Mental Health Equity in LLMs: Leveraging Multi-Hop Question Answering to Detect Amplified and Silenced Perspectives

Aman Chadha; Atmika Gorti; Batool Haider; Manas Gaur

REVIEW 5 major objections 5 minor 34 references

Mental Health Equity in LLMs: Leveraging Multi-Hop Question Answering to Detect Amplified and Silenced Perspectives

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A multi-hop question-answering framework exposes where LLMs amplify mental-health bias at demographic intersections, and few-shot debiasing cuts measured bias by 66–94%.

desk verdict A promising intersectional bias audit undermined by an undefined bias score and unsupported headline claims. read the letter →

arxiv 2506.18116 v1 pith:OVXEDT22 submitted 2025-06-22 cs.CL cs.AIcs.CY

Batool Haider , Atmika Gorti , Aman Chadha , Manas Gaur This is my paper · ORCID

classification cs.CLcs.AIcs.CY

keywords multi-hopquestionansweringintersectionalbiasmentalhealthLLMsdebiasingfew-shotpromptingBBQdatasetamplificationLLMevaluation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that standard single-axis bias checks miss the way language models mistreat people whose identities intersect, and that a multi-hop question-answering setup can expose those patterns. It builds on the IMHI mental-health dataset, tags posts for age, gender, race, socioeconomic status, and condition, then asks each model to reason across three tagged posts to answer questions such as how being a young low-income person with depression affects symptoms, coping, and relationships. Across Claude 3.5 Sonnet, Jamba 1.6, Gemma-3, and Llama-4, the framework finds systematic disparities and identifies 'amplification points' where early steps of reasoning magnify bias. Few-shot prompting with pre-classified BBQ examples cuts measured bias by 66–94%, while roleplay and explicit-debiasing prompts give mixed, sometimes counterproductive results.

What carries the argument

The machinery is the MHQA prompt template: three tagged source posts (real Reddit posts from IMHI, supplemented by Claude-generated posts when fewer than three exist for a demographic-condition pair) plus a question of the form 'How does being [demographic] positively/negatively affect the symptoms, coping mechanisms, and interpersonal relationships of a person with [condition]?' The tagging system (zero-shot and few-shot with BBQ examples) assigns age, gender, race, SES, and condition tags; the bias score measures sentiment/tone, demographic, and mental-health-condition dimensions of the model's answer. The 'amplification point' is the paper's term for the observation that bias grows as the model chains evidence across multiple sources and reasoning steps.

What would settle it

Re-run the MHQA evaluation with the synthetic evidence posts replaced by real held-out posts from the same demographic groups and conditions. If race and SES bias scores drop materially, the reported intersectional bias for those categories was an artifact of Claude 3.5 Sonnet's generated sources rather than a property of the four evaluated models. Alternatively, have independent human raters score the same model outputs for bias and compare with the paper's automatic scores; low agreement would falsify the claim that the framework detects bias rather than imposes a scoring convention.

Watch

Extended reading notes

Core claim

The central claim is that multi-hop question answering is a superior detector of subtle, intersectional bias in mental-health LLM responses. Concretely: take a Reddit post tagged with demographics and a mental-health condition, generate a question asking how that demographic factor positively or negatively affects symptoms, coping, and interpersonal relationships of someone with the condition, and supply three tagged posts as evidence that the model must corroborate. Score the generated answer for bias in three dimensions—sentiment/tone, demographic fairness, and mental-health-condition framing. The authors report that this procedure surfaces disparities that conventional single-factor checking misses, that bias can compound across reasoning hops, and that injecting BBQ few-shot examples into the prompt reduces measured bias by 66–94% depending on model and category, whereas direct debiasing instructions sometimes increase bias (Jamba's mental-health-condition bias rose from 0.344 to 0.520 under explicit prompting).

Load-bearing premise

The pipeline assumes that Claude 3.5 Sonnet tags demographic factors and mental-health conditions reliably, and that the artificial posts it writes for under-represented groups stand in for real user voices; if either fails, the bias scores for race, SES, and age reflect the generator, not the four models being evaluated.

Editorial extensions

If this is right

Mental-health chatbots should be evaluated with intersectional, multi-source prompts before deployment; single-factor audits are insufficient.
Cheap few-shot debiasing with BBQ-style examples can cut measured bias by 66–94%, so prompt-level mitigation is a viable first-line intervention.
Debiasing is model-specific: explicit instructions backfire on some models (Jamba's mental-health bias rose from 0.344 to 0.520), so blanket fairness prompts should not be assumed safe.
No evaluated model dominates on all bias axes, suggesting that equitable deployment may require model selection or ensembling per use case.
The persistence of measurable bias even after the best interventions implies that prompt fixes alone will not make LLMs safe for clinical mental-health use without oversight.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because 176 of the synthetic evidence posts were generated for the race category (versus 4 for gender), the race and SES bias scores may partly measure Claude 3.5 Sonnet's own stereotypes rather than the four evaluated models; swapping in real held-out posts for under-represented groups would isolate this.
The amplification-point claim implies a monotonic relationship between number of reasoning hops and measured bias; varying hop count while holding question content fixed would directly test it.
The reported 66–94% reductions use few-shot examples drawn from the same BBQ benchmark that defines the bias categories; an out-of-distribution test with novel demographic intersections would show how much of the gain transfers.
The framework's design mimics retrieval-augmented generation, so the method could be reused as a red-team audit for RAG-based mental-health assistants by feeding them adversarially selected source posts.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague, here is my read on arXiv:2506.18116.

The application of multi-hop QA to intersectional mental-health bias is genuinely new, and the pipeline of tagging Reddit posts and generating artificial posts to fill demographic gaps is not in the prior literature I've seen. The paper ships code and data links, which is credit-worthy, and the comparison of four models under two debiasing strategies is the kind of empirical audit that could be useful to practitioners. The ethical statement is thoughtful.

The problem is the central measurement. The bias score in Tables 2-4 is never defined: no scoring function, no aggregation over models, prompts, tags, or conditions, no range. Without that definition, every number in the results section is unauditable. This is not a minor omission; it is the whole evidential base. The abstract's claim of 66-94% reductions is also not supported by Table 2, where Gemma-3's sentiment/tone drops 0.771 to 0.341 (56%). The 'superior detection' claim floats without any baseline experiment against single-hop or direct prompting.

There is also a real circularity burden. Claude 3.5 Sonnet generates the demographic tags and the artificial posts, and then is one of the four evaluated models. For race, 176 of 306 artificial posts come from Claude (Figure 8). The paper explicitly acknowledges this skew, which is good, but the reported scores are not stratified by real versus artificial sources, so the influence cannot be assessed.

The reader's REJECT verdict is correct for the current manuscript. But I don't want to throw out the baby with the bathwater. The idea is worth pursuing, and the flaws are fixable in principle. A revision that defines the bias score, adds a single-hop baseline, stratifies results by source, and corrects the abstract's range would be a materially better paper.

I'd send this to a serious referee rather than desk reject, because the research direction matters and the pipeline is novel enough to deserve expert scrutiny. I would not cite it in its current form, but I'd keep an eye on a revised version.

Referee Report

5 major / 5 minor

Summary. The paper introduces a multi-hop question answering (MHQA) framework to detect intersectional bias in LLM mental-health responses. The authors tag IMHI/Dreaddit/MultiWD posts for age, gender, race, socioeconomic status, and mental-health condition using Claude 3.5 Sonnet; generate artificial source posts when fewer than three tagged posts exist; and prompt four LLMs (Claude 3.5 Sonnet, Jamba 1.6, Gemma-3, Llama-4) with positive/negative-effect questions. Bias scores are reported for zero-shot and few-shot settings in three categories (sentiment/tone, demographic, mental-health condition), and three debiasing interventions are compared: BBQ few-shot prompting, roleplay simulation, and explicit bias reduction. The abstract claims 66–94% bias reductions and superior detection of amplification points.

Significance. The paper addresses a socially important problem—bias in mental-health LLMs—and its core idea, probing intersectional identities through multi-hop QA, is worth exploring. I credit the authors for releasing code and modified data, evaluating multiple models, comparing several debiasing strategies, and including a thoughtful ethical statement. However, in its current form the empirical contribution cannot be audited: the bias score used in Tables 2–4 is never defined, the abstract's reduction range is inconsistent with the reported numbers, the 'superior detection' claim has no baseline or external validation, and the pipeline uses Claude 3.5 Sonnet both to create the evaluation data and as one of the evaluated models. If these issues are resolved, the framework could become a useful contribution to bias elicitation; as submitted, the central claims are unsupported.

major comments (5)

[Section 3, Tables 2–4] The central dependent variable is never defined. The manuscript reports scores such as 0.771 and 0.341 but does not state the scoring function, the range or units, the aggregation over posts, models, demographic tags, mental-health conditions, or question variants, or the measurement procedure (e.g., a classifier, an LLM judge, or manual annotation). Because every quantitative conclusion in the Results depends on this score, Tables 2–4 cannot be checked or reproduced. Please provide the exact formula, thresholds, and aggregation details, and include the scoring scripts in the released artifacts.
[Abstract and Section 3.2] The abstract's claim of '66–94% bias reductions' is contradicted by Table 2 and the surrounding text. In Table 2, Gemma-3's sentiment/tone score falls from 0.771 to 0.341 (56% reduction), its demographic score from 0.390 to 0.290 (26% reduction), and its mental-health-condition score from 0.502 to 0.338 (33% reduction). The text itself says 'Gemma-3 showed the strongest relative improvement ... decreasing from 0.771 to 0.341 (56%)' while also claiming reductions of 66–94%. The 94% figure appears only for Llama-4's demographic score in Table 4. Please recompute and correct the claimed range, and specify which table and prompt setting supports each percentage.
[Section 3, Introduction, Abstract] The claim that MHQA 'demonstrates superior detection of subtle bias patterns compared to conventional methods' is unsupported by any comparison. There is no single-hop or direct-prompt baseline, no alternative bias-elicitation method, and no human or external validation of 'amplified' versus 'silenced' perspectives. The term 'amplification point' is used repeatedly (Abstract, Section 3, Conclusion) but is never operationally defined or measured. Please add a quantitative baseline comparison and define the amplification metric before claiming superiority.
[Sections 2.3, 2.7, Figure 8] The evaluation pipeline is circular for Claude 3.5 Sonnet: the same model generates the demographic and mental-health tags used to build questions (Section 2.3), generates 176 of 306 artificial source posts for the race category (Section 2.7, Figure 8), and is also one of the four models scored in Tables 2–4. The paper acknowledges that categories with more artificial posts 'may skew results' but does not stratify any reported score by real versus artificial evidence. As a result, race, SES, and age scores may measure the content generator's priors rather than the target model's bias. Please report bias scores separately for real-only and artificial-only source sets, or re-run the evaluation with a different generator and exclude that generator from the model set.
[Section 3.3, Tables 3–4] The debiasing conditions are under-specified. The tables use 'Zero-Shot' and 'Few-Shot' labels, but the text does not provide the roleplay simulation prompts, the explicit bias reduction prompts, the number of BBQ examples used in each condition, or how these conditions differ from the few-shot condition in Table 2. Without these details, the reductions in Tables 3–4 cannot be interpreted or reproduced. Please include the full prompt templates and an ablation describing which components of each intervention are active.

minor comments (5)

[Section 3.1] The heading 'Elicitated Bias Patterns Across Models' contains a typo ('Elicitated' should be 'Elicited'), and Section 3.2 uses the non-standard phrase 'demographic bias being under-control'.
[Section 2.3, Figures 3–4] The counts in the text (e.g., White 13, Hispanic/Latino 11, Black/African American 4, American Indian 4) should be reconciled with Figure 3, whose bar heights and abbreviations are hard to read; please add a table or numerical labels.
[Section 3.4] Specific qualitative findings, such as 'responses to low-income young adults with depression were more negatively framed' and the 'Black, female, low-income, and depressed' example, are not tied to any table, figure, or statistical test; please add supporting evidence.
[Tables 2–4] Model names are inconsistent across tables ('Jamba 1.6' in Table 2 vs 'Jamba' in Table 3; 'Claude Sonnet' vs 'Claude 3.5 Sonnet'), and the Jamba few-shot mental-health value of 0.88 in Table 4 is not discussed; please verify this value and align naming.
[References] Reference [18] (Liu et al., MintQA) does not appear to support the citation about bias surfacing in active conversations, and references [32] and [35] share the same arXiv identifier (2406.12033) with different titles; these should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the paper's central claims are not shown to be reducible to their own inputs, though undefined metrics and Claude-generated evidence create serious validity risks.

full rationale

After walking the derivation chain, no step exhibits the kind of equation-level or statistical equivalence that defines circularity. The bias scores in Tables 2-4 are never defined, so the reported 66-94% reductions cannot be audited; however, an undefined metric is a reproducibility and correctness defect, not a circularity. The MHQA 'superior detection' claim lacks a conventional-method baseline, which weakens external validity but again is not a circular reduction. The closest concern is the closed-loop use of Claude 3.5 Sonnet: Section 2.3 uses Claude to tag demographic and mental-health attributes, Section 2.7 uses Claude to generate artificial source posts when fewer than three exist, and Section 2.4/Table 2 evaluates Claude as one of the four models. The paper itself flags this in Section 2.7: 'demographic categories with more artificial posts as sources may skew results in terms of bias evaluation.' This is a genuine confound, especially for race with 176 of 306 artificial posts, but the manuscript does not define the bias score, so there is no exhibited reduction showing that a reported score is forced to equal the tagged/generated input by construction. Debiasing via BBQ few-shot examples is a within-prompt comparison rather than a fitted parameter renamed as a prediction. No load-bearing self-citation or imported uniqueness theorem appears. Consequently, under the hard rule that circularity must be demonstrated by quote and specific reduction, the finding is no significant circularity; the serious methodological concerns belong to correctness and reproducibility risk rather than circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claims depend on LLM-generated labels, LLM-generated source posts, and an undefined bias metric; none of these is independently validated.

free parameters (4)

Number of source posts per MHQA question = 3
Set by hand in Section 2.7 to "diversify opinions"; the amount of context directly influences model answers and the resulting bias scores.
Maximum response length = 120 words
Set in Section 2.1 and repeated in the MHQA template; word limits can change sentiment and content, hence the measured bias.
Bias score formula and thresholds = not disclosed
Tables 2-4 depend on this undisclosed aggregation over sentiment, demographic, and mental-health dimensions; any classifier choice, threshold, or normalization is a free parameter.
Number of BBQ few-shot examples = 16 (implied by Figure 7)
The few-shot demonstrations are selected by hand and embedded at the start of prompts; their count and content control the debiasing effect.

assumptions (5)

ad hoc to paper Claude 3.5 Sonnet's demographic and mental-health tags are accurate enough to build bias-probing questions.
Section 2.3 uses Claude for all tagging with no human validation or agreement metric; any tag error propagates to the questions and scores.
ad hoc to paper Artificial posts generated by Claude 3.5 Sonnet faithfully represent the demographic and mental-health conditions they replace.
Section 2.7 creates synthetic posts for scarce race, SES, and age groups; if these posts differ in tone or content, the bias scores for those groups are confounded.
ad hoc to paper The undisclosed bias score is a valid quantitative measure of bias.
Tables 2-4 rest entirely on this score; no formula or rubric is given in Section 3.
domain assumption Three posts per question provide sufficient context to simulate internet-like retrieval and reveal bias.
Section 2.7 sets three sources without an ablation; the number of sources may determine whether bias is detected or hidden.
domain assumption BBQ few-shot examples act as implicit debiasing guidance without direct instruction.
Section 2.5 embeds BBQ examples at the start of prompts; whether this reduces bias or merely changes response style is not validated.

invented entities (1)

Amplification point
purpose: Explains where small biases in early multi-hop reasoning steps become magnified in final LLM outputs.
Named in the abstract and Section 1 as a key outcome, but never defined, located, or measured; no algorithm or example in the paper detects one.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Mental Health Equity in LLMs: Leveraging Multi-Hop Question Answering to Detect Amplified and Silenced Perspectives." pith.science (2026). https://pith.science/paper/OVXEDT22

@misc{pith2026250618116,
  author       = {Pith},
  title        = {Pith review of: Mental Health Equity in LLMs: Leveraging Multi-Hop Question Answering to Detect Amplified and Silenced Perspectives},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OVXEDT22}},
  note         = {Machine review of arXiv:2506.18116}
}

read the original abstract

Large Language Models (LLMs) in mental healthcare risk propagating biases that reinforce stigma and harm marginalized groups. While previous research identified concerning trends, systematic methods for detecting intersectional biases remain limited. This work introduces a multi-hop question answering (MHQA) framework to explore LLM response biases in mental health discourse. We analyze content from the Interpretable Mental Health Instruction (IMHI) dataset across symptom presentation, coping mechanisms, and treatment approaches. Using systematic tagging across age, race, gender, and socioeconomic status, we investigate bias patterns at demographic intersections. We evaluate four LLMs: Claude 3.5 Sonnet, Jamba 1.6, Gemma 3, and Llama 4, revealing systematic disparities across sentiment, demographics, and mental health conditions. Our MHQA approach demonstrates superior detection compared to conventional methods, identifying amplification points where biases magnify through sequential reasoning. We implement two debiasing techniques: Roleplay Simulation and Explicit Bias Reduction, achieving 66-94% bias reductions through few-shot prompting with BBQ dataset examples. These findings highlight critical areas where LLMs reproduce mental healthcare biases, providing actionable insights for equitable AI development.

Figures

Figures reproduced from arXiv: 2506.18116 by the authors.

**Figure 2.** The dataset instances were systematically tagged across four key dimen [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Dataset demographics categorized through large language model tagging, illustrating variations in age, gender, race, and income representation across the analyzed posts. Abbreviations used: Age categories - YA: Young Adult, SR: Senior, CH: Child, AD: Adult; Gender categories - M: Male, F: Female; Race categories - W: White, H/L: Hispanic/Latino, B/A: Black/African, A: Asian, AI/I: American Indian/Indigenous; Income… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Mental health conditions tag frequency. SAX: Social Anxiety, OCD: Obsessive Compulsive Disorder, EDO: Eating Disorder, DEP: Depression, BDO: Bipolar Disorder, ANX: Anxiety, ADD: Addiction. Gemma 3 (27B) includes builtin safety classifiers designed to detect harmf…

**Figure 5.** Figure 5: Positive and negative effect of demographic factors on LLM’s generated [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: MHQA template In order to diversify opinions, we attach three sources per question. However, as seen through the tagging frequencies per demographic, not all demographic factors are represented equally according to the mental health condition. Due to this limitation, …

**Figure 8.** Figure 8: SES: Socioeconomic Status. Artificial posts are defined as posts that imitate posts from the MultiWD and DR datasets to fulfill the 3 post requirement in the MHQA prompt [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

**Figure 7.** Figure 7: Examples of MHQA Template: Zero-shot and Few Shot Template for [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 29 canonical work pages

[1]

medRxiv (2024).https: //doi.org/10.1101/2024.10.29.24316368, https://www.medrxiv.org/ content/10.1101/2024.10.29.24316368v1.full, preprint

Anonymous: Socio-demographic biases in medical decision-making by large language models: A large-scale multi-model analysis. medRxiv (2024).https: //doi.org/10.1101/2024.10.29.24316368, https://www.medrxiv.org/ content/10.1101/2024.10.29.24316368v1.full, preprint

work page doi:10.1101/2024.10.29.24316368 2024
[2]

npj Digital Medicine (2025),https://www.nature.com/articles/s41746-025-01746-4

Anonymous: Racial bias in ai-mediated psychiatric diagnosis and treatment: a qualitative comparison of four large language models. npj Digital Medicine (2025),https://www.nature.com/articles/s41746-025-01746-4

work page 2025
[3]

Baskar, S., Gaur, M., Parthasarathy, S., Verlekar, T.T.: (cper) from guessing to asking: An approach to resolving persona knowledge gap in llms during multi-turn conversations. In: Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 4: Student Resea...

work page 2025
[4]

for the Protection of Human Subjects of Biomedical, U.S.N.C., Research, B.: The Belmont report: ethical principles and guidelines for the protection of human subjects of research, vol. 2. Department of Health, Education, and Welfare, National Commission for the ... (1978)

work page 1978
[5]

In: First Conference on Language Modeling

Cheng, J., Marone, M., Weller, O., Lawrie, D., Khashabi, D., Van Durme, B.: Dated data: Tracing knowledge cutoffs in large language models. In: First Conference on Language Modeling

work page
[6]

Deb, T.: AI Mental Health Market Forecast To Grow At a 32.1

work page
[7]

https://deepmind.google/technologies/gemma/ (2025), official model announcement with fine-tuning and safety details

DeepMind, G.: Gemma 3: Google’s new open model for safer and smarter ai. https://deepmind.google/technologies/gemma/ (2025), official model announcement with fine-tuning and safety details

work page 2025
[8]

Minds and machines28, 689–707 (2018)

Floridi, L., Cowls, J., Beltrametti, M., Chatila, R., Chazerand, P., Dignum, V., Luetge, C., Madelin, R., Pagallo, U., Rossi, F., et al.: Ai4people—an ethical framework for a good ai society: opportunities, risks, principles, and recommendations. Minds and machines28, 689–707 (2018)

work page 2018

Show all 34 references

[9]

Computational Lin- guistics50(3), 1097–1179 (2024)

Gallegos, I.O., Rossi, R.A., Barrow, J., Tanjim, M.M., Kim, S., Dernoncourt, F., Yu, T., Zhang, R., Ahmed, N.K.: Bias and fairness in large language models: A survey. Computational Lin- guistics50(3), 1097–1179 (2024). https://doi.org/10.1162/coli_ a_00524, https://direct.mit....

2024 doi
[10]

IEEE Internet Computing26(4), 5–11 (2022)

Gaur, M., Gunaratna, K., Bhatt, S., Sheth, A.: Knowledge-infused learning: A sweet spot in neuro-symbolic ai. IEEE Internet Computing26(4), 5–11 (2022)

2022
[11]

AI Magazine45(1), 139–155 (2024)

Gaur, M., Sheth, A.: Building trustworthy neurosymbolic ai systems: Con- sistency, reliability, explainability, and safety. AI Magazine45(1), 139–155 (2024)

2024
[12]

in armed conflict

Henckaerts, J.M.: Study on customary international humanitarian law: A contribution to the understanding and respect for the rule of law 18 Haider et al. in armed conflict. International Review of the Red Cross87(857), 175–186 (2005), https://www.onlinelibrary.iihl.org/wp-cont...

2005
[13]

Hsieh, C.P., Sun, S., Kriman, S., Acharya, S., Rekesh, D., Jia, F., Zhang, Y., Ginsburg, B.: Ruler: What’s the real context size of your long-context language models? arXiv preprint arXiv:2404.06654 (2024)

2024 arXiv
[14]

Kumar, S.H., et al.: Decoding biases: Automated methods and llm judges for gender bias detection in language models (2024),https://arxiv.org/ abs/2408.03907

2024 arXiv
[15]

ai21.com/blog/jamba-1-5-long-context-low-latency-open-source (2024), blog post discussing Jamba 1.5’s 256K token performance

Labs, A.: Jamba 1.5: Long context, low latency, open source.https://www. ai21.com/blog/jamba-1-5-long-context-low-latency-open-source (2024), blog post discussing Jamba 1.5’s 256K token performance

2024
[16]

https://doi.org/10.2196/ 59479,https://mental.jmir.org/2024/1/e59479

Lawrence, H.R., Schneider, R.A., Rubin, S.B., Matarić, M.J., McDuff, D.J., Bell, M.J.: The opportunities and risks of large language models in mental health.JMIRMentalHealth11,e59479(2024). https://doi.org/10.2196/ 59479,https://mental.jmir.org/2024/1/e59479

2024
[17]

Proceedings of the 2022 ACM Conference on Health, Inference, and Learning (CHIL) (2022)

Lin, H., Waseem, Z., De Choudhury, M.: Mental health insights from large language models: Bias, performance, and equity challenges. Proceedings of the 2022 ACM Conference on Health, Inference, and Learning (CHIL) (2022)

2022
[18]

Papers with Code (2024),https://paperswithcode.com/task/ multi-hop-question-answering/latest, accessed June 2025

Liu, Y., et al.: Mintqa: Multi-hop question answering on new and tail knowledge. Papers with Code (2024),https://paperswithcode.com/task/ multi-hop-question-answering/latest, accessed June 2025

2024
[19]

CHI Conference on Human Factors in Computing Systems Extended Abstracts (2021)

Magee, R., Golebiewski, M., Burke, M.: Intersectionality in ai bias research: A mental health use case. CHI Conference on Human Factors in Computing Systems Extended Abstracts (2021)

2021
[20]

arXiv preprint arXiv:2504.06160 (2025)

Magu, R., Kim, S., De Choudhury, M.: Navigating the rabbit hole: Emergent biases in llm-generated narratives about mental health. arXiv preprint arXiv:2504.06160 (2025)

2025
[21]

arXiv preprint arXiv:2306.05550 (2023)

Mei, K.X., Fereidooni, S., Caliskan, A.: Bias against 93 stigmatized groups in masked language models and downstream sentiment classification tasks. arXiv preprint arXiv:2306.05550 (2023)

2023 arXiv
[22]

In: Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP

Mohammadi, S., Raff, E., Malekar, J., Palit, V., Ferraro, F., Gaur, M.: Welldunn: On the robustness and explainability of language models and large language models in identifying wellness dimensions. In: Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting N...

2024
[23]

In: Findings of the Association for Computational Linguistics: ACL 2022

Parrish, A., Chen, A., Nangia, N., Padmakumar, V., Phang, J., Thompson, J., Htut, P.M., Bowman, S.: Bbq: A hand-built bias benchmark for question answering. In: Findings of the Association for Computational Linguistics: ACL 2022. pp. 2086–2105 (2022)

2022
[24]

arXiv preprint arXiv:2402.08721 (2024)

Qian, S., Lee, K., Sharma, T.: Debiasing large language models through reasoning strategies: Promise and pitfalls. arXiv preprint arXiv:2402.08721 (2024)

2024 arXiv
[25]

First Monday (2023) Bias Elicitation and Debiasing in Mental Health 19

Reagle, J.: Even pseudonyms and throwaways delete their reddit posts. First Monday (2023) Bias Elicitation and Debiasing in Mental Health 19

2023
[26]

Reagle, J., Gaur, M.: Spinning words as disguise: Shady services for ethical research? First Monday (2022)

2022
[27]

Research, M.A.: Llama 4 technical report.https://ai.meta.com/blog/ llama-4-open-foundation-models/ (2025), covers LLaMA 4’s expanded 10M context length and bias mitigation

2025
[28]

Frontiers in Artificial Intelligence6, 1229805 (2023)

Sarkar, S., Gaur, M., Chen, L.K., Garg, M., Srivastava, B.: A review of the explainability and safety of conversational agents for mental health to identify avenues for improvement. Frontiers in Artificial Intelligence6, 1229805 (2023)

2023
[29]

ACM Conference on Fairness, Accountability, and Transparency (FAccT) (2021)

Subramanian, A., Roberts, L., Singh, R.: Evaluating fairness metrics in mental health ai: An intersectional perspective. ACM Conference on Fairness, Accountability, and Transparency (FAccT) (2021)

2021
[30]

Journal of Medical Internet Research25(3), e45127 (2023)

Timmons, N., Choudhury, M.D., Kim, J.Y.: A call to action for fair and equitable mental health ai: Intersectionality, representation, and ethics. Journal of Medical Internet Research25(3), e45127 (2023)

2023
[31]

arXiv preprint arXiv:1911.00133 (2019)

Turcan, E., McKeown, K.: Dreaddit: A reddit dataset for stress analysis in social media. arXiv preprint arXiv:1911.00133 (2019)

2019 arXiv
[33]

In: Proceedings of the ACM Web Conference 2024

Yang, K., Zhang, T., Kuang, Z., Xie, Q., Huang, J., Ananiadou, S.: Men- tallama: Interpretable mental health analysis on social media with large language models. In: Proceedings of the ACM Web Conference 2024. p. 4489–4500. WWW ’24, ACM (May 2024). https://doi.org/10.1145/ 358...

2024
[34]

arXiv preprint arXiv:2405.01432 (2024)

Zeng, W., Kurniawan, D., Mullins, R., et al.: Shieldgemma: Safety-tuned language models for sensitive applications. arXiv preprint arXiv:2405.01432 (2024)

2024 arXiv
[35]

arXiv preprint arXiv:2406.12033 (2024)

Zhao, Y., Wang, Y., Keller, S.A., Hernandez-Boussard, T.: Role-play paradox in large language models: Reasoning and bias implications. arXiv preprint arXiv:2406.12033 (2024)

2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

medRxiv (2024).https: //doi.org/10.1101/2024.10.29.24316368, https://www.medrxiv.org/ content/10.1101/2024.10.29.24316368v1.full, preprint

Anonymous: Socio-demographic biases in medical decision-making by large language models: A large-scale multi-model analysis. medRxiv (2024).https: //doi.org/10.1101/2024.10.29.24316368, https://www.medrxiv.org/ content/10.1101/2024.10.29.24316368v1.full, preprint

work page doi:10.1101/2024.10.29.24316368 2024

[2] [2]

npj Digital Medicine (2025),https://www.nature.com/articles/s41746-025-01746-4

Anonymous: Racial bias in ai-mediated psychiatric diagnosis and treatment: a qualitative comparison of four large language models. npj Digital Medicine (2025),https://www.nature.com/articles/s41746-025-01746-4

work page 2025

[3] [3]

Baskar, S., Gaur, M., Parthasarathy, S., Verlekar, T.T.: (cper) from guessing to asking: An approach to resolving persona knowledge gap in llms during multi-turn conversations. In: Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 4: Student Resea...

work page 2025

[4] [4]

for the Protection of Human Subjects of Biomedical, U.S.N.C., Research, B.: The Belmont report: ethical principles and guidelines for the protection of human subjects of research, vol. 2. Department of Health, Education, and Welfare, National Commission for the ... (1978)

work page 1978

[5] [5]

In: First Conference on Language Modeling

Cheng, J., Marone, M., Weller, O., Lawrie, D., Khashabi, D., Van Durme, B.: Dated data: Tracing knowledge cutoffs in large language models. In: First Conference on Language Modeling

work page

[6] [6]

Deb, T.: AI Mental Health Market Forecast To Grow At a 32.1

work page

[7] [7]

https://deepmind.google/technologies/gemma/ (2025), official model announcement with fine-tuning and safety details

DeepMind, G.: Gemma 3: Google’s new open model for safer and smarter ai. https://deepmind.google/technologies/gemma/ (2025), official model announcement with fine-tuning and safety details

work page 2025

[8] [8]

Minds and machines28, 689–707 (2018)

Floridi, L., Cowls, J., Beltrametti, M., Chatila, R., Chazerand, P., Dignum, V., Luetge, C., Madelin, R., Pagallo, U., Rossi, F., et al.: Ai4people—an ethical framework for a good ai society: opportunities, risks, principles, and recommendations. Minds and machines28, 689–707 (2018)

work page 2018