{"id":"c5a342a8-7d7f-4641-a4b7-61f5d7c0e74c","arxiv_id":"2609.15855","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"K-Bench provides a validated, protected benchmark for evaluating LLM safety in high-risk mental health conversations across four risk domains with 125 model configurations.","lead":"K-Bench is a clinician-calibrated benchmark that evaluates how well large language models handle high-risk mental health conversations across suicide, self-harm, domestic violence, and substance misuse. It uses 200 multi-turn vignettes and a frozen GPT-4o judge that matched clinician consensus on 94.2% of items, enabling large-scale safety comparisons across 125 model configurations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Judge–consensus agreement of 94.2% is reported only for the full calibration set; the holdout subset reserved for validation is not separately evaluated, raising the risk of overfit.","rationale":"The reader's weakest assumption was the realism of synthetic patients in high-risk scenarios. While important, the paper acknowledges that limitation transparently, and the benchmark's utility does not collapse if realism in high-risk turns is imperfect, because the judge is calibrated on clinician-consensus ratings of the same synthetic transcripts. A more consequential gap is the lack of separate holdout validation for the automated judge agreement claim. The 94.2% figure is central to the assertion that the judge can replace clinician raters at scale. Reporting it only for the full calibration set—when prompt tuning used a held-out portion—leaves open the possibility of overfitting. However, this concern does not fundamentally invalidate the benchmark; it calls for more complete reporting. The reader's ACCEPT verdict remains appropriate, but the underlying evidence for judge reliability is slightly weaker than portrayed. Hence UNCHANGED with this caveat noted.","tokens_in":25160,"tokens_out":3093,"duration_ms":32934,"concrete_test":"Re-run the frozen GPT-4o judge on the specific holdout subset of clinician-rated transcripts (those explicitly excluded from prompt tuning). Compute exact agreement with clinician consensus on D1–D7 eligible items for this holdout subset alone. If holdout agreement is more than 2 percentage points below the reported 94.2% (i.e., below 92.2%), the calibration claim needs qualification and the judge's applicability to the full evaluation cohort is less certain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central validation claim for the automated judge is 94.2% exact agreement with clinician consensus. The Methods state clearly that 'Prompt-tuning transcripts were separated from a holdout set reserved for downstream validation; the holdout data were not used during prompt iteration' (Online Methods, Automated judge calibration and scoring). However, the Results report judge agreement only for the pooled 'primary D1–D7 calibration set' without any breakdown into prompt-tuning vs. holdout subsets. Because the judge prompt was iterated on a known subset, the overall figure may include instances that were indirectly tuned, yielding an optimistic estimate. The holdout performance is the true measure of generalization to unseen transcripts, which is exactly what the judge must do when scoring the 200-vignette evaluation cohort for the leaderboard. Without separate holdout agreement, the reader cannot assess whether the judge's reliability extends beyond the data used during prompt development. This is the most load-bearing concern because the entire benchmark's scalability depends on the automated judge's performance on new, unoptimized conversations.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"K-Bench is a clinician-calibrated benchmark for evaluating large language models in multi-turn, high-risk mental-health conversations. It couples a factorial vignette library (200 vignettes; up to 20 talk turns; suicide, self-harm, domestic violence, substance misuse, and no-risk presentations) with a seven-dimension, 47-component rubric, a clinician-consensus ground truth built from 478 rating rows over 151 transcripts (16,157 consensus cells), and a frozen GPT-4o judge. The headline validation figures are 89.4% clinician-clinician exact agreement (22,816 pairwise comparisons) and 94.2% judge-consensus exact agreement (6,751 eligible items). Synthetic patient turns are compared against the real-user Amanda corpus with embedding metrics (nearest-neighbour cosine distance 0.404-0.486, MMD 0.017-0.024, Frechet distance 0.056-0.076) under the no-risk condition. The paper then reports leaderboard results for 125 model-configuration entries (33 base models, 14 providers): overall scores 81.19-98.96, combined-risk scores 52.39-96.11, with D2 risk exploration the most discriminating dimension (mean 79.9; range 22.1-93.1). Matched paired analyses show configuration-specific rather than uniform effects of therapeutic prompting (61 pairs; mean +0.11 points) and elevated reasoning (30 pairs; mean -0.58 points), with model-cluster bootstrap intervals crossing zero.","tokens_in":25648,"tokens_out":14724,"duration_ms":113843,"significance":"If the judge's calibrated agreement generalises to unseen transcripts, K-Bench is a substantial step forward for safety evaluation of conversational AI in mental health: it is the largest clinician-calibrated, configuration-scale comparison to date in this domain, it covers co-occurring and gradually emerging risks with a granular rubric, and its protected test architecture is a principled answer to the contamination problem that open benchmarks face. The paper is unusually candid about its own limits: the no-risk-only realism reference, the not-applicable gating of unreached items, the 20-turn cap, and the descriptive rather than causal reading of severity and comorbidity differences are all stated explicitly. Methodological strengths include independent clinician ratings with adjudicated consensus, a deterministic and frozen judge, matched-pair comparisons with model-cluster bootstrap confidence intervals, and a versioned public leaderboard with full configuration metadata.","major_comments":[{"comment":"The central validation statistic, 94.2% judge-consensus agreement on 6,751 eligible items, is reported only for the pooled D1-D7 calibration set. The Online Methods state that 'Prompt-tuning transcripts were separated from a holdout set reserved for downstream validation; the holdout data were not used during prompt iteration,' but the Results give no agreement figure (or even the number of transcripts) for that holdout, and no explanation of how the pooled figure decomposes into tuning and holdout subsets. Because the judge's production role is to score unseen transcripts in the 200-vignette leaderboard cohort, the holdout agreement is the direct estimate of the generalization on which the benchmark depends; iterating the prompt on part of the data and reporting only the pooled number leaves the possibility that the headline figure is optimistic. Please report the holdout size and its judge-consensus agreement overall and by rubric section, or justify why the pooled figure suffices.","section":"Results, 'Automated-judge calibration'; Online Methods, 'Automated judge calibration and scoring'"},{"comment":"The realism validation compares synthetic turns with the Amanda corpus only under the no-risk condition, because, as the manuscript itself states, the reference corpus 'contained no-risk conversations because users presenting with risk were not eligible to continue in the original studies.' The claim that 'these findings establish synthetic patients as a realistic and controlled basis' for evaluating high-risk conversations is therefore an extrapolation: the suicide, self-harm, domestic-violence, and substance-misuse synthetic conditions are never compared with any real-risk reference. The rubric already contains clinician-rated patient-context items ('whether the patient's responses were realistic,' 'whether the emotional tone was believable') that were collected on the 151 transcripts but excluded from the leaderboard composites; reporting aggregate clinician ratings on those items would provide a direct, within-study anchor for the realism of risk-active synthetic patients. Alternatively, the Discussion claim that the overlap data support the realism of risk-active patient language should be softened.","section":"Results, 'Synthetic-patient realism'; Online Methods, 'Embedding-based realism assessment'"},{"comment":"The claim that the judge 'reproduced the clinician-derived scoring standard with greater item-level consistency than clinicians achieved with one another' compares judge-consensus agreement (94.2%) with clinician-clinician pairwise agreement (89.4%). These two targets are not directly comparable: consensus is the modal or central response, so a judge that matches the modal value will mechanically out-agree an average pair of clinicians whenever ratings split 2-1 (the judge agrees with two of the three raters, whereas the clinicians agree with one another only one time in three on such items). At least part of the reported 4.9 percentage-point advantage is therefore a property of the comparison target rather than evidence of the judge's superior consistency. I suggest additionally reporting the judge's agreement with individual clinician ratings (judge versus the 478 raw rating rows) or the agreement of a re-derived consensus with held-out clinicians, so that judge and human performance are measured against comparable targets.","section":"Table 1 note; Results, 'Automated-judge calibration'"},{"comment":"The judge is openai/gpt-4o, a model from one of the providers whose systems occupy the top leaderboard ranks, and the patient agent used for the public evaluation is also GPT-4o. Self-preference bias in LLM-as-judge evaluation is a documented phenomenon in the evaluation literature, so the cross-provider comparison rests on the assumption that the frozen judge's agreement with clinician consensus does not vary with the identity of the target model. No reported analysis addresses this: judge-consensus agreement is stratified by rubric section and risk domain but not by the provider or family of the model that generated each calibration transcript. A concrete check would be to stratify the 6,751 judge-consensus comparisons by target model provider and show that agreement and the distribution of discrepancies are stable; if the calibration transcripts came from a small or unrepresentative set of target models, that should be stated as a limitation.","section":"Results, 'Leaderboard composition and score meaning'; Online Methods, 'Automated judge calibration and scoring'"}],"minor_comments":[{"comment":"Several passages have lost spaces between words, e.g., 'K-Benchevaluatessevendimensions' in the Main text, 'itmeasuredrubric-concordantconversation' in the Discussion, and 'Reasoningrefersheretoaprovider-exposedinference-timeconfiguration' in the Online Methods; these appear to be typesetting artifacts and should be corrected.","section":"Throughout (typesetting)"},{"comment":"Supplementary Table 5 is titled 'Full Matched Effects of Therapeutic Prompting' but lists 16 pairs, whereas the text reports 61 matched default-versus-therapeutic pairs and a stricter 43-pair analysis; please clarify that the table is a selected subset and state where the complete 61-pair dataset is published.","section":"Supplementary Table 5; Results, 'Therapeutic-prompt effects'"},{"comment":"The top overall scores on the leaderboard differ by only 0.01-0.03 points (ranks 1-3 of Supplementary Table 1), and the bootstrap intervals in Figure 2 presumably overlap for such configurations; please report rank-stability statistics (e.g., the proportion of bootstrap replicates preserving adjacent ranks) or explicitly warn that near-tied overall rankings do not support ordinal comparisons.","section":"Figure 2; Supplementary Table 1"},{"comment":"Exact agreement on the four-category Q1 severity item is only 65.7% among clinicians for suicide and 75.3% overall; reporting adjacent-category or weighted agreement for Q1 as well would clarify how much of the disagreement is ordinal rather than categorical, and would place the judge's 82.5% figure in a fairer context.","section":"Table 1; 'Risk-severity operationalisation' in Online Methods"},{"comment":"The realism reference corpus (Amanda) is described only by a citation to a manuscript under review (Vowels et al., 2026a); please state where the de-identified corpus is archived or how reviewers can access it, since the embedding-based realism comparison cannot otherwise be reproduced.","section":"Online Methods, 'Embedding-based realism assessment'"},{"comment":"Figure 4 and the surrounding text use 'blended text cost' without defining the input/output token mix and pricing assumptions; a one-sentence definition in the Online Methods would make the cost comparison reproducible.","section":"Figure 4; Online Methods, 'Data analysis and reporting'"}],"recommendation":"major_revision","confidential_remarks":"A major revision that is close to acceptance: the central claims are defensible, and the requested additions (holdout-only agreement, provider-stratified judge agreement, and clinician-rated realism aggregates for risk-active transcripts) all use data the authors already hold. Two editorial points. First, the protected design means no reader can verify the headline validation statistics; I suggest requesting an embargoed review package (transcripts, consensus records, and judge outputs) for reviewers under a confidentiality agreement, in parallel with the holdout analysis. Second, the disclosure of interests is transparent and adequate, but given that the judge is an OpenAI model and OpenAI configurations occupy the top leaderboard ranks, the provider-stratified agreement analysis should be treated as a condition of acceptance rather than optional polish, to pre-empt any appearance of evaluator-model preference."},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Two things you should know. K-Bench is the first mental-health benchmark I've seen that combines four co-occurring risk domains with multi-turn conversations up to 20 turns, 125 model configurations, a clinician-calibrated judge, and protected test materials. Second, the validation is generally more careful than most work in this area—but the headline judge-agreement number is reported only on the full calibration set, and that is the soft spot worth pushing on.\n\nWhat is genuinely new: nobody else has evaluated this many configurations (33 base models, 14 providers) on a fixed cohort of risk vignettes that include co-occurring and gradually disclosed risk. The factorial vignette expansion from lived-experience-derived source material is a real effort to control clinical complexity, not just a generic persona bank. The rubric is granular: 47 components across seven dimensions, with risk exploration separated from clinical judgement. That separation pays off—D2 is the discriminating dimension, and that is a clinically meaningful finding rather than an artifact.\n\nThe judge validation is thorough in most respects. Six clinicians produced 22,816 pair comparisons at 89.4% exact agreement; the frozen GPT-4o judge reached 94.2% on 6,751 item comparisons, with the expected pattern of lower agreement on severity and higher on clinical judgement and global dimensions. The matched prompting and reasoning comparisons are also done carefully, with model-cluster bootstrap intervals and clean baselines that exclude relabelled provider defaults.\n\nWhere I'd push back: the stress-test note is right. The Methods say prompt-tuning transcripts were separated from a holdout set, but the Results report only the pooled 'primary D1–D7 calibration set.' Since the judge prompt was iterated on a known subset, the pooled figure could be optimistic, and we cannot see how the judge generalizes to the holdout—which is exactly what it does on the live benchmark. The fix is simple: report judge–consensus agreement on the holdout subset separately. I suspect it remains strong, but the paper should not ask readers to take the pooled number on faith.\n\nThe realism validation is honest but limited: the reference corpus contained no-risk conversations, so the embedding overlap evidence supports the no-risk condition, and the authors say the high-risk generalization is an assumption. That is a limitation, not a flaw, and it is acknowledged. Bigger structural issue: all operational materials, judge prompts, and code are withheld, which is defensible for a protected benchmark but means external verification is impossible at item level. For a benchmark whose entire value is its validity, a little more disclosure—for example, holdout agreement, per-item confusion matrices, more transparency about the 'stealth/ox-alpha' entries—would strengthen trust.\n\nWho should read this: anyone making deployment decisions about LLMs in mental-health contexts, regulators, and benchmark designers. It deserves a serious referee. My recommendation: send it to peer review, and make the holdout breakdown a required revision rather than a suggestion.","headline":"A clinically serious, well-validated mental-health benchmark worth refereeing, with one number that needs to be reported properly: the judge's held-out agreement, not just the pooled calibration figure.","tokens_in":26192,"tokens_out":1752,"would_cite":true,"duration_ms":16642,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"K-Bench provides a clinically calibrated benchmark that evaluates large language models in high-risk mental health conversations using synthetic patients and an automated judge that agrees with clinicians on 94.2% of items.","keywords":["large language models","mental health","benchmark","risk assessment","synthetic patients","clinician calibration","conversational AI","patient safety"],"falsifier":"Recording real high-risk conversations from crisis helplines or mental health services and comparing them (after de-identification) with synthetic conversations for the same risk presentations would directly test whether the synthetic patient language and disclosure patterns generalize to genuine high-risk interactions.","tokens_in":26801,"feed_emoji":"🧠","tokens_out":4511,"duration_ms":45716,"temperature":0.7,"pith_summary":"K-Bench is a benchmark designed to evaluate how safely large language models handle mental health conversations where risk can emerge gradually. The paper establishes that synthetic patient conversations, generated from lived-experience-derived material and a factorial design, are realistic enough to substitute for real users when compared with a no-risk reference corpus. It then shows that a frozen GPT-4o automated judge, calibrated against the consensus of six clinicians, reproduces clinician ratings with higher consistency than the clinicians themselves. Using this evaluation system on 125 model configurations from 33 models, K-Bench finds that leading models score above 95 on a composite of clinical judgement and risk exploration, while risk exploration separates strong from weak models more sharply than other competencies. These results matter because people increasingly turn to chatbots for mental health support, and existing benchmarks test single-turn or single-risk scenarios.","feed_headline":"K-Bench: 94% agreement with clinicians in mental health risk conversations","feed_subtitle":"A benchmark evaluates 125 model configurations across four risk domains using synthetic patients and a frozen GPT-4o judge.","key_machinery":"The central object is K-Bench itself: a protected benchmark pipeline that combines (1) a 122-variable factorial vignette framework derived from lived-experience material, (2) synthetic patient-agent conversation generation with validated realism, (3) a 47-component clinician-developed rubric across seven dimensions, (4) item-level clinician consensus ground truth from 151 transcripts, and (5) a frozen GPT-4o automated judge calibrated to that consensus. The benchmark evaluates each model configuration on the same 200-vignette cohort in conversations of up to 20 turns, and withholds operational test materials to prevent gaming.","core_discovery":"The paper develops and validates K-Bench, a clinician-calibrated benchmark that evaluates LLM performance across seven dimensions of conversational competence (clinical judgement, risk exploration, ethical reasoning, supportive conversation, psychological knowledge, cultural competence, and autonomy and boundaries) using 200 multi-turn vignettes spanning suicide, self-harm, domestic violence, substance misuse, and no-risk presentations. The central discovery is that a frozen GPT-4o judge, when prompted with the same rubric used by clinicians, achieves 94.2% exact item-level agreement with the clinician consensus across 6,751 eligible comparisons, outperforming the 89.4% agreement among the clinicians themselves. This makes large-scale, reproducible evaluation feasible without requiring a clinician to read every transcript.","pith_inferences":["If synthetic patients validated only under no-risk conditions do not fully generalize to high-risk conversations where language may be more urgent or culturally specific, then the realism guarantees may be weaker than the benchmark claims. A direct empirical test using real high-risk conversations from crisis helplines would settle this.","The benchmark's restriction to API configurations rather than complete consumer products means that real-world safety may differ when moderation, memory, and escalation layers are present.","The 20-turn cap may miss how models handle deferred risks in longer interactions; future iterations with more turns could reveal additional failure modes.","The strong judge-clinician agreement suggests the rubric captures a reproducible standard for contemporaneous risk formulation, but not for predicting actual suicide attempts or self-harm episodes, which remains a separate challenge."],"forward_implications":["K-Bench provides a standard that future models can be measured against, enabling direct comparison of safety performance across providers, versions, prompts, and reasoning settings.","The benchmark's risk exploration dimension exposes safety-relevant weaknesses that overall scores alone would hide, guiding developers toward targeted improvements.","Therapeutic prompting helps weaker models but not stronger ones, suggesting that prompt engineering alone is insufficient for clinical safety.","The finding that elevated reasoning does not consistently improve risk performance implies that inference-time compute investments may not translate to safer mental health conversations.","Protected operational materials allow the benchmark to remain valid for new model releases, tracking genuine progress over time."],"supporting_citations":[{"why":"VERA-MH: prior benchmark for suicide risk evaluation in multi-turn conversations that K-Bench extends to broader clinical coverage and larger model comparison.","marker":"Bentley et al. 2026"},{"why":"SIM-VAIL: clinician-calibrated automated auditing for mental health chatbots, providing prior validation approach that K-Bench builds upon with more risk domains and configurations.","marker":"Weilnhammer et al. 2026"},{"why":"Amanda corpus: real human-AI conversations used as the reference distribution for assessing synthetic patient realism.","marker":"Vowels et al. 2026a"},{"why":"Lived-experience psychosocial-risk material: source data for vignette narratives that ground the benchmark in genuine patient experiences.","marker":"Vowels et al. 2026b"},{"why":"AEF-1 standards: guidelines for independent third-party AI evaluation that justify the withholding of operational test materials to prevent benchmark gaming.","marker":"Stosz et al. 2025"},{"why":"Provides clinical framework for safe conversational AI in psychotherapy, informing the rubric dimensions and evaluation criteria.","marker":"Miner et al. 2019"}],"fun_headline_variants":["GPT-4o judge hits 94% clinician agreement on mental health risk","125 LLMs scored: frozen GPT-4o matches experts 94%","K-Bench: LLM safety benchmark achieves 94% clinical calibration","Benchmark reveals LLM risk gaps: GPT-4o judge 94% accurate"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The realism of synthetic patient conversations is validated only against a no-risk reference corpus (the Amanda dataset, which excluded users presenting with risk), yet the benchmark assumes that embedding-space overlap under no-risk conditions implies similar realism for conversations involving suicide, self-harm, domestic violence, and substance misuse.","fun_headline_variants_meta":{"raw":{"variants":["GPT-4o judge hits 94% clinician agreement on mental health risk","125 LLMs scored: frozen GPT-4o matches experts 94%","K-Bench: LLM safety benchmark achieves 94% clinical calibration","Benchmark reveals LLM risk gaps: GPT-4o judge 94% accurate"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000593,"raw_usage":{"cost":0.000593488,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000593488,"upstream_inference_prompt_cost":9.7328e-05,"upstream_inference_completions_cost":0.00049616},"total_tokens":2672,"prompt_tokens":900,"completion_tokens":1772,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":1687}},"tokens_in":900,"tokens_out":1772,"cost_source":"openrouter_usage","duration_ms":17062,"image_pages":[],"temperature":1.0,"reasoning_tokens":1687,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-09-15T04:11:34.646679+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"falsifier":"Recording real high-risk conversations from crisis helplines or mental health services and comparing them (after de-identification) with synthetic conversations for the same risk presentations would directly test whether the synthetic patient language and disclosure patterns generalize to genuine high-risk interactions.","supporting_citations":[{"cited_title":"2025 , note =","cited_arxiv_id":null,"evidence_quote":"AEF-1 standards: guidelines for independent third-party AI evaluation that justify the withholding of operational test materials to prevent benchmark gaming."}],"review_version":1}