{"id":"0a50ac78-9410-4d4d-a8e9-014494c8c0c7","arxiv_id":"2607.13036","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A new benchmark of 1,048 psychiatric notes shows LLMs frequently diagnose too early under incomplete evidence, and safety prompting only trades premature diagnoses for excessive abstention.","lead":"Safe-Psych is a new benchmark that feeds psychiatric clinical notes to LLMs piece by piece and checks whether the model asks for more information, diagnoses, or abstains. Across nine models, it finds that LLMs often commit to diagnoses before enough evidence exists, and that \"be careful\" prompts mostly shift errors into over-abstention.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Retrospective sufficiency labels are the unvalidated pivot: all sequential timing/calibration claims depend on them, yet the protocol lets annotators see the full note before marking the earliest sufficient section.","rationale":"The reader's weakest-assumption analysis points to the retrospective 'earliest sufficient section' labels, and this is indeed the most load-bearing threat to the paper's sequential claims. The labels are generated after full-note review, so they are a retrospective counterfactual about when a clinician could have diagnosed, not a record of prospective clinical decisions. The paper itself discloses this in Appendix B.6, which is a point in its favor, but disclosure does not remove the dependency: every timing category, the expert sufficiency curve, and the premature-versus-on-time accuracy comparison are computed relative to these labels. I agree with the reader that this is the central vulnerability. The concern is not fatal, for two reasons. First, the paper's own direction-of-bias analysis suggests the benchmark is lenient: if anything, it would under-count premature diagnoses, so the qualitative finding that models diagnose early is likely conservative. Second, the under/over-abstention tradeoff in Figure 3 is based on the final sufficient/insufficient judgment, not on the earliest-section timing, so that part of the central claim is less exposed. But the precise quantitative claims—that most diagnoses are premature, that on-time diagnoses are more accurate, and that abstention prompting shifts the timing curve—remain conditional on the validity of the retrospective thresholds. A prospective annotation study would settle the question directly. Given that the reader already assigned CONDITIONAL and flagged this exact issue, my stress-test does not move the verdict; it reinforces the condition. The paper should either add prospective validation or soften the precise timing/accuracy claims to reflect the known retrospective bias.","tokens_in":20615,"tokens_out":9317,"duration_ms":97332,"concrete_test":"Draw a random sample of ~200 cases. Present sections incrementally to independent psychiatrists who cannot see future sections, and at each step ask whether they would diagnose, clarify, or abstain. Compare the prospective 'sufficient to diagnose now' section against the retrospective earliest-sufficient-section label used in the benchmark. Recompute the timing shares (Figure 5), cumulative sufficiency curves (Figure 6), and the premature-versus-on-time accuracy comparison (Figure 7) using the prospective labels. If timing category shares shift by more than 5 percentage points or the premature/on-time accuracy gap reverses or attenuates materially, the retrospective-label concern lands and the benchmark's central calibration claims need re-qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The sequential benchmark's central timing and calibration results rest on a single retrospective label: the earliest section at which psychiatrists, after viewing the full note, judged information sufficient for diagnosis (Sec. 2.3, Fig. 2). The protocol explicitly gives annotators the complete record before they mark sufficiency, and Appendix B.6 concedes this 'may mark information as sufficient earlier than it would be judged in a prospective clinical setting, making the benchmark relatively lenient.' Disagreements on the earliest sufficient section are resolved by taking the earlier annotation (Sec. 2.3), which further shifts the label distribution toward early sufficiency. This directly determines which model actions count as premature versus on-time in Figures 5-7 and shapes the headline claim that premature diagnoses are less accurate. The paper argues the bias is conservative for already-premature labels, but it does not quantify how many diagnoses at the boundary are misclassified, and it does not validate the labels against prospective decision points. If true prospective sufficiency thresholds are later, the observed premature rates are understated and the Figure 7 accuracy gap may be distorted; if some labels are also too late, the premature group is contaminated by appropriate early diagnoses. The final under/over-abstention tradeoff in Figure 3 is less affected because it uses only the final sufficient/insufficient decision, but the sequential timing story and the 'on-time versus premature' accuracy comparison depend entirely on the unvalidated counterfactual thresholds.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Safe-Psych is a sequential benchmark for evaluating LLM decision behavior under evolving clinical information in psychiatry. It releases 1,048 anonymized, translated psychiatric notes segmented into five sections, with psychiatrist-derived ground-truth labels for diagnosis, the earliest sufficient section, and sufficiency of the full note. Nine LLMs are evaluated under full-information and sequential settings, with responses mapped by a validated LLM judge to DIAGNOSE, CLARIFY, or ABSTAIN. The paper's main claims are that strong models under-abstain (diagnose despite insufficient evidence), rarely seek clarification unless explicitly prompted, diagnose prematurely relative to expert sufficiency labels, that premature diagnoses are less accurate than on-time ones, and that abstention-aware prompting shifts errors toward over-abstention rather than improving calibration.","tokens_in":20882,"tokens_out":6479,"duration_ms":69275,"significance":"If the central claims hold, Safe-Psych is a valuable resource: it addresses a gap left by static QA and abstention benchmarks by measuring how model behavior changes as evidence accumulates. The dataset is large, real-world, and externally annotated; the judge is validated against human annotations; prompting, evaluation, and reproducibility code are described in unusual detail. The distinction between capability and calibration, and the explicit measurement of over-abstention alongside under-abstention, are strengths. However, the sequential timing and calibration results depend on retrospectively assigned sufficiency labels, and the paper's own limitation section concedes that these may be systematically earlier than prospective clinical judgments. That pivot requires additional validation or sensitivity analysis before the headline timing claims can be regarded as fully established.","major_comments":[{"comment":"The expert stopping step tau_exp is the load-bearing label for all premature/on-time classifications, yet it was assigned after annotators viewed the full note, and disagreements were resolved by taking the earlier section. Appendix B.6 concedes this 'may mark information as sufficient earlier than it would be judged in a prospective clinical setting.' The paper argues the bias is conservative for cases already labeled premature, but it does not quantify boundary misclassification. Please add a sensitivity analysis: re-annotate a subset with section-by-section prospective judgments, or re-run the timing metrics with tau_exp shifted one section later and with adjudicated disagreements excluded. Without this, Figures 5–7 do not fully support the sequential calibration story.","section":"§2.3, Appendix B.6"},{"comment":"The claim that 'on-time diagnoses are consistently more accurate than premature ones' is presented without confidence intervals, case counts, or a statistical test. Since accuracy is computed only for cases where the model produced a diagnosis, the premature and on-time groups may differ systematically in case difficulty or diagnostic mix. A model that diagnoses early may do so disproportionately on easy cases, while a model that waits may only diagnose cases it is confident about. Please report n and 95% CIs for each timing group, and stratify by expert tau_exp stage or diagnosis category. A paired comparison using cases where the same model eventually reaches an on-time diagnosis would directly test whether waiting itself improves accuracy.","section":"§4.4, Fig. 7"},{"comment":"The under-abstention and over-abstention rates rely on only 9% insufficient cases (roughly 95 of 1,048). Error bars in Figure 3 are shown only for the four models with repeated seed runs; the remaining five models have no uncertainty quantification. Because decoding is stochastic and the insufficient-case denominator is small, the statement that 'under-abstention exceeds 60% for most models' needs binomial or bootstrap confidence intervals for every model, or at least a statement of whether single-run estimates for MedGemma-27B, Gemma-3-4B, Claude-Opus-4.6, and Gemini-Flash-2.5 are robust to sampling variability. The appendix currently gives repeated-run data only for the full-information setting, not for the sequential setting most central to the paper.","section":"§4.1, Fig. 3; §6"},{"comment":"The LLM judge is GPT-5.4, which is also one of the evaluated models. Its macro-F1 on the final action label is 0.84, below the human agreement of 0.89. The validation set of 149 responses is deliberately enriched for boundary cases, but no per-model or per-strategy judge agreement is reported. Since every behavioral metric passes through this judge, differential judge accuracy across models or inference strategies could confound cross-model comparisons. Please report judge agreement stratified by evaluated model (at least for GPT-5.4 vs. the other models), and include a sensitivity analysis using human labels on a random subset of responses.","section":"§3.4.1, Table 1, Appendix C.1.1"}],"minor_comments":[{"comment":"Please state explicitly whether an intermediate ABSTAIN terminates the trajectory or whether the model continues to receive subsequent sections. The current text records 'premature abstention' but does not define the resulting transition, which is needed for exact reproduction.","section":"§3.1"},{"comment":"The axis label 'Final S' is ambiguous; rename it 'Final section' to match the earlier sections S1–S4.","section":"Appendix Fig. 1b"},{"comment":"F10.2 (alcohol dependence syndrome) is listed twice in the diagnosis-specific agreement table; remove the duplicate.","section":"Appendix Table 3"},{"comment":"References [24] and [25] both cite the Qwen3 Technical Report with the same arXiv identifier; one should be removed or merged.","section":"References"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this is a genuinely useful artifact: a 1,048-note psychiatric benchmark that presents evidence incrementally and asks models to clarify, diagnose, or abstain. The authors validate their LLM judge against human annotations, report inter-rater agreement, release code and data, and disclose their limitations in unusual detail. Second, the main empirical claim—that strong models under-abstain under incomplete information and that safety prompts merely shift errors into over-abstention—is credible and likely robust. That finding alone makes the paper worth reading.\n\nThe under/over-abstention result (Figure 3) is the cleanest contribution because it depends only on the final sufficient/insufficient label, which has 90% raw agreement and is less sensitive to timing assumptions. The sequential results—clarification rates, premature-diagnosis timing, and the premature-vs-on-time accuracy gap—are more fragile because they rest on the psychiatrists' retrospective \"earliest sufficient section\" labels. Annotators saw the full note before marking that section, and disagreements were resolved by taking the earlier annotation. That biases the benchmark toward early sufficiency, which the paper correctly notes is lenient. In practice, this likely makes the premature-diagnosis finding conservative: some diagnoses labeled \"on-time\" would be premature under stricter prospective thresholds, so the accuracy gap in Figure 7 is probably understated, not overstated. Still, the paper does not quantify how much the timing or accuracy conclusions would shift under different threshold choices, and Figure 7 has no confidence intervals or significance tests. That is the main soft spot, and it is a moderate one, not fatal.\n\nOther soft spots are minor and mostly disclosed: only 9% of cases are insufficient-information cases, so under-abstention estimates for that subset are unstable; the judge reaches macro-F1 0.84 on action labels, so small cross-model differences may be noise; and the data come from one Romanian hospital with a 77% female sample, translated to English. None of these undermine the central story.\n\nWho is this for? Anyone building or evaluating clinical LLM safety benchmarks. It extends AbstentionBench and MedAbstain in a way that is more than cosmetic: the clarify option and the sequential disclosure actually change what you measure. I would bring it to a reading group and would cite it if I worked in this area.\n\nRecommendation: send it to peer review. The referees should ask for a sensitivity analysis on the sufficiency thresholds—for example, using the later annotation instead of the earlier one, or a small prospective validation—and for uncertainty intervals on Figure 7. But this is a solid, honest piece of benchmark work that deserves a serious referee, not a desk rejection.","headline":"A careful, useful benchmark with a credible main finding; the retrospective sufficiency labels make the timing results approximate, but the paper handles that honestly and deserves refereeing.","tokens_in":21438,"tokens_out":2306,"would_cite":true,"duration_ms":27909,"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":"LLMs commit to psychiatric diagnoses before the evidence is in, a new benchmark shows.","keywords":["LLM safety","psychiatric diagnosis","sequential evaluation","abstention","calibration","clarification-seeking","clinical decision support","benchmark"],"falsifier":"A prospective study in which psychiatrists judge sufficiency while reading a note section-by-section, without seeing later sections, could check whether the retrospective labels shift. If under real-time labeling premature diagnoses are no less accurate than on-time ones, or under-abstention drops below the reported rates, the central timing and calibration story would need revision.","tokens_in":20511,"feed_emoji":"🧠","tokens_out":3440,"duration_ms":34650,"temperature":0.7,"pith_summary":"Safe-Psych is a benchmark for whether language models know when not to diagnose in psychiatry. It presents 1,048 real clinical notes one section at a time, with psychiatrist-defined labels telling whether a model should diagnose, ask for clarification, or abstain at each step. Across nine models, the paper finds that capability does not track calibration: most models diagnose more than 60% of the cases that psychiatrists judged insufficient, even when explicitly allowed to abstain, and they rarely request more information unless prompted. Telling models to be cautious delays diagnoses and lowers premature commitment, but only by making them abstain on many cases that did have enough information. The paper argues that recognizing incomplete evidence is a separate skill from diagnostic accuracy and that current prompting does not supply it.","feed_headline":"LLMs commit to psychiatric diagnoses before evidence is in","feed_subtitle":"A 1,048-note sequential benchmark shows most models diagnose despite insufficient evidence, and caution prompts only shift the failure.","key_machinery":"The load-bearing object is the expert action sequence derived from psychiatrist annotations: for each case, the earliest section where information becomes sufficient yields CLARIFY for earlier steps, DIAGNOSE from that step onward, and ABSTAIN at the end if it never becomes sufficient. Model trajectories are judged by an LLM judge that extracts whether the response commits to a diagnosis, requests specific information, or gives none, and maps that to the same three actions. This setup converts a one-shot QA task into a timing-and-calibration task, letting the paper measure under-abstention (diagnosing insufficient cases), over-abstention (abstaining on sufficient cases), and premature versus","core_discovery":"The paper's central claim is that LLMs used for psychiatric decision support are poorly calibrated when clinical evidence arrives over time: they commit to diagnoses before enough information exists, under-abstain even when abstention is explicitly allowed, and only shift error patterns rather than improve them when prompted to be cautious. In the benchmark's sequential protocol, models that diagnose at the expert-defined 'sufficient' stage are consistently more accurate than those that diagnose early, so premature commitment carries a concrete accuracy cost. The authors find diagnostic capability and uncertainty calibration decouple: strong models can be accurate on full notes while still f","pith_inferences":["If the same failure generalizes beyond psychiatry, the sequential-disclosure design could be adapted to emergency triage or primary care, where information also arrives over time; that transfer is untested.","The under-/over-abstention tradeoff resembles a calibration frontier; future work could treat abstention thresholds as tunable and evaluate whether models can be calibrated per case difficulty rather than only per prompt.","Because sufficiency labels were made retrospectively, a prospective study where clinicians mark sufficiency in real time would test whether the reported premature-diagnosis accuracy gap survives stricter timing labels.","The benchmark's judge maps 'committed plus specific question' to CLARIFY; this definition likely rewards models that hedge a diagnosis with a question, so results may be sensitive to response style, an artifact worth testing."],"forward_implications":["Deploying current LLMs as intake or triage assistants in psychiatry risks early commitment: diagnoses made before evidence is sufficient are measurably less accurate than on-time ones.","Adding abstention instructions does not fix calibration; it just trades unsafe early diagnoses for missed ones, so prompting alone is not a safety mechanism.","Models generally will not ask for missing information on their own; clarification-seeking must be explicitly prompted or trained.","Diagnostic accuracy on full notes overstates real-world usefulness, because in practice evidence arrives piecemeal; sequential evaluation is needed to see the gap.","Psychiatrist disagreement on a case correlates with lower model accuracy, suggesting some benchmark errors reflect genuine case ambiguity rather than model fault alone."],"fun_headline_variants":["LLMs Jump to Psychiatric Diagnoses Before Evidence Is In","Asking Models to Be Cautious Just Shifts Their Mistakes","Psychiatric AI Diagnoses Too Early, New Benchmark Finds","Capability Doesn't Equal Calibration in Psychiatric LLMs","Premature Diagnoses Are Less Accurate, Says Safe-Psych"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The claim rests on treating psychiatrists' retrospective 'earliest sufficient section' labels as true decision points; if those labels mark information sufficient earlier than it would be in real time, premature diagnoses can be mislabeled as on-time or late.","fun_headline_variants_meta":{"raw":{"variants":["LLMs Jump to Psychiatric Diagnoses Before Evidence Is In","Asking Models to Be Cautious Just Shifts Their Mistakes","Psychiatric AI Diagnoses Too Early, New Benchmark Finds","Capability Doesn't Equal Calibration in Psychiatric LLMs","Premature Diagnoses Are Less Accurate, Says Safe-Psych"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000148,"raw_usage":{"total_tokens":1029,"prompt_tokens":753,"completion_tokens":276,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":206}},"tokens_in":497,"tokens_out":276,"duration_ms":3219,"temperature":1.0,"reasoning_tokens":206,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T14:46:23.691612+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A prospective study in which psychiatrists judge sufficiency while reading a note section-by-section, without seeing later sections, could check whether the retrospective labels shift. If under real-time labeling premature diagnoses are no less accurate than on-time ones, or under-abstention drops below the reported rates, the central timing and calibration story would need revision.","supporting_citations":[],"review_version":1}