{"id":"04512cea-c7f0-40c7-b60e-79e5302e0afe","arxiv_id":"2412.00800","paper_version":2,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":1.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey-style XAI book with code examples, covering standard interpretability methods and models, but no new scientific contributions.","lead":"This preprint is a book-length guide to explainable AI, from classical models to LLMs, with Python examples. It compiles existing techniques and application case studies rather than presenting new research findings.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4.3.1's 'hidden state' plot is actually the model's output, not RNN hidden states; this code error directly undermines the book's central claim to equip readers with practical XAI skills.","rationale":"The reader's weakest_assumption correctly identifies that the code examples must be reliable for the guide's central educational claim to hold, and the RNN example in §4.3.1 is a clear, verifiable violation. I agree with that identification and found no more load-bearing issue: the book is explicitly a tutorial/review, not a research contribution, so its novelty is intentionally low and its correctness risk centers on practical reproducibility. The code error is concrete and independently checkable, and it directly affects the 'practical skills' promise. Because the reader already issued UNVERDICTED (neither accepting nor rejecting a non-research guide), my finding does not move the verdict; it reinforces the same conclusion. I therefore recommend UNCHANGED, while flagging that this specific error should be corrected in any revised version.","tokens_in":48405,"tokens_out":2363,"duration_ms":22983,"concrete_test":"Run the Section 4.3.1 code verbatim and print `hidden_states.shape` and `np.allclose(hidden_states, y_pred)`. If the shape is (1, 100, 1) and the arrays are equal, the 'hidden states' are in fact the Dense output. Then modify the code to extract the SimpleRNN layer's actual outputs, e.g., `Model(inputs=model.input, outputs=model.layers[0].output).predict(X)`, and regenerate Figure 4.4. If the new plot shows 10 distinct traces that differ from the single output trace, the book's code is confirmed incorrect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The book's central claim (Abstract, §1.4) is that it provides a comprehensive, practical guide to XAI, with Python code examples that bridge theory and practice. That claim is load-bearing on the correctness of the code examples. Section 4.3.1 violates this: the code defines a SimpleRNN followed by a Dense layer (lines 15–18) and then computes both `y_pred` and `hidden_states` as `model.predict(X)` (lines 24–25). These are identical calls, so `hidden_states` is the Dense output, not the RNN hidden-state activations. Concretely, `hidden_states.shape` is (1, 100, 1), so the loop `for i in range(hidden_states.shape[-1])` draws only one trace, not the activations of all 10 hidden units as the figure caption claims. A reader following the code would not learn how to visualize hidden states; they would see the model's prediction plotted under a misleading label. This is not a stylistic issue: it is a factual error in the core deliverable of the guide, and it undermines the promise of 'hands-on understanding' in §1.4. The same pattern—code that does not implement what the text claims—is the weakest point in an otherwise survey-style resource.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a book-length survey/guide to Explainable AI, covering theoretical foundations, interpretability of classical models (decision trees, linear models, SVMs, rule-based systems, GAMs, Bayesian models), deep learning models (CNNs, RNNs, Transformers), large language models (BERT, GPT, T5, LLaMA), a wide range of XAI techniques (SHAP, LIME, Integrated Gradients, Grad-CAM, LRP, counterfactuals, causal methods, graph and multimodal methods), applications, evaluation metrics, tools, and future directions. The stated goal, in the Abstract and §1.4, is to provide a comprehensive and practical guide that 'bridges theory and practice' through Python code examples and a companion GitHub repository.","tokens_in":48609,"tokens_out":4489,"duration_ms":37694,"significance":"If the code examples and organizational structure were reliable, this would be a useful educational resource: the breadth is substantial, and the practical emphasis with runnable Python snippets for SHAP, LIME, Integrated Gradients, DeepLIFT, Grad-CAM, LRP, and several counterfactual methods is a genuine strength. The book also gives credit to many standard references and includes a counterpart GitHub repository, which supports reproducibility in principle. However, the central value proposition—a hands-on, trustworthy guide—depends critically on the correctness of the code and the coherence of the presentation, and both are currently undermined by the errors described below.","major_comments":[{"comment":"The code defines a SimpleRNN followed by a Dense layer (lines 15–18) and then computes both `y_pred` and `hidden_states` as `model.predict(X)` (lines 24–25). These are identical calls, so `hidden_states` is the Dense output, not the RNN hidden-state activations. Concretely, `hidden_states.shape` is (1, 100, 1), so the loop `for i in range(hidden_states.shape[-1])` draws only one trace, not the activations of all 10 hidden units as the figure caption claims. A reader following the code would not learn how to visualize hidden states; they would see the model's prediction plotted under a misleading label. This error directly contradicts the promise in §1.4 of providing 'hands-on understanding' and must be corrected, for example by building a separate model that outputs `model.layers[0].output`.","section":"§4.3.1, lines 24–25"},{"comment":"The subsection titled 'Advantages and Disadvantages of Logistic Regression' appears in the middle of the Decision Trees section, immediately after the discussion of pruning and before the decision-tree feature-importance code example. This is a structural error: the content belongs in §3.3, which already contains a logistic regression discussion, and its placement here is confusing for a reader using the book as a reference. The guide's claim to be comprehensive and reliable is weakened by such organization mistakes, and this passage should be relocated or removed as a duplicate.","section":"§3.2"},{"comment":"The book gives inconsistent definitions of its two central terms. In §1.2, 'interpretability' is defined as 'the degree to which a human can understand the cause of a decision,' and 'explainability' is defined as 'the extent to which the internal mechanics of a machine learning model can be understood,' with explainability said to go further by focusing on 'why.' In §5.4.3, however, 'interpretability' is described as understanding internal workings (e.g., neurons and layers), while 'explainability' is described as providing human-understandable reasons such as feature importance or visualizations. These are essentially swapped. Because the book explicitly introduces these terms as foundational concepts, the inconsistency is pedagogically misleading and should be reconciled in a revision.","section":"§1.2 vs. §5.4.3"}],"minor_comments":[{"comment":"Lines 8 and 14 of the SHAP example appear in the manuscript text as bare sentences ('Load the dataset', 'Split the dataset...') without the '#' comment prefix; if reproduced as written, the code is syntactically invalid. Please ensure all comment lines are properly prefixed in the printed code.","section":"§2.4.1, code listing"},{"comment":"The text states that the layer-activation plot shows 'all 13 layers of BERT,' but BERT-base has 12 transformer layers; the 13 hidden states returned by `output_hidden_states=True` include the embedding layer. The wording should be clarified to avoid confusion.","section":"§5.6.2"},{"comment":"The attention-heatmap example uses a synthetic 3×3 weight matrix, but the preceding text discusses the phrase 'The cat sat' and alignment in machine translation; the figure caption and the result explanation should make clear that the heatmap is illustrative and not computed from a real model on that phrase.","section":"§4.4.3"},{"comment":"The claim that GPT-4 supports multimodal inputs is referenced to a generic citation [102]; a primary source or clearer specification of the model version would be more helpful to readers.","section":"§5.2, GPT-4 bullet"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a survey-style book with a very large author list, and it does not claim novel research contributions. Its main value would be pedagogical, but the reviewing standard for a comprehensive guide should still require that the code examples are runnable and that the text is internally consistent. The RNN hidden-state bug and the misplaced logistic-regression section are concrete, fixable problems; the definitional inconsistency is also fixable. I would encourage the editor to ask the authors to verify every code listing—ideally by executing the companion repository—and to perform a systematic editorial pass for structural and terminological consistency before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a survey, not a research paper. It covers a lot: decision trees, SHAP, LIME, Grad-CAM, LLM interpretability, even causal inference and counterfactuals. The organization is sensible, the citations look standard, and many of the code examples are clear and reproducible. If you want a quick map of the XAI landscape, this is a reasonable place to start.\n\nThe problems are where the book makes its strongest promise: hands-on practical skills. The stress-test concern is real. In Section 4.3.1 the code defines a SimpleRNN followed by a Dense layer, then sets both y_pred and hidden_states to model.predict(X). Those calls are identical, so 'hidden_states' is the Dense output, not the RNN hidden-unit activations. The plot labeled 'Activations of All Hidden Units Over Time' actually draws a single trace of the model's prediction. A reader following along would not learn how to visualize hidden states. That is a genuine bug in a core deliverable.\n\nThere are also structural slips. The 'Advantages and Disadvantages of Logistic Regression' block appears inside the Decision Trees section (3.2), and the same block is duplicated in 3.3. Some chapters, like Applications and Tools, are little more than lists; they don't go deep enough to be called comprehensive. These are soft spots, but they are fixable.\n\nThe theoretical descriptions of standard XAI methods are largely correct. Shapley values, LIME's local surrogate objective, Integrated Gradients, and Grad-CAM are all presented accurately. The authors are not inventing anything, but they are also not misrepresenting established work. The citation pattern seems honest; nothing looks self-serving or circular.\n\nThe bottom line: as a research contribution, this is a zero. As a textbook draft, it is usable but needs revision. The code error in 4.3.1 should be corrected, the misplaced Logistic Regression material moved, and the thin chapters either expanded or honestly labeled as overviews. I would send it to a serious referee, not for the science, but because it has the potential to be a genuinely useful educational reference once the errors are fixed.","headline":"A broad, mostly sound XAI survey whose practical value is undercut by a handful of fixable code and structural errors.","tokens_in":49157,"tokens_out":1743,"would_cite":false,"duration_ms":18570,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This book-length guide argues that explainable AI is a teachable, unified field spanning classical interpretable models, post-hoc attribution methods, and large-language-model probing, and supports the argument with Python examples.","keywords":["Explainable AI","interpretability","feature attribution","SHAP","LIME","Grad-CAM","large language models","counterfactual explanations"],"falsifier":"Run the Section 4.3.1 RNN code and inspect what `model.predict(X)` returns: if the plotted curves are the network's final predictions rather than the 10 hidden-unit activations the caption claims, the example does not demonstrate hidden-state interpretation. More generally, executing each chapter's code and checking that the printed outputs and figures match the described behavior would settle whether the practical promise holds.","tokens_in":48232,"feed_emoji":"📘","tokens_out":3988,"duration_ms":71219,"temperature":0.7,"pith_summary":"This work is a book-length guide attempting to establish that explainable AI (XAI) can be presented as one coherent curriculum, from intrinsically interpretable models (decision trees, linear models, GAMs, Bayesian models) through post-hoc attribution techniques (SHAP, LIME, Integrated Gradients, Grad-CAM, LRP) to the special interpretability challenges of CNNs, RNNs, and large language models. A sympathetic reader would take the central claim to be that a reader with basic machine learning knowledge can, by following this guide, acquire both the theoretical distinctions (transparency, interpretability, explainability, fairness) and the practical skills needed to explain real models. The book also claims that the same XAI toolkit applies across healthcare, finance, and policy, and that explanation quality can be assessed with metrics like fidelity, stability, and faithfulness. If this is right, the volume is a practical educational resource rather than a research monograph, and its value lies in how well its code examples and organization deliver that curriculum.","feed_headline":"XAI guide spans classical models to LLMs","feed_subtitle":"A Python-backed tour of SHAP, LIME, Grad-CAM, counterfactuals, and LLM probing methods.","key_machinery":"The organizing machinery is a two-axis view of interpretability: intrinsic versus post-hoc, and model-based versus model-agnostic techniques. The main working objects are feature attribution methods — SHAP values from cooperative game theory, LIME's local surrogate models, Integrated Gradients' path integrals, Grad-CAM's gradient-weighted activation maps, and Layer-wise Relevance Propagation's backward relevance decomposition — together with attention-weight visualization and embedding/probing analysis for transformers and LLMs. Code examples are the load-bearing mechanism for transferring the techniques to readers; each chapter pairs a method with a minimal Python implementation, and the book's promise of practical mastery rests on those implementations being correct.","core_discovery":"On the paper's own terms, the central discovery is organizational: the many threads of XAI — intrinsic model interpretability, feature attribution, counterfactual and causal explanation, attention and embedding analysis, evaluation metrics, and tooling — belong in a single narrative that runs from classical models to LLMs. The guide asserts that this narrative can be made hands-on, with Python code for each technique, and that the resulting competence lets practitioners debug models, satisfy regulatory demands for explanation, and audit fairness. No new algorithm or empirical result is claimed; the contribution is the synthesis and its pedagogical packaging.","pith_inferences":["The guide's own Section 4.3.1 example plots the model's prediction output as if it were hidden-state activations; if similar mismatches occur elsewhere, readers following the code would internalize a misleading picture of what the code computes.","A natural extension the author leaves implicit is a curated, testable companion suite that verifies each code example against the method it claims to illustrate, especially the RNN and attention examples.","The synthesis suggests a research hypothesis: a practitioner trained on classical post-hoc methods can transfer that training to LLM interpretability faster than one starting directly with LLM-specific tools; this could be tested with a controlled learning study."],"forward_implications":["A reader who works through the guide is expected to be able to apply SHAP, LIME, and related methods to tabular, image, and text models, and to distinguish local from global explanations.","The book implies that interpretability is not one property: transparency, interpretability, explainability, and fairness are related but distinct, so evaluation should use multiple metrics.","LLM interpretability is treated as an extension of existing machinery (gradient attribution, probing classifiers, attention analysis) rather than a separate field, so skills transfer from classical models to BERT, GPT, and T5.","Case studies in healthcare, finance, and policy are presented as evidence that XAI methods have practical decision-support value beyond model debugging."],"supporting_citations":[{"why":"Supplies SHAP, the central game-theoretic feature attribution method used throughout the guide.","marker":"[22]"},{"why":"Supplies LIME, the local surrogate model that anchors the post-hoc interpretability chapters.","marker":"[23]"},{"why":"Supplies Grad-CAM, the gradient-weighted activation mapping technique for CNN explanations.","marker":"[24]"},{"why":"Supplies the Transformer and self-attention architecture that underlies the LLM interpretability chapters.","marker":"[36]"},{"why":"Supplies Integrated Gradients, the path-integral attribution method used for deep networks and LLMs.","marker":"[89]"},{"why":"Identifies BERT as a canonical bidirectional LLM whose embeddings and layers are probed in examples.","marker":"[92]"},{"why":"Identifies T5 as the unified text-to-text LLM used to illustrate task-agnostic architecture.","marker":"[93]"},{"why":"Identifies GPT-3 and the few-shot learning paradigm that motivates LLM interpretability challenges.","marker":"[95]"},{"why":"Identifies LLaMA as an open-weight model enabling transparency and research access.","marker":"[103]"},{"why":"Supplies Layer-wise Relevance Propagation, the backward relevance decomposition technique for deep models.","marker":"[166]"}],"fun_headline_variants":["XAI unboxed: classical models to LLMs","From SHAP to GPT: full XAI guide","XAI tour: decision trees to LLM probing","One guide to explainable AI, all models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guide's value as a practical resource depends on its code examples actually implementing the techniques they describe, and the RNN hidden-state example in Section 4.3.1 already plots the model's output as though it were hidden states, so this premise is not fully met.","fun_headline_variants_meta":{"raw":{"variants":["XAI unboxed: classical models to LLMs","From SHAP to GPT: full XAI guide","XAI tour: decision trees to LLM probing","One guide to explainable AI, all models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000146,"raw_usage":{"total_tokens":1157,"prompt_tokens":892,"completion_tokens":265,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":202}},"tokens_in":508,"tokens_out":265,"duration_ms":3537,"temperature":1.0,"reasoning_tokens":202,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:58:37.858721+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the Section 4.3.1 RNN code and inspect what `model.predict(X)` returns: if the plotted curves are the network's final predictions rather than the 10 hidden-unit activations the caption claims, the example does not demonstrate hidden-state interpretation. More generally, executing each chapter's code and checking that the printed outputs and figures match the described behavior would settle whether the practical promise holds.","supporting_citations":[],"review_version":1}