SDARE-Bench: Evaluating Large Language Models on Conversational Stigma Detection and Response in Dyadic and Group Dialogue

AI summary

Overview

Research area: Natural Language Processing / large language model safety evaluation, specifically stigma detection and stigma response in conversational settings, drawing on psychological stigma theory.

Technical level: Intermediate.

Scope: This paper introduces SDARE-Bench, a benchmark of 1,138 dyadic queries and 1,388 group dialogues that tests whether eight large language models can detect stigma and its components and respond appropriately to stigma in both one-to-one and four-speaker settings.

What This Paper Is About

Existing LLM safety benchmarks mostly test overtly harmful content or use static prompts with fixed answer formats such as multiple choice, and they almost always model one-to-one exchanges. The authors argue that stigma is often polite, indirect, and free of profanity, so it slips past harmful-content filters, and that it frequently unfolds in groups where speakers reinforce or resist each other. The goal of SDARE-Bench is to measure both whether models recognise stigma and how they behave when asked to respond to it, across dyadic and multi-speaker dialogues.

Key Contributions

1. A first-of-its-kind scenario-based stigma benchmark. SDARE-Bench is described as the first benchmark to evaluate stigma in both dyadic queries and group dialogues, covering 93 stigma types drawn from psychological literature and operationalised through stereotypes, prejudice, and discrimination labels, four stigma sources, and interactional speaker roles.

2. Evaluation of open-ended response generation, not just static judgement. The benchmark moves beyond closed-form detection tasks to assess how models respond in prose as conversational assistants, in both one-to-one queries and four-speaker, eight-turn dialogues.

3. An expert-annotated response classifier for scalable evaluation. A multi-label classifier was trained to score model responses, built on 1,392 annotations from four human experts, enabling automated analysis of response failures.

4. A quality-controlled construction pipeline. Items were generated with an expert-in-the-loop, schema-guided process, screened by five harmful content detectors, calibration-rated by two psychologists with over 25 years of clinical experience, and judged at scale by Llama-3.1-70B-Instruct.

Main Findings

• Stigma presence is detected better than stigma components. Across the eight evaluated models, mean stigma presence accuracy was 69.05 in dyadic settings and 74.29 in group settings under Table 1, while mean hierarchy-aware macro accuracy was only 52.05 dyadic and 51.22 group. Stereotype, prejudice, and discrimination classification were consistently lower still.

• Component detection gets worse in group dialogues. Although stigma presence was detected more accurately in group dialogues, models performed consistently worse on stereotype, prejudice, and discrimination classification in that setting, which the authors attribute to greater conversational structure and social role complexity.

• DeepSeek-V3.1 was the strongest performer on detection. It led most individual classification dimensions, reaching 88.58 dyadic and 98.13 group stigma presence accuracy and 57.59 / 57.96 HMacroAcc. Qwen3-8B and Mistral-7B showed the weakest overall performance, especially on finer-grained component classification.

• Open-ended responses show substantially more stigma in group settings. In Table 2, mean stigma present rates were 31.04 in dyadic queries versus 69.18 in group dialogues. Discrimination rose from 27.08 to 67.52, stereotype from 25.40 to 64.23, and prejudice from 18.91 to 47.10.

• Pushback weakens in groups, unrealistic advice rises. Mean active pushback was 14.55 in dyadic settings but only 4.77 in group dialogues. Unrealistic advice rose from 14.42 to 42.07 and quality issues from 2.05 to 20.04. Overly generalised advice was the exception, with no consistent pattern (6.74 dyadic versus 7.69 group).

• Fixed-format benchmarks miss this. As an illustrative comparison, SocialStigmaQA produced a stigma rate of only 2.37%, against 31.04% for SDARE-Bench dyadic queries. The authors note the benchmarks differ in format and design, but read the contrast as evidence that open-ended conversational evaluation surfaces stigma expression that fixed response formats do not capture.

• Models reinforce stigma most when the user is the stigmatiser. In dyadic settings, responses expressed stigma 69.0% of the time when the user occupied a stigmatiser role and 63.5% for a reinforcer role, compared with 12.8% when responding to stigmatised targets and 25.0% for self-stigmatising target-stigmatisers.

• Constructed group pressure pushes stigma expression to near ceiling. Replacing the target and defender with additional reinforcers raised the mean stigma present rate from 79.9% to 97.5%, 95% CI [16.43, 18.73], Fisher's exact p < .001. Adjusted logistic regression estimated group pressure increased the odds of stigma expression by a factor of 12.0, 95% CI [7.57, 19.16], p < .001.

• Models rarely introduce stigma into clean prompts. Across all models there were 11 flagged dyadic responses and 3 flagged group responses where stigma appeared despite non-stigmatising inputs.

• The response classifier was validated and checked for shortcutting. Mean Gwet's AC1 was 0.823 for dyadic and 0.716 for group responses. In a sanity check holding 100 group pressure dialogues constant while replacing only the model response with a non-stigmatising one, the stigma positive rate dropped from 98.0% to 0% and mean predicted probability from .983 to .052, indicating predictions depended on response content rather than the stigmatising context alone.

• Providing label definitions did not change the picture. When models were re-evaluated with explicit stigma label definitions, dyadic source and role accuracy increased by 7.9 and 6.0 percentage points respectively, while group role accuracy decreased by 10.0 percentage points, and overall conclusions were unchanged.

Methodology in Plain English

The researchers built the benchmark in a staged pipeline. They started from 1,997 everyday activities in the American Time Use Survey Activity Lexicon (2024) and had three judges (GPT-5-mini, Claude-Haiku-4.5, and Gemini-2.5-Flash) independently rate each for plausibility, keeping 188 dyadic and 127 group scenario contexts. For each scenario the same judges ranked stigma types from a 93-category taxonomy, and the top five per scenario were retained.

Each item was then described by a structured schema covering five elements: the social scenario, the stigma type, the stigma source (public, self, structural, or associational), the stigma component (stereotype, prejudice, discrimination), and the conversational role. Roles follow bullying group dynamics: stigmatiser, target, reinforcer, defender, and bystander, with special variants that merge target and stigmatiser and variants that concentrate reinforcers to create group pressure. Single-turn schemas became user queries to an AI assistant; group schemas became four-speaker, eight-turn dialogues ending in an advice or wording request. Matched control items were generated from the same schema with no stigmatising content.

Prompts and schemas were iteratively refined in consultation with an anthropologist and a psychologist. To keep costs down, smaller models in each family (GPT-5-mini and Gemini-2.5-Flash) produced schemas, and larger counterparts (GPT-5 and Gemini-2.5-Pro) expanded them into text. Quality control ran in four steps: removal of any item flagged by five safety models, which removed 46 dyadic and 21 group items; human expert rating of a 100-item calibration subset by two psychologists; scaled review by an LLM judge, where Llama-3.1-70B-Instruct (MAE = 0.27) matched expert ratings more closely than Claude-Sonnet-4.5 (MAE = 0.36) and was used for all items; and pairwise selection between GPT and Gemini outputs, removing a further 3 dyadic and 117 group items. Human expert agreement on the calibration subset was MAE = 0.15.

Evaluation used eight models — DeepSeek-V3.1, Qwen2.5-72B-Instruct, Qwen3-8B, Nemotron-3-Super-120B-A12B, Mistral-Small-24B-Instruct-2501, Mistral-7B-Instruct-v0.3, Phi-4, and GLM-4.7-Flash — with one completion per prompt and deterministic decoding where supported (temperature = 0). Task I asked models to return six fields (stigma presence, source, stereotype, prejudice, discrimination, role). Task II asked for open prose responses with no response guidelines or stigma correction, to reflect ordinary deployment. Responses were scored by a DeBERTa-v3-large classifier that was domain-adapted with masked-language-model training on an unlabelled in-domain corpus, then fine-tuned with eight binary classification heads on 1,392 expert annotations using class-weighted cross-entropy and five-fold cross-validation.

Why This Matters

Research impact. The paper argues that closed-form, static stigma evaluations underestimate failure. It shows a detection-versus-response gap: models can often say stigma is present but cannot characterise its components, and when generating responses they frequently reinforce stigma, resist it weakly, give unrealistic advice, and occasionally introduce stigma where none existed. It also reframes multi-speaker context as a distinct safety surface rather than a variation of the one-to-one case.

Real-world applications:

• Group-facing AI assistants, such as group chats in ChatGPT, Meta AI in WhatsApp groups, and Claude for Slack, where the group pressure condition produced a 97.5% mean stigma expression rate.
• Workplace and collaborative tools, where models may align with whoever is speaking rather than with the person being stigmatised.
• Health and clinical decision support, given that group dialogues produced higher stigma expression across stigma sources, scenario categories, and stigma clusters, including legal, healthcare, childcare, and employment scenarios.
• Advice-seeking use in general, where responses showed unrealistic advice rates of 42.07 in group settings and 14.42 in dyadic.

Industry relevance. The benchmark supplies a classifier, rubric, and labelled data that teams can reuse to audit models for stigma-related failures that standard harmful-content filters do not flag, and it points to a specific failure mode — compliance with a socially dominant framing in the input — that current alignment work does not directly target.

Future Directions

• Multilingual and cross-cultural extension. The benchmark is English, text-only, and the authors note stigma is culturally situated, so results should not be treated as a universal measure of stigma safety.

• Controlled ablations on what drives the group gap. SDARE-Bench compares dyadic and group formats as realistic interaction styles rather than isolating the effects of speaker number, context length, turn-taking, and distributed social cues; the authors call for controlled ablations to separate these contributions.

• Moving from measurement to mitigation. The rubric measures whether models reinforce or resist stigma but does not prescribe ideal responses or develop interventions. The authors suggest finer-grained response guidelines that help models challenge stigma while remaining helpful and context-sensitive.

• Investigating sycophancy as a stigma mechanism. The role-based and group pressure results suggest models prioritise alignment with the user's or group's apparent goal rather than protecting the stigmatised speaker, which the authors flag as an additional failure mode worth targeted study.

Target Audience

Researchers and practitioners in LLM safety, alignment, and evaluation; computational social scientists and psychologists studying stigma, prejudice, and group dynamics; teams building or auditing conversational assistants deployed in multi-user or advice-giving settings; and benchmark developers interested in open-ended, scenario-based evaluation methods. Readers should note the paper's warning that it contains stigma and offensive content solely for research purposes, and its stated intent that the benchmark be used for research and safety evaluation rather than to justify deployment decisions or rank models as safe for use.

Authors’ abstract

Large Language Models (LLMs) are increasingly used in advice seeking and decision making that may affect social judgements. Despite stigma's profound effects on people and communities, benchmarks remain scarce. Existing general-domain evaluations typically rely on static prompts and fixed-format tasks, overlooking conversational contexts and audience effects in everyday communication. To address these gaps, we introduce SDARE-Bench, the first scenario-based benchmark evaluating both stigma detection and open-ended response generation in LLMs, comprising 1,138 dyadic queries and 1,388 group dialogue. Empirical results across 8 LLMs consistently demonstrate poor identification of stigma components, especially in group dialogues. In open-ended response generation, stigma expression was substantially higher in group settings than in dyadic, with weaker resistance to stigma and more unrealistic advice. Responses were evaluated using a classifier trained on 1,392 human annotated responses. In constructed group pressure settings, stigma expression rates further increased to a striking average of 97.5%. Our findings identify stigma response as a recurring LLM safety vulnerability, especially in socially complex conversational contexts.