{"id":"64ec23fc-4177-4901-9c42-648dab22f5a3","arxiv_id":"2507.16840","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"CASPER claims a triplet-view contrastive learning method with an equal-angle similarity vector improves smart Ponzi scheme detection over SourceP, especially with only 25% labels.","lead":"A new AI system called CASPER uses contrastive learning on unlabeled smart contract code to detect Ponzi schemes with less labeled data. On the XBlock benchmark it reports higher F1 scores than the SourceP baseline at every label fraction, but the evaluation has several biases.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The medium augmentation in §III-B replaces function bodies with return statements, so the contrastive loss trains the encoder to be invariant to the exact Ponzi-defining logic; the reported 92.4 F1 is difficult to reconcile with the described method unless an unstated factor compensates.","rationale":"The reader's weakest_assumption is exactly that the three augmentations preserve Ponzi-relevant semantics, and the medium augmentation is the clearest violation: replacing function bodies with `return value;` deletes the logic that defines a Ponzi scheme. The contrastive objective in Eq. 12 then actively trains the encoder to be invariant to that deletion, and the DFG of the stripped view is empty, so data-flow cues are also suppressed. This is not merely a missing analysis; it is an internal design inconsistency with the claimed results. The data-overlap concern raised by the reader is important, but it is secondary: even without overlap, the described objective should hurt rather than help, so the reported 92.4 F1 cannot be attributed to the method as written. The other flagged issues (headline number mismatch, invalid proof in Section IV, missing code) are serious but do not by themselves explain the mechanism; the augmentation flaw does. A controlled ablation of the view set would settle whether the described method can deliver the claimed gain. Since this reinforces the reader's rejection rather than changing it, the verdict remains UNCHANGED.","tokens_in":21792,"tokens_out":8360,"duration_ms":107104,"concrete_test":"Run the XBlock experiment exactly as described in §V-A but vary only the self-supervised view set: (a) all three views as in §III-B; (b) drop the function-body-replacement view and use only the two semantics-preserving views; (c) replace the medium view with a semantics-preserving augmentation (e.g., comment insertion) that keeps function bodies intact. Keep the 25% labeled subset, the self-training loop, and the temporal split fixed, and report F1 over at least five seeds. If (b) or (c) matches or exceeds the 92.4 F1 of (a), the body-replacement augmentation is not the source of the reported gain and the method as written is not validated; if (a) remains best with low variance, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B defines three views: variable splitting (Eq. 4), function-body replacement (Eq. 6), and variable renaming (Eq. 8). The medium view takes every function body and replaces it with `return value;`, and Section III-B then computes a fresh AST/DFG for each augmented view. The contrastive loss in Eq. 12 maximizes similarity between the original and this body-stripped view, so the InfoNCE-style objective will drive the encoder to discard any feature that is not shared with a contract whose entire payout, withdrawal, and transfer logic has been deleted. The DFG of the stripped view is essentially edge-free, so data-flow information is likewise aligned against an empty graph. Thus the learned representation is blind to the very cues that separate Ponzi from non-Ponzi contracts. If the method truly reached F1 92.4 with only 25% labels, that success cannot come from the described contrastive objective; it would have to come from an unstated source, such as overlap between the 10,051-contract pre-training corpus and XBlock, or an implementation that does not match the text. The paper's own ablation in Figure 5 says DFG is the more important input, which makes the empty-DFG alignment even more damaging. This makes the central label-efficiency claim unsupported as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CASPER, a semi-supervised framework for detecting smart Ponzi schemes in Ethereum contracts. The method first pre-trains a GraphCodeBERT encoder with a contrastive objective on unlabeled contract source code, using three code augmentations (variable splitting, function-body replacement, and variable renaming) to generate views, and a three-view similarity measure based on an 'intermediate vector' with equal angles to the three view representations. A classifier is then trained with labeled plus pseudo-labeled data via confidence-thresholded self-training. The authors report that on the XBlock dataset, CASPER reaches F1 92.4 with only 25% labels and F1 95.2 with 100% labels, outperforming SourceP and several other baselines, and they further report generalization and transfer results on EPSD, EBD, honeypot, and phishing datasets. The paper also contains a proof and derivation for the multi-vector cosine similarity method.","tokens_in":22121,"tokens_out":7850,"duration_ms":65285,"significance":"If the reported results are correct, CASPER would be a practically valuable contribution: it would show that self-supervised pre-training on unlabeled source code plus pseudo-label self-training can substantially reduce the label cost of smart Ponzi detection while improving accuracy over fully supervised source-code baselines. The proposed three-view contrastive objective with an O(N^3) negative sample pool is conceptually interesting, and the authors provide a new similarity formulation. However, the significance is heavily conditional: the central experimental claims are undermined by internal inconsistencies (abstract vs. table, text vs. table), a lack of error bars, and, most importantly, the described medium augmentation destroys the function-body semantics that the paper's own ablation suggests are crucial. These issues cast doubt on whether the reported performance can be attributed to the described method.","major_comments":[{"comment":"The medium augmentation replaces every function body with 'return value;' and the contrastive loss (Eq. 12) trains the encoder to be invariant to this replacement. This removes exactly the payout, withdrawal, and transfer logic that distinguishes a Ponzi contract from a legitimate one, and the DFG of the stripped view is essentially edge-free. The paper's own ablation in Figure 5 shows that DFG (i.e., data-flow context between functions) is the more important input. As described, the method should therefore discard the most discriminative features; the reported F1 of 92.4 at 25% labels is not credible unless an unstated compensating factor exists. Please run a control experiment with a semantics-preserving medium augmentation (e.g., dead-code insertion or statement reordering) and report whether all three views remain necessary; also probe the learned encoder with a linear classifier on function-body-level semantics to verify that the augmented views retain Ponzi-relevant information.","section":"Section III-B (Eq. 5–6) and Section V-B (Fig. 5)"},{"comment":"The headline numbers in the abstract do not match the results in Table III. The abstract states that with 100% labels CASPER outperforms the baseline by 2.3% in F1 score, but Table III shows a difference of 4.9 percentage points (95.2 vs. 90.3); with 25% labels the abstract claims 'nearly 20% higher,' but Table III shows a difference of 13.1 points (92.4 vs. 79.3). The text in Section V-B.1 also says SourceP's F1 is higher at 25% than at 50%, yet Table III lists 79.3 at 25% and 84.2 at 50%. These contradictions must be resolved before the reported performance can be assessed.","section":"Section V-B.1, Table III vs. abstract"},{"comment":"The self-supervised pre-training corpus (10,051 contracts crawled from Etherscan) and the XBlock evaluation set (6,498 contracts from Etherscan) are both built from Etherscan, but no overlap check is reported. If any evaluation contract appears in the pre-training corpus, the reported label-efficiency gains could be partly due to test-time memorization rather than representation quality. The same concern applies to the EPSD, EBD, honeypot, and phishing evaluation sets. Please report a hash-based or address-based overlap analysis between the pre-training corpus and each evaluation set, and re-run the experiments after removing any overlapping contracts.","section":"Section V-A.1 and Section V-B.1"},{"comment":"The paper explicitly states that the block-height-based split 'can provide better model performance' than a random division. Choosing the split strategy that maximizes performance for the main comparison, without reporting the random-split result or any variance across seeds, makes the headline F1 values non-representative. Report mean and standard deviation over at least five random seeds for both the temporal split and a random split, for both CASPER and SourceP, and state which split was used for each table.","section":"Section V-B.1"},{"comment":"The proof of the existence of an equal-angle intermediate vector is invalid. In Section IV-A, the plane H perpendicular to an arbitrary vector v need not intersect all three given vectors, and the circumcenter argument equates distances from a line to points A, B, C, not the angles between the direction vector v and the original vectors. The subsequent derivation is restricted to R^3 and contains typographical errors (Eq. 21 repeats 'αxa' in all three terms; Eq. 19's third denominator appears to use z_b instead of z_c). Since the actual feature vectors produced by GraphCodeBERT are 768-dimensional, the paper must provide a correct existence and construction argument for R^d, or state explicitly which approximation is used in higher dimensions.","section":"Section IV-A and IV-B"}],"minor_comments":[{"comment":"'Detecter' is a typo; also the abstract says 'smart Ponzi scheme detectER' while the title uses the same misspelling.","section":"Title/Abstract"},{"comment":"The naming of augmentations is inconsistent: the text calls the function-body-replacement augmentation 'medium,' the renaming augmentation 'weak,' but then uses subscripts w (weak) for the function-body view and m (medium) for the renaming view. Align the names and notations.","section":"Section III-B"},{"comment":"The bar chart text is not legible in the PDF and the numbers are not referenced; also the claim about SourceP instability is contradicted by Table III, as noted in the major comments.","section":"Section V-B.1, Figure 3"},{"comment":"The pseudo-label loss uses the predicted probability yhat_unlabeled,j as both the target and the log input; this is unusual. Please clarify whether hard pseudo-labels are used, and if so, define the loss with the hard label.","section":"Equation (15)"},{"comment":"Reference [13] is a duplicate of [12]; reference [14] is listed as an ACM URL but is not properly formatted; several in-text citations appear as '[?]' (e.g., in Section I and Section VI). The reference list also contains unrelated entries (e.g., flood susceptibility modeling).","section":"References"},{"comment":"The ablation results are only in a radar figure with garbled text; the exact F1/Precision/Recall values are not given. Report the numeric table for all ablation combinations so the claims can be verified.","section":"Section V-B.4, Figure 5"}],"recommendation":"reject","confidential_remarks":"As the reader's report notes, the central problem is not a formal circularity but an internal inconsistency: the described contrastive augmentation should erase the features that the model's own ablation shows are essential. Unless the authors can show that a semantics-preserving version of the medium augmentation yields the same performance, the results as written cannot be trusted. The abstract/table mismatches and the lack of overlap checks further reinforce the rejection. I see no path within a revision that would not require redesigning the method and rerunning all experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the label-efficiency result is almost certainly an artifact of a bad augmentation choice, not a property of your method. The medium augmentation in Section III-B replaces every function body with a return statement, then computes a fresh DFG; the stripped DFG is essentially edge-free. The contrastive loss in Eq. 12 aligns all three views to an intermediate vector, so the encoder is trained to be invariant to the very function bodies where payout, withdrawal, and transfer logic lives. Your own ablation in Figure 5 says DFG is the most important input, and the DFG of the stripped view is empty. Those two statements cannot both be true. Unless there is unstated leakage between the 10,051-contract pre-training corpus and XBlock, an F1 of 92.4 with 25% labels is not credible from the described pipeline.\n\nWhat the paper gets right: the three-view contrastive loss with N^3-N negatives is a genuine, if small, extension of InfoNCE. The experiments cover multiple datasets (EPSD, EBD, honeypot, phishing) and the low-label setting is a real operational problem for blockchain forensics. The writing is candid about the block-height split, though the candor makes the flaw worse.\n\nThe evaluation has more problems: the abstract says '2.3%' and 'nearly 20%' but Table III shows 4.9-point and 13.1-point gains. No error bars. No overlap check between the pre-training corpus and the evaluation sets. The 'proof' of the equal-angle vector in Section IV is wrong (rank 2 leaves a 1-dimensional nullspace, not only the trivial solution) and the equations have typos. No code or data is provided.\n\nBottom line: the idea is worth testing, but this version does not test it. The central claim is undermined by the augmentation design. Reject in current form; if the authors resubmit, they need to show the augmented views preserve Ponzi-relevant semantics, release code and data, fix the numbers, and use a defensible split.\n\nI would still give this a serious referee rather than a desk reject, because the failure mode is instructive and a good referee could force the authors to confront the augmentation problem. For yourself, read Section III-B and Figure 5 together; the tension jumps out.","headline":"The label-efficiency gains are almost certainly an artifact of an augmentation that deletes the Ponzi-defining logic; the paper has a plausible idea but the central empirical claim is not supported by the described pipeline.","tokens_in":22631,"tokens_out":5335,"would_cite":false,"duration_ms":58937,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CASPER shows that contrastive pre-training on unlabeled smart-contract source code detects Ponzi schemes at F1 92.4 with only 25% of the labels, beating a fully supervised baseline.","keywords":["smart Ponzi scheme detection","contrastive learning","semi-supervised classification","smart contract security","Ethereum","pseudo-label self-training","multi-vector cosine similarity","GraphCodeBERT"],"falsifier":"A direct check: train the encoder on the 10,051 unlabeled contracts with only the medium augmentation (function bodies replaced by a return statement), then see whether a linear probe on the resulting representations can still separate the 318 known Ponzi contracts from the non-Ponzi contracts in XBlock; if it cannot, the contrastive objective has learned away the body-level payout logic that distinguishes Ponzi contracts. A second, cheaper check: search for address or source-code hash overlaps between the 10,051 pre-training contracts and the XBlock, EPSD, EBD, honeypot, and phishing sets—if overlap exists, the label-efficiency advantage may be partially inherited rather than learned.","tokens_in":1783,"feed_emoji":"🕵️","tokens_out":1668,"duration_ms":90578,"temperature":0.7,"pith_summary":"The paper is trying to show that smart Ponzi scheme detection does not have to be a fully supervised problem. It proposes CASPER, which first learns representations of smart-contract source code from 10,051 unlabeled contracts using a three-view contrastive objective, then trains a semi-supervised classifier that expands its own training set with confident pseudo-labels. On the XBlock dataset, CASPER reaches F1 92.4 with only 25% of the labels, against 79.3 for the SourceP baseline under identical conditions, and 95.2 with full labels. The same pre-trained encoder transfers to other Ponzi datasets and to honeypot and phishing contracts. If these numbers hold, the main bottleneck of scarce labels can be pushed off the critical path.","feed_headline":"Ponzi detection hits F1 92.4 with a quarter of the labels","feed_subtitle":"CASPER needs only 25% of the labels to beat the fully labeled baseline by 2.1 F1 points.","key_machinery":"The central object is a three-view contrastive objective over augmented smart-contract source codes. Each contract is transformed by three augmentations: splitting variables into two to five sub-variables, replacing function bodies with a simple return statement, and renaming variables. Each view's source code is parsed into an abstract syntax tree and a data flow graph, and GraphCodeBERT takes the source code plus the data flow graph as input to produce a feature vector. The contrastive loss maximizes a multi-vector cosine similarity among the three views of the same contract while minimizing similarity against the $N^3 - N$ negative pairs formed by views of different contracts in the batch, with the equiangular intermediate vector representing the similarity of each group. A self-training classifier then assigns pseudo-labels to confident unlabeled predictions and retrains iteratively.","core_discovery":"The central claim is that a contrastive representation-learning stage over unlabeled code, followed by self-trained semi-supervised classification, yields smart Ponzi detection that is both more accurate and far less label-hungry than fully supervised source-code analysis. The authors compare CASPER against SourceP with the same label settings: with 100% labels CASPER's F1 is 95.2 versus 90.3, and with 25% labels it is 92.4 versus 79.3. They further report that the learned encoder generalizes to the EPSD and EBD Ponzi datasets, to unseen Ponzi scheme types, and to unrelated fraud types such as honeypot and phishing contracts, which they read as evidence that the representation captures general contract semantics rather than a memorized dataset. The paper also contributes a multi-vector cosine similarity measure, built on an intermediate vector that makes equal angles with all input vectors, and reports that it converges faster than centroid or weighted-average similarity.","pith_inferences":["A testable consequence the paper does not report: if the 10,051-contract pre-training corpus has no overlap with the evaluation sets, then the same three-view contrastive recipe could be dropped onto other blockchain fraud labels without architectural changes.","The augmentations' semantics deserve scrutiny: replacing function bodies with a return statement could teach the encoder to ignore the payout logic that distinguishes Ponzi contracts, so a synthetic check of whether augmented views retain Ponzi-relevant cues would settle whether the label-efficiency gain is robust to augmentation choice.","The paper attributes part of its gain to the multi-vector cosine similarity, but it does not isolate the equiangular intermediate vector from the larger negative pool; ablating the intermediate vector while keeping the $N^3-N$ negative pairs would clarify which component matters.","Because CASPER works from static source code and data flow graphs, it can in principle screen contracts before deployment, unlike transaction-based detectors that only react after funds have moved."],"forward_implications":["Smart Ponzi detection can be built with far fewer hand-labeled contracts: with 25% labels CASPER already exceeds the F1 of SourceP trained on 100% labels.","Unlabeled source code from public blockchain explorers can carry the representation-learning burden, reducing annotation cost and making large-scale screening more practical.","The pre-trained encoder transfers to other Ponzi datasets and to other fraud types, so one contrastive pre-training run may serve multiple detection tasks.","The learned representation is classifier-agnostic: SVM, XGBoost, and a multilayer perceptron all reach F1 scores above 93, indicating that the representation, not the final classifier, drives the gain.","Expanding the negative-sample pool from $2N-1$ (SimCLR) and $N^2-N$ (CLIP) to $N^3-N$ improves downstream classification in the reported comparison."],"supporting_citations":[{"why":"Supplies GraphCodeBERT, the pre-trained code representation backbone that turns source code and data flow graphs into feature vectors.","marker":"[57]"},{"why":"Provides the SourceP baseline and the source-code-plus-data-flow comparison methodology used throughout the experiments.","marker":"[20]"},{"why":"Supplies the XBlock dataset and the classical baselines Ridge-NC, SVM-NC, XGBoost-TF-IDF, and MulCas.","marker":"[39]"},{"why":"Provides SadPonzi, the bytecode-based semantic-analysis baseline used in the main and representation-learning comparisons.","marker":"[7]"},{"why":"Provides SimCLR, the contrastive baseline with $2N-1$ negative samples against which the larger negative pool is tested.","marker":"[51]"},{"why":"Provides CLIP, the contrastive baseline with $N^2-N$ negative samples used in the same comparison.","marker":"[86]"},{"why":"Supplies the EPSD Ponzi dataset used to test generalization to another labeled Ethereum Ponzi collection.","marker":"[61]"},{"why":"Supplies the EBD Ponzi dataset used as a second generalization target.","marker":"[16]"}],"fun_headline_variants":["CASPER: contrastive learning spots smart Ponzi schemes with 25% of labels","Smart Ponzi detection with a quarter of labels outperforms fully supervised","Contrastive approach cuts label needs 75% while boosting Ponzi detection F1","CASPER beats full-supervision Ponzi detection using just 25% of the labels","Ponzi scheme detector CASPER: 25% labels, 2.1 F1 gain over supervised"],"cache_read_input_tokens":24704,"weakest_assumption_plain":"The load-bearing premise is that the three code augmentations—splitting variables, replacing function bodies with a return statement, and renaming variables—preserve the semantic signals that mark a contract as a Ponzi scheme; if an augmentation discards payout or withdrawal logic, the contrastive encoder will learn representations blind to the very cues the classifier needs, and the 10,051-contract pre-training corpus must also be assumed not to overlap the evaluation sets.","fun_headline_variants_meta":{"raw":{"variants":["CASPER: contrastive learning spots smart Ponzi schemes with 25% of labels","Smart Ponzi detection with a quarter of labels outperforms fully supervised","Contrastive approach cuts label needs 75% while boosting Ponzi detection F1","CASPER beats full-supervision Ponzi detection using just 25% of the labels","Ponzi scheme detector CASPER: 25% labels, 2.1 F1 gain over supervised"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000374,"raw_usage":{"total_tokens":2021,"prompt_tokens":994,"completion_tokens":1027,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":914}},"tokens_in":610,"tokens_out":1027,"duration_ms":10944,"temperature":1.0,"reasoning_tokens":914,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:57:12.348317+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct check: train the encoder on the 10,051 unlabeled contracts with only the medium augmentation (function bodies replaced by a return statement), then see whether a linear probe on the resulting representations can still separate the 318 known Ponzi contracts from the non-Ponzi contracts in XBlock; if it cannot, the contrastive objective has learned away the body-level payout logic that distinguishes Ponzi contracts. A second, cheaper check: search for address or source-code hash overlaps between the 10,051 pre-training contracts and the XBlock, EPSD, EBD, honeypot, and phishing sets—if overlap exists, the label-efficiency advantage may be partially inherited rather than learned.","supporting_citations":[{"cited_title":"SourceP: Detecting Ponzi Schemes on Ethereum with Source Code","cited_arxiv_id":"2306.01665","evidence_quote":"Provides the SourceP baseline and the source-code-plus-data-flow comparison methodology used throughout the experiments."},{"cited_title":"Securing the ethereum from smart ponzi schemes: Identification using static features","cited_arxiv_id":null,"evidence_quote":"Supplies the XBlock dataset and the classical baselines Ridge-NC, SVM-NC, XGBoost-TF-IDF, and MulCas."},{"cited_title":"A simple framework for contrastive learning of visual rep- resentations","cited_arxiv_id":null,"evidence_quote":"Provides SimCLR, the contrastive baseline with $2N-1$ negative samples against which the larger negative pool is tested."},{"cited_title":"Exploiting Blockchain Data to Detect Smart Ponzi Schemes on Ethereum","cited_arxiv_id":null,"evidence_quote":"Supplies the EBD Ponzi dataset used as a second generalization target."}],"review_version":1}