DistressBench: Evaluating Multi-Dimensional Crisis Support in Large Language Models

Drew Rein; Patrick Oathout; Vishal Kumar; Udari Madhushani Sehwag

cs.CLOct 8, 2026

DistressBench: Evaluating Multi-Dimensional Crisis Support in Large Language Models

Authors: Drew Rein, Patrick Oathout, Vishal Kumar, Udari Madhushani Sehwag

Organizations: Scale AI · Work done while at Scale AI

Abstract

Language models have been adopted faster than evaluation practices that reliably assess their risks, and a growing share of use is affective rather than informational. Within that traffic users disclose suicide and self-harm (SSH) crises. Existing safety benchmarks score models primarily on refusal, but refusal is not the desired behavior here; the conversation requires a substantive supportive response. We introduce DistressBench, 718 English conversations with clinician-authored user turns spanning 23 SSH subcategories, each paired with a task-specific rubric of weighted atomic criteria across seven dimensions of crisis support and a clinician reference response. A model's score is the weighted average over the criteria its reply satisfies, reading directly as the fraction of required care delivered. The best of 25 frontier models satisfies 88.4% of weighted criteria and the median model 71.4%, against 98.7% for clinician references scored on the same criteria. Pooled across all 25 models, in 35.3% of conversations where a model satisfied every criterion for recognizing the crisis, it failed to route the user toward human help. The shortfall concentrates in the dimensions requiring the model to act rather than withhold. Contrasting each dimension's median model against its best separates capability limits from choices: de-escalation reaches 74.6% in the strongest model against a median of 46.0%, and safety disclaimers 84.6% against 8.3%. Multi-turn conversations score below single-turn ones (68.5% against 74.2%). We release the corpus, rubrics, judge calibration labels, and a companion annex of 20 long-horizon red-team transcripts. These results argue for evaluating crisis safety on the care a model delivers rather than on what it declines to say, and for treating de-escalation and resource routing as deployment requirements.

Figures & tables

Figure 1: Expert effort is concentrated in construction, and the clinician calibration set validates the judge panel rather than feeding the leaderboard. Fill encodes who produced each artifact.

Figure 2: Every model falls short of the clinician reference ceiling, and most of the roster is statistically inseparable. Weighted rubric score over 718 tasks, 95% bootstrap CIs; dashed lines mark gaps surviving a paired bootstrap. Ranks run left panel then right.

	Disclaimers	De-escalation	Resources	Identification	Compassion	Harmless	Moralization
Median	0.083	0.460	0.703	0.825	0.852	0.945	0.982
Best	0.846	0.746	0.866	0.933	0.969	0.970	0.992
$\Delta_{d}$	+0.763	+0.286	+0.163	+0.107	+0.117	+0.025	+0.010
Reading	Choice	Choice	Mixed	Mixed	Mixed	Saturated	Saturated

Table 1: The two largest deficits are each achieved by at least one deployed model, which is inconsistent with a frontier limit. Choice-capability gap

\Delta_{d}

(Section 3.2 ) by dimension over 25 models. Columns round independently, so

\Delta_{d}

can differ by

0.001

from their difference. Appendix A defines the dimensions.

Figure 3: Every model, including the leader, leaves users it has correctly recognized without a route to help. Share of fully recognized conversations that fail an actionable-resources criterion.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Subcategory	$n$	Subcategory	$n$
Help-Seeking Signals	41	Trauma-Triggered Distress	29
Multi-Domain Distress	41	Chronic Distress	27
Cognitive Patterns	41	Masked Distress	26
Suicidal Ideation	40	Acute Emotional Psychological Distress	22
Thanatophobia	40	Impulsivity	21
Suicide Planning & Preparation	40	Loneliness & Isolation	21

Appendix

Table 2: DistressBench subcategory counts (

n=718

Field	Values	Role
Subcategory	23 SSH types	Scenario diversity (e.g., NSSI, psychosis, grief)
Age	U18 / O18	Developmental framing (209 / 509 tasks)
Turn structure	Single / Multi (3–5 turns)	Short vs. sustained disclosure
Intent	Borderline / Harmful	Gray-area vs. explicit harm content
Progression	Ideation, Planning, Action, Post-Action	Acuity trajectory
Explicit type	Implicit / Explicit distress	Signal directness

Appendix

Table 3: DistressBench task metadata axes (English v1).

Dimension	What it measures
Identification	Recognizes user distress and specific distress type
Compassionate engagement	Empathy without sycophancy or excessive anthropomorphism
De-escalation	Actionable steps to reduce distress
Actionable resources	Correct, locale-appropriate crisis resources (e.g., 988)
Disclaimers	Non-clinician status; not a replacement for professional care
Moralization	Avoids shaming, condescension, or nannying

Appendix

Table 4: DistressBench rubric dimensions (v1).

Figure 4: DistressBench corpus metadata distributions (

n=718

		Clinicians			Panel
Dimension	Base rate	Agree	$\kappa$	AC1	$\kappa$	AC1
Actionable resources	0.64	0.895	0.771	0.806	0.845	0.868
De-escalation	0.45	0.809	0.614	0.623	0.673	0.672
Compassionate engagement	0.75	0.797	0.457	0.675	0.575	0.775
Identification	0.75	0.751	0.332	0.604	0.488	0.725
Harmless response	0.85	0.844	0.387	0.791	0.525	0.873

Appendix

Table 5: Agreement by dimension on the calibration set. Human columns are agreement among the three clinicians; panel columns are agreement with each clinician, averaged. Where a base rate is extreme,

\kappa

and AC1 diverge sharply and AC1 is the more faithful summary.

Model	Primary	LOO	$\Delta$	Artifact	Excess	Pooled excess
Muse Spark 1.2	0.884	0.815	$-0.069$	$-0.076$	$+0.007$	$-0.016$
Muse Spark 1.3	0.849	0.795	$-0.054$	$-0.076$	$+0.022$	$-0.001$
Muse Glimmer 30B	0.806	0.740	$-0.066$	$-0.076$	$+0.010$	$-0.013$
Claude Fable 5.1	0.716	0.653	$-0.062$	$-0.061$	$-0.001$	$-0.009$
Claude Opus 4.8	0.714	0.651	$-0.063$	$-0.061$	$-0.002$	$-0.010$
Claude Fable 5	0.694	0.632	$-0.062$	$-0.061$	$-0.001$	$-0.009$

Appendix

Table 6: No model gains materially from having its own provider family on the panel, including the leader. Leave-one-out scoring for the eleven models whose provider family sits on the judge panel. Excess is the drop beyond the artifact for the specific judge removed; negative excess indicates the own-family judge was favorable. The final column gives the excess that a single pooled

-0.053

artifact would have reported, which is a leniency difference rather than a family effect.

Judge	Global leniency	Diagonal residual	SD units
Claude Opus 4.8	$-0.016$	$+0.012$	$+0.76$
Llama 4 Maverick	$+0.075$	$-0.009$	$-0.56$
Gemini 3.1 Pro	$-0.059$	$-0.004$	$-0.26$

Appendix

Table 7: Judge leniency versus own-family residual. Residual standard deviation is

0.016

; the largest off-diagonal residual is

0.058

Axis	Value	Threads
Subcategory	Psychosis	10
	Cognitive Patterns	4
	Behavioral Drift	2
	Escalation Pattern	2
	Masked Distress	2
Age	O18	16

Appendix

Table 8: DistressBench-LH composition (

n=20

threads).

Category	Analytic role
Sycophantic engagement in the face of crisis	Validation without intervention
Fostering dangerous emotional dependency and isolation	Rapport-capture failure
Reinforcing delusions and psychosis	Signal mishandling / severe unsafe
Facilitating or instructing harm	Severe unsafe behavior
Romanticizing or justifying suicide & self-harm	Severe unsafe behavior

Appendix

Table 9: DistressBench-LH violation taxonomy.

Figure 5: Violation categories across the 20 DistressBench-LH threads, on a single frozen reference path.

UID	Subcategory	Violation categories
lh_mt_6_8-1	Behavioral Drift	None (control)
lh_mt_6_8-2	Cognitive Patterns	All five categories
lh_mt_6_8-3	Cognitive Patterns	Sycophantic engagement
lh_mt_6_8-4	Escalation Pattern	None (control)
lh_mt_6_8-19	Masked Distress	Facilitating harm; sycophantic engagement
…	…	…

Appendix

Table 10: DistressBench-LH task inventory (excerpt; full list ships with the release).

Band	Model	Score	95% CI	All met
1	Muse Spark 1.2	0.884	[0.875, 0.893]	40.4%
2	Muse Spark 1.3	0.849	[0.839, 0.858]	18.3%
3	Muse Glimmer 30B	0.806	[0.796, 0.815]	11.3%
4	Inkling	0.785	[0.774, 0.795]	4.6%
5	Qwen3.8 Max	0.776	[0.766, 0.787]	4.5%
6	GPT-5.6 Terra	0.764	[0.753, 0.775]	3.8%

Appendix

Table 11: One model separates from the field, and complete rubric satisfaction is rare everywhere else. Weighted rubric score over three attempts,

n=718

tasks, 95% bootstrap CI. Bands are separated by paired-bootstrap tests on adjacent pairs, and models within a band are not statistically distinguishable. All met is the share of conversations in which every rubric criterion is satisfied.

Figure 6: No subcategory is handled well by every model, and the weakest are non-suicidal self-injury and trauma-triggered distress. Subcategory pass rates for the top eight models.

Figure 7: Wide median-to-best gaps identify behaviors that are demonstrably achievable and commonly omitted. Median against best model by dimension, over 25 models.

Figure 8: Clinician references nearly saturate the rubric, so the residual model shortfall is not an artifact of impossible criteria. Best model score against the clinician reference ceiling on the same judge pipeline.

Figure 9: Multi-turn conversations score lower for the large majority of models. Single-turn against multi-turn weighted rubric scores, with per-model points.

Subcategory	Blank	Attempts	Rate
Escalation Pattern	39	114	34.2%
Trauma-Triggered Distress	18	87	20.7%
Uncontrolled Anger	17	111	15.3%
Psychosis	14	108	13.0%
Acute Emotional Psychological Distress	8	66	12.1%
Cultural/Spiritual Crisis	13	114	11.4%

Appendix

Table 12: The filter fires on agitation and is nearly silent on the subcategories carrying the most direct lethal risk. Claude Fable 5 empty-response incidence by SSH subcategory, three attempts pooled. Subcategories with no blanks: non-suicidal self-injury, eating disorders, masked distress, self worth, chronic distress, and existential despair.

Dimension	$n$	Production	Implicit	Divergence	Limit
Disclaimers	1,219	0.005	0.043	0.885	0.067
De-escalation	1,062	0.073	0.019	0.886	0.023
Actionable resources	642	0.091	0.016	0.851	0.042
Identification	336	0.059	0.114	0.792	0.035
Compassionate engagement	238	0.120	0.051	0.773	0.056
All	3,687	0.071	0.039	0.850	0.041

Appendix

Table 13: Standard divergence accounts for 85% of forfeited criterion weight, and discrimination failure for 4%. Share of forfeited criterion weight by mechanism, 21 models, 3,687 criterion units. Llama 4 Maverick is excluded as an uninformative evaluator. Production gap : prefers the compliant reply and names the missing behavior unprompted. Implicit preference : prefers it without naming the behavior. Standard divergence : applies the criterion correctly when told, does not prefer the reply that satisfies it. Discrimination limit : cannot identify the compliant reply.

Explore similar work

EMPATH: A Multilingual Auditor-Judge Benchmark for Safety Evaluation of Emotional-Support Chatbots
Jun 29, 2026Camilo Chacón SartoriLLM Safety Benchmarks Multilingual Language Model Evaluation
MindSurf
K-Bench: a clinically calibrated benchmark for evaluating large language models in high-risk mental health conversations
Sep 14, 2026Laura M. Vowels, Matthew J. Vowels, Shivali Sharma +9LLM Safety Benchmarks Mental Health Counseling
School of Psychology, University of Roehampton, London, United Kingdom · Kivira Health, United Kingdom · University of Hertfordshire, Hatfield, United Kingdom +4
Expert-Level Crisis Detection in Mental Health Conversations
Jun 9, 2026Grace Byun, Abigail Lott, Rebecca Lipschutz +3Mental Health Suicide Risk Modeling
Department of Computer Science, Emory University · Department of Psychiatry and Behavioral Sciences, Emory University

Jun 29, 2026J/K move · Enter open · S save

EMPATH: A Multilingual Auditor-Judge Benchmark for Safety Evaluation of Emotional-Support Chatbots

Camilo Chacón Sartori

MindSurf

LLM Safety Benchmarks Multilingual Language Model Evaluation LLM-as-a-Judge Multi-Turn Dialogue Evaluation LLM Safety Evaluation Language Model Auditing

Open paper AskPDF arXiv

Safety benchmarks often buy scalability by fixing the prompt, the language, and the turn structure. For emotional-support chatbots, that bargain hides precisely where safety failures emerge: across a multilingual, multi-turn crisis conversation. We present EMPATH, a benchmark for safety evaluation of emotional-support chatbots. An auditor model role-plays help-seeking users, generating multi-turn conversations from 140 seed instructions and 34 personas. A judge model scores each full transcript against 19 metrics across five dimensions: crisis handling, therapeutic quality, conversational integrity, emotional safety, and cultural adaptation. EMPATH is built for Mexican Spanish and US English; the studies reported here run in Mexican Spanish. Auditor and judge are drawn from different model families, and the judge is treated as an instrument to be calibrated rather than trusted. A strict per-criterion rubric reveals material score inflation on 10 of the 19 metrics and restores discrimination. We study the measurement properties of the benchmark through judge calibration and cross-family inter-judge agreement. We also illustrate EMPATH on three frontier models, one of them open-weight. Aggregate scores sit within 0.74 points of one another, but per-metric profiles diverge by up to six points in model-specific places. Under the standard rubric, both the ranking and the weak spots are stable across a second, cross-family judge: 93% of scores fall within plus or minus 1. A five-run test-retest adds a second axis: even the steadiest model swings from 2 to 10 on a crisis metric across identical re-runs, and deepseek-v4-pro returns a different conversation on every run even at temperature 0. Run-to-run reliability is therefore a per-model safety property, not noise to average away. EMPATH is system-agnostic; the pipeline, seeds, personas, and rubrics are released for reuse.

Abstract

Explore similar work

EMPATH: A Multilingual Auditor-Judge Benchmark for Safety Evaluation of Emotional-Support Chatbots

K-Bench: a clinically calibrated benchmark for evaluating large language models in high-risk mental health conversations

Expert-Level Crisis Detection in Mental Health Conversations