architectures are plausible targets but remain unverified.","If the proportional rule holds, DLBacktrace could be a cheap way to compute counterfactual-style positive and negative attributions for bias detection in tabular models, though the paper only sketches that use."],"forward_implications":["One implementation of DLBacktrace could replace separate tabular, image, and NLP explainers with a single deterministic backtrace.","Layer-wise relevance, including bias-to-input ratios and saturation, could support network pruning and bias audits without retraining.","Global feature importance can be computed by averaging normalized local scores, giving a Shapley-like ranking without the cost of sampling.","In regulated settings, the deterministic and baseline-free nature would make explanations reproducible for audits.","The reported improvements on image metrics would mean DLBacktrace is more faithful to a ResNet's actual behavior than Grad-CAM, vanilla gradient, SmoothGrad, or Integrated Gradients."],"supporting_citations":[{"why":"Defines LIME, the local surrogate baseline compared against on tabular and text data.","marker":"[10]"},{"why":"Defines SHAP, the Shapley-value baseline used in the tabular comparisons.","marker":"[11]"},{"why":"Defines Integrated Gradients, a gradient-based baseline compared on image and text tasks.","marker":"[12]"},{"why":"Defines Grad-CAM, the gradient-heatmap baseline compared on image and text tasks.","marker":"[18]"},{"why":"Defines the vanilla gradient saliency baseline used in the image benchmark.","marker":"[19]"},{"why":"Defines SmoothGrad, the noise-averaged gradient baseline compared on image and text tasks.","marker":"[20]"},{"why":"Supplies the evaluation toolkit behind the image-modality faithfulness, sensitivity, and pixel-flipping metrics.","marker":"[27]"},{"why":"Defines the attention mechanism and multi-head attention that Section 3.4 extends for relevance propagation.","marker":"[29]"},{"why":"Introduces the token-perturbation LeRF, MoRF, and Delta AUC metrics used for the text benchmark.","marker":"[32]"}],"fun_headline_variants":["DLBacktrace: one explainer for MLPs, CNNs, and BERT","No baseline needed: DLBacktrace traces any deep model","Model-agnostic explainability: DLBacktrace works everywhere","Backtrace any deep net without extra data or gradients","From tabular to text: DLBacktrace explains it all"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a neuron's relevance can be split into positive and negative parts in exact proportion to the signed pre-activation inputs $W_{ij}x_{ij}$ across every layer, including nonlinear, normalized, and attention layers, but the paper gives no proof that this split preserves the model's actual computation.","fun_headline_variants_meta":{"raw":{"variants":["DLBacktrace: one explainer for MLPs, CNNs, and BERT","No baseline needed: DLBacktrace traces any deep model","Model-agnostic explainability: DLBacktrace works everywhere","Backtrace any deep net without extra data or gradients","From tabular to text: DLBacktrace explains it all"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000372,"raw_usage":{"total_tokens":2028,"prompt_tokens":1024,"completion_tokens":1004,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":916}},"tokens_in":640,"tokens_out":1004,"duration_ms":9703,"temperature":1.0,"reasoning_tokens":916,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:17:55.906635+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a two-layer ReLU network with known weights and bias, compute the sum of input relevances from DLBacktrace and compare it to the output relevance after subtracting bias shares; if the two diverge for a range of inputs, the rule is lossy. For the attention formulas, scale $Q$ and $K$ by a constant and check whether $r_{QK}$ remains invariant; a change would show the stated equations are dimensionally unstable.","supporting_citations":[{"cited_title":"Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, and Dhruv Batra","cited_arxiv_id":null,"evidence_quote":"Defines Grad-CAM, the gradient-heatmap baseline compared on image and text tasks."}],"review_version":1}