{"id":"117bb699-6779-453c-9e8e-8831106a76dd","arxiv_id":"2504.16913","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Adding chain-of-thought explanations to a fine-tuned BERT improved human-vs-AI text detection from 0.742 to 0.898 F1 in a shared task, but the explanations were generated with the true labels.","lead":"The paper describes COT_Finetuned, a system that combines chain-of-thought explanations with fine-tuned BERT to classify text as human or AI-written and to identify the generating LLM. It reports 0.898 F1 for human/AI classification in the DeFactify 4.0 shared task, though the explanation generation step may leak the true labels.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported CoT gain is not evaluable because reasoning is generated using the true label at training time and no label-free test-time procedure is defined; the method leaks the answer.","rationale":"The reader's weakest assumption identifies exactly the load-bearing flaw: the CoT reasoning used as input is generated from the true label, so the central BERT+COT comparison cannot be a clean test of whether reasoning helps. I agree with that assessment. Additional corroborating weaknesses include the loss function in Eq. (3), which omits the reasoning loss promised in Section 3.2.3, so there is no evidence the model is trained to produce reasoning at all; and the lack of released code or data, which prevents independent verification of the leaderboard scores. The Task B result is also modest and far below the top leaderboard entries. None of this requires questioning author intent; the paper is simply under-specified at the point where its central claim depends on a test-time procedure that is never defined. The concrete test above would settle the issue directly.","tokens_in":6756,"tokens_out":3719,"duration_ms":35677,"concrete_test":"Ask the authors to release the exact test-time inference script and run BERT+COT on the official test set with reasoning generated by a label-free prompt (e.g., 'Why might this text be AI-generated or human-written?') instead of p(d_i, y_i). If the Task A F1 drops materially from 0.898 toward the plain BERT baseline 0.742, the reported improvement is an artifact of label leakage. Alternatively, if the released code reveals that y_i is used during evaluation, the result is invalid by construction.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim—that adding CoT reasoning to BERT raises Task A F1 from 0.742 to 0.898—rests on the training pipeline in Section 3.2.4. Equations (4)-(5) define the reasoning r_i as LLaMA's output to the prompt p(d_i, y_i) = 'Why is this particular d_i generated by y_i?', where y_i is the true Task A label. The same r_i is then fed to BERT as an input feature. The paper says only that after training, 'pass the test dataset to the model, which will return Label_A and Label_B' (Section 3.2.4). It never specifies how reasoning is obtained for test documents, whose labels are withheld (Section 4.1). There are only two possibilities, and both break the claim. If y_i is used to generate test-time reasoning, the model is effectively given the answer it is supposed to predict, so the 0.898 score is not an honest detection result. If some unstated label-free reasoning generator is used instead, then the evaluated system differs from the method described and the ablation comparing BERT vs BERT+COT is not reproducible from the paper. Either way, the causal statement 'COT reasoning improves performance' is unsupported, and the Task B gain (0.249 to 0.307) is even weaker and subject to the same leakage.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes COT_Finetuned, a two-task framework for the DEFACTIFY 4.0 CT2 shared task: Task A classifies text as human-written or AI-generated, and Task B identifies which LLM generated an AI text. The method first uses LLaMA to generate chain-of-thought reasoning for each document, then fine-tunes BERT or RoBERTa on the document, its true label, and the reasoning. The authors report that adding CoT improves Task A F1 from 0.742 to 0.898 for BERT and Task B F1 from 0.249 to 0.307, and they conclude that CoT reasoning improves performance and interpretability.","tokens_in":7169,"tokens_out":3818,"duration_ms":35491,"significance":"The paper addresses a timely problem—detecting AI-generated text and attributing it to a specific model—and the idea of using free-form reasoning as an auxiliary feature is interesting. If the claims were valid, the approach would provide a transparent detection method. However, the central experimental result is undermined by a probable label leak in the reasoning-generation step, and the paper does not specify a reproducible test-time procedure. The reported F1 gains therefore cannot be attributed to CoT reasoning, which is the paper's core contribution. On the positive side, the authors participate in a shared task and report official leaderboard scores, but that does not compensate for the methodological flaws.","major_comments":[{"comment":"The prompt p(d_i, y_i) used to generate reasoning r_i contains the true Task A label y_i (\"Why is this particular d_i generated by y_i?\"). For Task B, the same construction is implied using y^B_i. The paper never specifies how reasoning is produced for unlabeled test documents; the only test-time instruction is to \"pass the test dataset to the model.\" If the true label is used at test time, the classifier is effectively given the answer it is supposed to predict, so the reported F1 of 0.898 for Task A is not an honest detection score. If some unstated label-free reasoning generator is used instead, then the evaluated system is not the one described, and the comparison between BERT and BERT+COT is not reproducible. Either way, the claim that CoT improves performance is unsupported. Table 7 further confirms the leak: the reasoning outputs in the examples explicitly contain the classification labels ('Human' and 'LLM').","section":"3.2.4, Eqs. (4)-(5)"},{"comment":"The text states that the combined loss includes a \"reasoning loss component,\" but Eq. (3) sums only the two classification losses for Task A and Task B. No reasoning loss is defined or optimized anywhere else. Consequently, the claim that the model \"learns to generate high-quality reasoning\" is not backed by any training signal, and the reasoning strings function only as input features rather than as an output the model is trained to produce. This is a load-bearing discrepancy between the method description and the actual loss function.","section":"3.2.3, Eq. (3)"},{"comment":"The experimental setup is internally inconsistent: the main text reports a batch size of 32 and epochs ranging from 50 to 250, while Appendix A.2 reports a batch size of 5 and epochs from 10 to 250. All results in Table 3 are single-run F1 scores with no variance, confidence intervals, or significance tests. Given the modest Task B improvement (0.249 to 0.307), the possibility that the observed gains are due to run-to-run variability cannot be ruled out. These issues make the central result difficult to evaluate or reproduce.","section":"4.2 and A.2"}],"minor_comments":[{"comment":"There is a typo in the final sentence: \"Labe_A\" should be \"Label_A\".","section":"3.2.4"},{"comment":"Equation (1) defines Label_B via a likelihood maximization over models, but this likelihood is never computed or used anywhere else in the paper; the actual Task B labels are simply taken from the dataset.","section":"3.2.2, Eq. (1)"},{"comment":"The phrase \"the future text d_i\" appears to be a typo; it should read \"the text d_i\".","section":"3.2.4, step 5"},{"comment":"The hyperparameter row \"k 6 layer\" is unclear and should be removed or explained, and the learning rate format is inconsistent with the rest of the table.","section":"Appendix A.2, Table 6"},{"comment":"The column header says \"Score\" but the values are F1 scores; this should be stated explicitly in the caption.","section":"Table 2"},{"comment":"Reference [13] has \"????\" in place of a year, and the email address in the footnote is corrupted; these should be corrected.","section":"References"}],"recommendation":"reject","confidential_remarks":"This is a short shared-task system description whose central experimental comparison is invalid as described. The reasoning generation step leaks the true label, and no label-free test-time procedure is specified, so the claimed CoT improvement cannot be taken at face value. I do not see a path to acceptance without re-running the experiments with a properly specified protocol, which goes beyond a typical revision. Hence my recommendation is reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the large F1 gain the paper attributes to chain-of-thought reasoning is not credible, because the reasoning is generated from the true label. The stress-test note is right. Section 3.2.4 builds prompt p(d_i, y_i) = \"Why is this particular d_i generated by y_i?\" using the true label, feeds LLaMA's output to BERT, and never defines a label-free procedure for test documents. So the method either leaks the answer or uses an unstated, different generator. The central claim is unsupported.\n\nWhat is new: applying CoT rationales as input features for AI-text detection is a reasonable idea, and I have not seen it in the cited literature. The dual-task framing and leaderboard context are useful for shared-task participants.\n\nSoft spots, in order: (1) label leakage in the training pipeline; (2) missing test-time reasoning specification; (3) Eq. (3) omits the reasoning loss promised in the text; (4) single-run results with no variance and no code/data. Task B at 0.307 F1 is weak even if the method worked. The paper's claim that CoT improves both tasks is overstated given the small Task B difference.\n\nThis is not a paper to build on as written. With a corrected setup—label-free reasoning generation and a proper loss—the idea might be worth a follow-up, but the current version does not deserve publication. I would not cite it, and I would not send it to a serious referee without a major rewrite.\n\nFor a reading group, it is a good example of accidental label leakage in shared-task systems.","headline":"The reported CoT gain is an artifact of label leakage; the underlying idea is plausible but the evaluation as presented is broken.","tokens_in":7551,"tokens_out":3719,"would_cite":false,"duration_ms":32839,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Appending chain-of-thought explanations to BERT improves detection of AI-generated text and identifies the generating LLM.","keywords":["chain-of-thought reasoning","AI-generated text detection","LLM attribution","BERT fine-tuning","text classification","interpretability","LLaMA"],"falsifier":"Run the same COT_Finetuned pipeline on unlabeled test documents where the LLaMA reasoning is produced by a prompt that omits the true label, then compare the F1 scores; if the score falls back to the plain-BERT level, the reported CoT gain is an artefact of label leakage.","tokens_in":6572,"feed_emoji":"🧠","tokens_out":3527,"duration_ms":28132,"temperature":0.7,"pith_summary":"The paper proposes COT_Finetuned, a dual-task framework that appends chain-of-thought reasoning to a fine-tuned BERT to classify text as AI-generated or human-written (Task A) and to identify which LLM produced an AI-generated document (Task B). The central claim is that the CoT reasoning, generated by LLaMA from a prompt that asks why the document was produced by its true label, substantially improves performance: BERT + CoT reaches 0.898 F1 on Task A and 0.307 on Task B, versus 0.742 and 0.249 for plain BERT. The authors argue the reasoning process adds interpretability as well as accuracy, and that CoT helps the model pick up stylistic patterns distinctive of different LLMs.","feed_headline":"CoT reasoning lifts AI-text detection F1 to 0.898","feed_subtitle":"Adding LLaMA-generated explanations to BERT beats plain BERT and RoBERTa on both detection and attribution.","key_machinery":"The core mechanism is the chain-of-thought (CoT) explanation: for each document $d_i$, the prompt $p(d_i, y_i) = \\text{'Why is this particular } d_i \\text{ generated by } y_i\\text{?'}$ is passed to LLaMA to produce reasoning $r_i$, and the document, its label, and $r_i$ are passed to a fine-tuned BERT whose output layer applies sigmoid for Task A and softmax for Task B. This CoT-augmented input is what carries the claimed improvement, together with a total loss that sums binary cross-entropy for Task A and cross-entropy for Task B.","core_discovery":"The discovery is that a classifier can be made more accurate and more transparent by conditioning it on generated reasoning: for each training document, the method prompts LLaMA with 'Why is this particular document generated by this label?' and uses the resulting explanation as an additional input to a fine-tuned BERT. Trained with a joint loss over the binary human/AI decision and the multi-class LLM attribution, COT_Finetuned outperforms both RoBERTa and BERT alone. On the shared task's test set, BERT + CoT achieves an F1 of 0.898 for Task A and 0.307 for Task B, and the paper concludes that chain-of-thought reasoning improves both tasks.","pith_inferences":["The prompt in Eq. (4) includes the true label $y_i$, so at test time the method would need an unstated label-free way to generate reasoning; if instead the label leaks through the reasoning, the reported improvement may be inflated.","A testable extension: generate reasoning from a label-free prompt (e.g., 'Why might this text be AI-generated?') and compare F1; a large drop would confirm label leakage.","CoT reasoning may act as a stylometric feature extractor: the explanations could encode the very stylistic cues that distinguish GPT-4.0 from DeBERTa or FalconMamba, so the method could be combined with contrastive learning to push Task B beyond 0.307."],"forward_implications":["If CoT-augmented classification works as claimed, adding explainable reasoning inputs is a cheap way to boost detector accuracy without changing the classifier architecture.","The dual-task framing means a single fine-tuned model can serve both human/AI filtering and forensic attribution of the generating model, which would matter for academic-integrity tools.","Since Task B scores remain low (0.307), the paper implies that identifying the exact LLM is inherently harder and needs more than current CoT inputs to reach usable accuracy.","The method's reliance on LLaMA-generated explanations implies detector quality depends on the explanatory capabilities of the reasoning generator, not just on the text classifier."],"supporting_citations":[{"why":"Task overview and leaderboard that define the two tasks and provide the official F1 scores used for comparison.","marker":"[1]"},{"why":"Supplies the DEFACTIFY-text dataset with human and AI-generated documents and labels for both tasks.","marker":"[21]"},{"why":"Provides the BERT and fine-tuned RoBERTa baselines that COT_Finetuned extends and compares against.","marker":"[5]"}],"fun_headline_variants":["CoT reasoning lifts AI-text detection F1 to 0.898","Reasoning boosts BERT for AI-text detection and attribution","Chain-of-thought explanations make AI-text ID transparent","Why-questions help BERT identify AI text better"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reasoning fed into the classifier is generated from a prompt that already contains the true answer (the label), and the paper does not say how test documents get reasoning without leaking that label.","fun_headline_variants_meta":{"raw":{"variants":["CoT reasoning lifts AI-text detection F1 to 0.898","Reasoning boosts BERT for AI-text detection and attribution","Chain-of-thought explanations make AI-text ID transparent","Why-questions help BERT identify AI text better"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000228,"raw_usage":{"total_tokens":1425,"prompt_tokens":843,"completion_tokens":582,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":514}},"tokens_in":459,"tokens_out":582,"duration_ms":5461,"temperature":1.0,"reasoning_tokens":514,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:51:32.724905+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same COT_Finetuned pipeline on unlabeled test documents where the LLaMA reasoning is produced by a prompt that omits the true label, then compare the F1 scores; if the score falls back to the plain-BERT level, the reported CoT gain is an artefact of label leakage.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DEFACTIFY-text dataset with human and AI-generated documents and labels for both tasks."}],"review_version":1}