{"id":"bd16b9d5-bf16-4611-be0d-a75c10f9e7aa","arxiv_id":"2509.21530","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Constraining LLM rewriting with a biomedical NER model improves medical entity preservation and reduces hallucinations in synthetic clinical notes, with modest downstream gains on MIMIC-III tasks.","lead":"This paper presents a method for generating synthetic clinical notes that combines a small medical-entity detector with a large language model rewriter. The detector tells the rewriter which medical terms to keep unchanged, which reduces hallucinated content and improves downstream prediction accuracy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PR/HR evidence is circular: the same NER that issues the preservation constraints also scores the outputs, so the headline safety advantage needs an independent entity-level check.","rationale":"The reader's CONDITIONAL verdict is right in direction, but the most load-bearing flaw is sharper than 'NER may miss tokens': the evaluation metric is defined by the same model that produces the constraint set, which makes the headline PR/HR comparison partly tautological. Section A.3 explicitly states that E(·) uses the same tool used to create K_i, and the prompt in A.1 forbids altering those tokens; a method can score near-perfect PR without demonstrating that clinically relevant information survives in a form that preserves the label. Baselines are not given the same constraint, so their lower PR may reflect lexical change rather than semantic loss. This does not prove the method is useless — the downstream experiments are a genuine effort to show utility — but those results are presented without significance testing and are often within one standard error of each other. The concrete test above would settle whether the safety claim is substantive or an artifact. If the independent evaluation reproduces the advantage, the paper's central claim stands and the conditional concerns are minor; if it does not, the contribution reduces to a plausible but unvalidated heuristic. Therefore the appropriate verdict remains CONDITIONAL pending that check.","tokens_in":16728,"tokens_out":4859,"duration_ms":46236,"concrete_test":"Take 100–300 MIMIC-III notes. Independently annotate label-relevant facts (or use a different validated clinical NER plus clinician adjudication) and compute PR/HR against that ground truth, counting synonymous or paraphrased mentions as preserved, and run the exact same acceptance filter for Naive, CATO, and Ours. Also have a clinician (or an LLM judge with a rubric) decide for each generated note whether the downstream label would still be correct. If Ours's advantage over CATO disappears or label-relevant facts are lost despite verbatim token preservation, the central claim is an artifact of measuring with the constraint-generating model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest evidence for the safety claim is Table 1, but the PR/HR measurement in A.3 uses E(·) = entities extracted by 'the same tool used to create K_i' — the biomedical NER that also generates the constraints. The A.1 prompt then instructs the generalist to preserve exactly those strings verbatim. Consequently, Ours is scored on whether it keeps the exact tokens it was told to keep, while baselines that legitimately rephrase (e.g., 'hypertension' to 'high blood pressure') are marked as deletions. A.3 additionally describes an acceptance filter on PR/HR; if it is applied only to Ours, the Table 1 numbers are selected rather than representative. Token-level preservation also does not imply label preservation: the LLM can keep the entity string while changing its negation, scope, or temporality, so the claimed 'reduced hallucinations at token and concept levels' is not established by these metrics. Downstream gains (Table 2) are small relative to standard errors and lack significance tests, so the utility claim rests on the same circular safety evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a model-collaboration framework for clinical text augmentation: a lightweight clinical NER model (the weak expert) extracts safety-critical tokens, and a general-purpose LLM (the strong generalist) rewrites clinical notes while preserving those tokens verbatim. The authors evaluate the quality of the generated notes using preservation rate (PR) and hallucination rate (HR), and measure downstream utility by training Qwen-3 and BERT models on augmented MIMIC-III notes for readmission, mortality, and length-of-stay prediction, plus zero/few-shot phenotyping and ICD coding. They report that their method outperforms naive LLM paraphrasing and the CATO style-only baseline on both safety metrics and downstream tasks, and they additionally show that the guidance can be distilled into a single model via DPO. The core claim is that weak-expert constraints provide a practical safety layer for LLM-based augmentation in high-stakes domains.","tokens_in":16898,"tokens_out":2612,"duration_ms":23028,"significance":"If the safety and utility claims hold, the framework is attractive: it is lightweight, requires no additional training of the weak expert, and is compatible with small open-weight LLMs. The authors also state that prompts, acceptance decisions, and scripts will be released, which would aid reproducibility. The zero/few-shot phenotyping and ICD retrieval evaluations are a useful complement to supervised fine-tuning. However, the central safety evidence is weakened by a circular evaluation: PR and HR are measured with the same NER tool that produces the preservation constraints, and the acceptance filter is not described for baselines. Downstream gains are often within one standard error of the strongest baseline, and no significance tests are reported. These issues are fixable but currently leave the main claims insufficiently supported.","major_comments":[{"comment":"The headline safety results in Table 1 are partly by construction. Equation (3) defines PR and HR using E(·), which A.3 states is \"the same tool used to create K_i\"—i.e., the biomedical NER that generates the constraints. The prompt in A.1 then instructs the LLM to preserve exactly those strings verbatim. Thus, Ours is rewarded for keeping the exact tokens it was told to keep, while baselines that legitimately rephrase (e.g., \"hypertension\" to \"high blood pressure\") are penalized as deletions. Moreover, A.3 describes an acceptance filter with thresholds τ_PR and τ_HR, but the thresholds are never specified and it is not stated whether the same filter is applied to the Naive and CATO baselines. If the filter is applied only to Ours, Table 1 reports selected outputs, not representative ones. Please provide an independent entity-level evaluation (e.g., a different NER or manual review), report the thresholds, and apply identical filtering to all methods.","section":"§5.3, §A.3, Eq. (3)"},{"comment":"Token-level preservation does not establish label preservation. The framework assumes that K_i = W(x_i) is exactly the label-relevant variable set V, but no evidence is provided that the NER is complete or accurate with respect to the clinical labels. Section 6 itself concedes: \"When weak experts under-detect, medical facts change; when they over-detect, augmentation variation is limited.\" An LLM can keep an entity string while changing its negation, scope, or temporality (e.g., \"no fever\" to \"fever\"). Therefore the claim of \"reduced hallucinations at both the token and concept levels\" is not supported by PR/HR alone. Please add a direct check of label consistency on augmented notes, or an error analysis showing that semantic changes beyond token deletion are absent.","section":"§4.2, §6"},{"comment":"The downstream utility claim is weaker than stated. In Table 2, most differences between Ours and the best baseline are within one standard error: for BERT mortality, Ours is 0.929 ± 0.03 versus CATO 0.923 ± 0.003; for Qwen readmission, Ours is 0.599 ± 0.03 versus CATO 0.552 ± 0.04, which is less than two standard errors. No significance tests, confidence intervals, or paired comparisons are reported, so \"consistent performance gains\" is not statistically established. Additionally, hyperparameters in A.2 were selected by grid search on the readmission task, which may inflate that task's result; please clarify whether the same fixed protocol was used for all augmentation methods and report significance tests or effect sizes.","section":"Table 2, Figure 5, Table 3"}],"minor_comments":[{"comment":"There is a typo: \"G receives explicit guidance on domain-critical tokens, t avoids distorting\" should read \"it avoids distorting.\"","section":"§4.2"},{"comment":"The word \"rresults\" appears in the sentence \"medical specialization yields the best rresults\"; please correct it.","section":"§6"},{"comment":"The column header \"Pred.\" should be \"Prec.\" for precision, to match the text and standard notation.","section":"Table 3"},{"comment":"The model name is inconsistently written as \"LLama\" and \"Llama\"; please standardize to \"Llama\" or \"LLaMA\" throughout.","section":"§5.2, Table 1"},{"comment":"The hyperparameter analysis in Tables 6–8 reports accuracy/F1 for readmission only; if these hyperparameters were used for all downstream tasks, please state that explicitly, since the mortality and length-of-stay results may have different optimal settings.","section":"§A.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's reproducibility appendix promises release of scripts and prompts, but no link or repository is provided in the manuscript; the editor may wish to ask the authors to include the code/data availability statement. The main concern is the circular safety evaluation, which the authors can address with an independent entity-level check; the downstream claim would also benefit from significance testing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid incremental paper, not a breakthrough. The new bit is simple and plausible: use a lightweight biomedical NER as a 'weak expert' whose extracted tokens become hard constraints in an LLM rewrite prompt, and pair that with DPO to distill the behavior into a single model. That combination is genuinely new in the clinical augmentation literature, and the paper is honest about the trade-off between under- and over-detection.\n\nWhat it does well: the framework is clearly described, the choice of baselines (naive paraphrase, CATO style-only) is reasonable, and the downstream evaluations cover several tasks and two model architectures. The DPO extension is a nice addition. The paper also admits its main weakness in Section 6: if the expert under-detects, medical facts change; if it over-detects, augmentation variety collapses. That is the right kind of self-awareness.\n\nThe soft spots are real, and they are the headline ones. Preservation Rate and Hallucination Rate are computed with the same NER model that generates the constraint set K_i. So the method is scored on whether it keeps the exact tokens it was explicitly told to keep, while baselines that legitimately paraphrase (e.g., 'hypertension' to 'high blood pressure') are penalized. That makes the Table 1 safety advantage partly by construction. The appendix also mentions an acceptance filter on PR/HR; if that filter is applied only to Ours, the numbers are selected, not representative. Token preservation also doesn't guarantee label preservation: the LLM can keep the entity string but change negation, scope, or temporality. On downstream tasks, Ours is often best on mean, but the gains are small relative to the reported standard errors and there are no significance tests. The pattern is consistent enough that I doubt the method is useless, but the utility claim rests on the same circular safety metrics.\n\nFor a reader: this is worth a look if you work on LLM-based augmentation for clinical text or safety-constrained synthetic data. It deserves a proper review, not a desk reject. The fixes are addressable: independent entity ground truth (or human annotation), equal treatment of baselines, and significance testing. With those, the paper could be solid. As it stands, I'd treat the central safety claim as plausible but unproven.","headline":"A plausible incremental method for safer LLM-based clinical augmentation, but the headline safety numbers are partly circular and the downstream gains need significance testing.","tokens_in":17433,"tokens_out":1816,"would_cite":true,"duration_ms":15358,"reading_group":"yes","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 having a lightweight clinical named-entity recognition model extract medical tokens and requiring the LLM to preserve them during rewriting yields safer augmented clinical notes and consistent gains on downstream…","keywords":["clinical text augmentation","large language models","named entity recognition","model collaboration","hallucination prevention","clinical prediction","MIMIC-III","preference optimization"],"falsifier":"Build an augmented corpus from clinical notes whose label-bearing content includes entities the NER model is known to miss, such as negation cues (\"no fever\"), rare drug-dose pairs, or abbreviations, and have clinicians judge whether each augmented note still supports its original label. If a substantial fraction of notes change meaning while Preservation Rate and Hallucination Rate look acceptable, the token-preservation assumption fails; the same test in the over-detection direction should show that near-zero style variation produces no downstream gain.","tokens_in":16507,"feed_emoji":"🏥","tokens_out":8946,"duration_ms":72896,"temperature":0.7,"pith_summary":"Data augmentation for clinical text is unsafe when a general-purpose LLM decides on its own which parts of a note matter: paraphrases delete diagnoses, symptoms, and medication details, and add content that was never in the record. This paper proposes a query-based model collaboration in which a lightweight biomedical named-entity recognition model first extracts the medical tokens that must not change, and the LLM is then instructed to rewrite the note while preserving those tokens verbatim. Across MIMIC-III notes, the authors report that this guidance raises entity preservation and cuts hallucinated entities compared with naive paraphrasing and with CATO, a style-only augmentation method. If the result holds, small expert models can act as a practical safety layer on top of strong LLMs for high-stakes domains, improving downstream clinical predictions without retraining the LLM or paying for human annotation.","feed_headline":"A small term detector reins in LLM clinical note rewrites","feed_subtitle":"Preserving drug names and symptoms during LLM paraphrasing cuts medical hallucinations and boosts downstream predictions.","key_machinery":"The load-bearing object is the query-based model collaboration loop. A weak expert $W(\\cdot)$ extracts safety-critical tokens $K_i = W(x_i)$; a prompt template passes both the original text and $K_i$ to a strong generalist $G(\\cdot)$; and $G$ generates $\\tilde{x}_i = G(\\mathrm{prompt}(x_i, K_i))$ under an explicit instruction to alter no token in $K_i$. The paper instantiates $W$ as a DistilBERT-based biomedical NER model and $G$ as an instruction-tuned LLM (Qwen-3 or Llama-3), and evaluates quality with Preservation Rate and Hallucination Rate. The framework's extension replaces the two-model loop with a single \"strong expert\" trained by direct preference optimization on expert-guided versus naive rewrites.","core_discovery":"The central discovery is that the token-level constraint, not the LLM's own medical knowledge, is what keeps augmented notes faithful. Given an original note $x_i$, the weak expert returns $K_i = W(x_i)$; the prompt tells the strong generalist $G$ to rewrite $x_i$ while keeping every token in $K_i$ verbatim; the output $\\tilde{x}_i$ is paired with the original label $y_i$ to form the augmented dataset. On 300 MIMIC-III samples the paper measures Preservation Rate rising from 0.48-0.62 for naive or CATO augmentation to 0.66-0.79 under the collaborative method, and Hallucination Rate falling from 0.59-0.77 to 0.33-0.43. Training downstream models on these notes improves readmission accuracy from 0.552 (CATO) to 0.599 with Qwen-3, and improves mortality and length-of-stay metrics; zero/few-shot phenotyping and ICD coding retain or improve signal while baseline augmentations degrade. The authors also show that the dual-model guidance can be distilled into a single model via preference learning, though with less consistent gains across backbones.","pith_inferences":["The paper leaves implicit that this framework is portable: any domain where a small extractor can flag invariant tokens (contract clauses, regulated financial terms) could reuse the same prompt-collaboration recipe, because the LLM never needs retraining.","A testable extension the authors mention only in passing is to use PR/HR thresholds as an acceptance gate, selecting only augmented notes that pass before training; that would turn the method into a data-selection pipeline rather than a pure generator.","A limitation not explored in the paper is contextual meaning: tokens like \"no fever\" or \"without chest pain\" carry negation that verbatim token preservation cannot protect, so a dependency-aware constraint would be needed to fully guarantee label preservation."],"forward_implications":["Augmented clinical notes produced under expert-token constraints are safer training data than naive or CATO rewrites, with higher Preservation Rate and lower Hallucination Rate on the reported benchmarks.","Downstream clinical predictors trained on these notes improve readmission and mortality accuracy and reduce length-of-stay RMSE compared with baselines, and the gains replicate across Qwen-3 and BERT backbones.","Zero-, one-, and few-shot phenotyping and ICD coding retain or improve performance on augmented notes, whereas naive and CATO augmentations degrade these tasks.","A single LLM fine-tuned with direct preference optimization can partially reproduce the dual-model guidance, giving an alternate deployment mode when medical priors are already present.","Even a general-purpose entity extractor as the weak expert improves augmentation quality, indicating the framework tolerates imperfect expert signals."],"supporting_citations":[{"why":"Supplies the biomedical NER \"weak expert\" that extracts the preserved token set $K_i$.","marker":"(Raza et al., 2022)"},{"why":"Defines the CATO baseline and the causal formulation of augmenting only non-causal features $U$, which the paper argues is unsafe without supervision.","marker":"(Feder et al., 2023)"},{"why":"Provides the MIMIC-III clinical notes used in all augmentation-quality and downstream evaluations.","marker":"(Johnson et al., 2016)"},{"why":"Defines the Preservation Rate and Hallucination Rate metrics used to measure synthetic-note safety.","marker":"(Liu et al., 2024)"},{"why":"Supplies the retrieval-based framing for zero-shot ICD coding evaluation.","marker":"(Boyle et al., 2023)"},{"why":"Provides DPO, used to distill the dual-model collaboration into a single strong-expert model.","marker":"(Rafailov et al., 2024)"},{"why":"Supplies the Llama-3 strong generalist and inference models used in experiments.","marker":"(Grattafiori et al., 2024)"},{"why":"Supplies the Qwen-3 models used as default strong generalist and downstream backbone.","marker":"(Yang et al., 2025)"}],"fun_headline_variants":["Expert-guided LLM rewriting preserves clinical facts","Token-level leash stops clinical note hallucinations","Collaborative augmentation lifts clinical prediction accuracy","Keeping drug names verbatim reduces medical hallucinations","Dual-model notes beat single LLM for clinical training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the token set $K_i = W(x_i)$ extracted by the biomedical NER model is exactly the label-relevant content of the note, so preserving those tokens verbatim leaves the label $y_i$ unchanged; if the extractor misses a label-bearing term, the LLM can quietly alter it and the augmented pair becomes mislabeled.","fun_headline_variants_meta":{"raw":{"variants":["Expert-guided LLM rewriting preserves clinical facts","Token-level leash stops clinical note hallucinations","Collaborative augmentation lifts clinical prediction accuracy","Keeping drug names verbatim reduces medical hallucinations","Dual-model notes beat single LLM for clinical training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000232,"raw_usage":{"total_tokens":1477,"prompt_tokens":920,"completion_tokens":557,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":490}},"tokens_in":536,"tokens_out":557,"duration_ms":4877,"temperature":1.0,"reasoning_tokens":490,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:45:56.341165+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build an augmented corpus from clinical notes whose label-bearing content includes entities the NER model is known to miss, such as negation cues (\"no fever\"), rare drug-dose pairs, or abbreviations, and have clinicians judge whether each augmented note still supports its original label. If a substantial fraction of notes change meaning while Preservation Rate and Hallucination Rate look acceptable, the token-preservation assumption fails; the same test in the over-detection direction should show that near-zero style variation produces no downstream gain.","supporting_citations":[{"cited_title":"Large-scale application of named entity recognition to biomedicine and epidemiology","cited_arxiv_id":null,"evidence_quote":"Supplies the biomedical NER \"weak expert\" that extracts the preserved token set $K_i$."}],"review_version":2}