W&C-Sent: Warmth & Competence in SCM

• W&C-Sent is a sentence-level framework that operationalizes the Stereotype Content Model by assessing social perceptions along warmth (trust and sociability) and competence dimensions.
• It provides both a benchmark protocol and an annotated dataset, enabling the mapping of social groups through structured judgments rather than undifferentiated sentiment.
• Applications span LLM evaluation, debiasing assessments, and human–AI interaction studies, addressing complex issues like ambivalence in stereotype content.

Searching arXiv for papers on Warmth and Competence Sentences, SCM-based stereotype analysis, and related sentence-level resources. Warmth and Competence Sentences (W&C-Sent) denotes sentence-level resources and evaluation protocols that represent social perception along the two principal dimensions of the Stereotype Content Model (SCM): warmth and competence. In the recent literature, the term has two closely related usages. One is a practically specified benchmark design, derived from StereoMap, for eliciting keyword-supported and reason-supported sentence outputs from LLMs about social groups and mapping those outputs into a Warmth–Competence space (Jeoung et al., 2023). The other is the first formally released sentence-level dataset bearing the name W&C-Sent: 1,633 English sentence–target pairs annotated for trust, sociability, and competence, where trust and sociability are treated as facets of warmth (Ayesh et al., 9 Jan 2026). Across both usages, W&C-Sent operationalizes social evaluation not as undifferentiated sentiment, but as structured judgments about perceived intent, morality, sociability, and capability.

1. Conceptual foundations in the Stereotype Content Model

SCM organizes social perception along two primary dimensions. Warmth concerns perceived intent: whether a target is friendly, trustworthy, sincere, sociable, or, conversely, adversarial and immoral. Competence concerns perceived capability: whether a target is competent, skilled, intelligent, confident, assertive, or instead incapable and dependent (Jeoung et al., 2023). In the sentence-level dataset literature, warmth is further decomposed into trust and sociability. Trust covers moral or personal regard, including honesty and sincerity, whereas sociability covers relational regard, including friendliness, generosity, and helpfulness (Ayesh et al., 9 Jan 2026).

A central SCM claim is that stereotypes are frequently ambivalent rather than uniformly positive or negative. High warmth with low competence corresponds to paternalistic or pity-based evaluations; high competence with low warmth corresponds to envy; high values on both dimensions correspond to admiration; and low values on both correspond to contempt or hate. This quadrant structure is also tied to the BIAS Map, which links warmth and competence to characteristic emotions and behavioral tendencies, such as active or passive facilitation, exclusion, and harm (Jeoung et al., 2023).

Computational work preceding W&C-Sent translated this theory into embedding-space geometry. Warmth and competence axes were defined by difference-of-means vectors over positive and negative seed lexicons, and lexical or group representations were projected into a two-dimensional SCM subspace (Fraser et al., 2021). This established a technical foundation for later sentence-level work: once warmth and competence are treated as explicit dimensions rather than latent sentiment, they become measurable at the level of prompts, reasons, and target-conditioned utterances.

2. From lexical SCM resources to sentence-level W&C-Sent

Before W&C-Sent, most computational SCM work operated at the word or embedding level. "Understanding and Countering Stereotypes: A Computational Approach to the Stereotype Content Model" defined warmth and competence axes in semantic embedding space and validated them against expanded lexicons, reporting that FastText Common Crawl subword embeddings achieved 85.0% accuracy for warmth and 85.8% for competence on a 3,159-word validation set (Fraser et al., 2021). "Social-Group-Agnostic Word Embedding Debiasing via the Stereotype Content Model" then used 15 warmth antonym pairs and 15 competence antonym pairs to construct an SCM bias subspace for debiasing static embeddings across gender, race, and age (Omrani et al., 2022).

These earlier efforts did not provide a sentence dataset. The 2022 debiasing paper explicitly stated that it did not introduce a sentence-level resource or templates named W&C-Sent (Omrani et al., 2022). The shift to sentence-level analysis therefore marks a methodological transition from lexical polarity and projection diagnostics to contextualized judgments that must handle negation, sarcasm, target specificity, and pragmatic framing.

Resource type	Representation	Main contribution
SCM lexicon/projection work	Words and embedding axes	Warmth/competence subspaces and quadrant structure (Fraser et al., 2021)
SCM-based debiasing	Antonym pairs in embeddings	Social-group-agnostic bias subspace using warmth and competence pairs (Omrani et al., 2022)
StereoMap-derived W&C-Sent protocol	Prompted keywords, reasons, ratings	Sentence-level elicitation and group mapping for LLMs (Jeoung et al., 2023)
W&C-Sent dataset	Sentence–target pairs	First released sentence-level corpus annotated for trust, sociability, and competence (Ayesh et al., 9 Jan 2026)

This progression suggests that W&C-Sent is best understood as the sentence-level instantiation of SCM: it retains the theory’s dimensional structure while moving from isolated lexical cues to contextual language about explicit targets.

3. StereoMap and the benchmark-style formulation of W&C-Sent

StereoMap operationalizes SCM for LLMs through prompt-based elicitation of ratings, keywords, and reason sentences about socio-demographic groups (Jeoung et al., 2023). In the benchmark-oriented formulation derived from this work, W&C-Sent consists of sentence-level outputs paired with numerical judgments. The core prompt family asks how a group is viewed by society, rates warmth and competence items on a 5-point Likert scale, and requests keywords plus explanatory sentences.

The primary warmth items are Friendly, Sociable, Trustworthy, Honest, and the competence items are Competent, Skilled, Confident, Assertive. An extended version adds Traditional, Conservative, Wealthy, High-status, together with emotion and behavior items such as Contempt, Disgust, Admire, Proud, Pity, Sympathy, Envious, Jealous and Help, Protect, Fight, Attack, Cooperate, Associate, Exclude, Demean (Jeoung et al., 2023). The protocol aggregates results over 10 runs per model to stabilize estimates.

StereoMap computes per-group scores by averaging items within each dimension and then averaging across runs:

$$W_g^{(r)} = \frac{1}{M_W}\sum_{m \in \text{Warmth}} s_{g,m}^{(r)}, \qquad W_g = \frac{1}{R}\sum_{r=1}^R W_g^{(r)}$$

$$C_g^{(r)} = \frac{1}{M_C}\sum_{m \in \text{Competence}} s_{g,m}^{(r)}, \qquad C_g = \frac{1}{R}\sum_{r=1}^R C_g^{(r)}$$

It optionally normalizes these scores across groups by either $z$-score or min–max scaling, and computes inter-group distances in the Warmth–Competence plane as

$D_{ij} = \sqrt{(W_i - W_j)^2 + (C_i - C_j)^2}.$

For structure discovery, StereoMap uses K-means with $k=5$ over $(W_g, C_g)$, and paired $t$-tests assess whether competence exceeds warmth, warmth exceeds competence, or the two are statistically indistinguishable within a cluster at $p < .001$ (Jeoung et al., 2023). Correlational analyses primarily use Spearman’s $\rho$ to connect stereotype dimensions with emotion and behavior ratings.

The sentence component is not incidental. Models are explicitly asked to list keywords and provide one-sentence reasons. The reported example outputs include warmth-oriented explanations such as “Caring: They are dedicated to helping others and providing them with the best possible care,” and competence-oriented explanations such as “Skilled: They demonstrate strong clinical skills and problem-solving in fast-paced environments” (Jeoung et al., 2023). This makes the protocol a sentence-level resource in practice, even before the release of a manually annotated corpus.

4. The first formal W&C-Sent dataset

"Annotating Dimensions of Social Perception in Text: The First Sentence-Level Dataset of Warmth and Competence" introduced the first released dataset explicitly named W&C-Sent (Ayesh et al., 9 Jan 2026). It contains 1,633 English sentence–target pairs, not merely sentences in isolation. This target-conditioned design is central: the same sentence may be paired with multiple targets, and annotations are given with respect to one specified target at a time.

The dataset draws primarily from social media. Of the 1,633 pairs, 1,482 come from selected subsets of SemEval-2016 Task 6 Stance in Tweets and 151 from the ABCDE corpus. Targets include three individuals—Hillary Clinton, Donald Trump, Barack Obama—and four social groups—women, religious people, nonreligious people or atheists, and climate change activists or environmentalists (Ayesh et al., 9 Jan 2026). Sentences were retained only when they expressed attitudes toward humans, and original stance targets were sometimes reassigned to human referents, as when climate-change posts were reinterpreted as targeting environmentalists.

Each pair is annotated independently for trust, sociability, and competence on a 7-point ordinal scale from $-3$ to $$C_g^{(r)} = \frac{1}{M_C}\sum_{m \in \text{Competence}} s_{g,m}^{(r)}, \qquad C_g = \frac{1}{R}\sum_{r=1}^R C_g^{(r)}$$0, where $$C_g^{(r)} = \frac{1}{M_C}\sum_{m \in \text{Competence}} s_{g,m}^{(r)}, \qquad C_g = \frac{1}{R}\sum_{r=1}^R C_g^{(r)}$$1 denotes neutral, not applicable, or not expressed. The annotation guidelines instruct raters to distinguish morality from capability, to consider sarcasm and irony, and to remain target-specific. This is especially important because competence is explicitly treated as morally valence-agnostic: a sentence may portray a target as highly competent while also portraying that target as untrustworthy or antisocial (Ayesh et al., 9 Jan 2026).

Dataset property	Value
Size	1,633 sentence–target pairs
Dimensions	Trust, Sociability, Competence
Label scale	$$C_g^{(r)} = \frac{1}{M_C}\sum_{m \in \text{Competence}} s_{g,m}^{(r)}, \qquad C_g = \frac{1}{R}\sum_{r=1}^R C_g^{(r)}$$2 to $$C_g^{(r)} = \frac{1}{M_C}\sum_{m \in \text{Competence}} s_{g,m}^{(r)}, \qquad C_g = \frac{1}{R}\sum_{r=1}^R C_g^{(r)}$$3, with $$C_g^{(r)} = \frac{1}{M_C}\sum_{m \in \text{Competence}} s_{g,m}^{(r)}, \qquad C_g = \frac{1}{R}\sum_{r=1}^R C_g^{(r)}$$4 for neutral/not expressed
Annotators	4–7 per pair per dimension; 216 total
Split	60/20/20 train/dev/test with grouped splits by text

Quality control was substantial. Annotators were recruited on Prolific, had fluent English and high prior approval rates, and were paid USD $$C_g^{(r)} = \frac{1}{M_C}\sum_{m \in \text{Competence}} s_{g,m}^{(r)}, \qquad C_g = \frac{1}{R}\sum_{r=1}^R C_g^{(r)}$$5 per hour. The study used attention checks with predetermined labels and 229 author-preannotated sentences for spot checks (Ayesh et al., 9 Jan 2026). Reliability was moderate and dimension-dependent: split-half reliability was 0.76 for trust, 0.68 for sociability, and 0.56 for competence; Krippendorff’s $$C_g^{(r)} = \frac{1}{M_C}\sum_{m \in \text{Competence}} s_{g,m}^{(r)}, \qquad C_g = \frac{1}{R}\sum_{r=1}^R C_g^{(r)}$$6 was 0.60, 0.50, and 0.30 respectively. When labels were coarsened into Low, Neutral, and High, agreement increased, which indicates that fine-grained competence judgments are particularly difficult at sentence level.

5. Empirical patterns revealed by W&C-Sent-style resources

StereoMap showed that LLMs produce mixed-dimension stereotype placements that resemble human SCM regularities (Jeoung et al., 2023). Across models, Asians, Jews, and rich people were repeatedly placed in high-competence, low-warmth clusters; elderly people were consistently placed in high-warmth, low-competence clusters; homeless people were consistently low on both dimensions; and middle-class, educated, and professional groups often appeared in high-warmth, high-competence regions. In extended prompts, nurses and doctors also occupied high–high positions. These placements were accompanied by BIAS Map-consistent emotion and behavior patterns: high–high clusters attracted the highest admiration, low–low clusters showed higher contempt and pity, warmth correlated positively with active facilitation and negatively with harm, and competence correlated with passive facilitation.

The same study also characterized stylistic differences in reasoning. Bard frequently cited statistics and institutional sources such as the U.S. Census Bureau, Pew Research Center, and BLS projections, whereas text-davinci-003 and GPT-3.5 emphasized public perception, media framing, education, resources, and inequality (Jeoung et al., 2023). This indicates that W&C-Sent outputs can encode not only stereotype content but also a model’s preferred justificatory register.

The formal W&C-Sent dataset exposed complementary regularities in human-annotated sentence data (Ayesh et al., 9 Jan 2026). Negative judgments dominate, especially for trust and sociability, reflecting the contentious social-media domains from which the data were drawn. Coarse-label distributions are skewed toward Low for all three dimensions: 57.87% of trust labels, 61.97% of sociability labels, and 47.34% of competence labels are Low. Yet the dimensions are not reducible to one another. Aggregated score correlations are positive but imperfect—$$C_g^{(r)} = \frac{1}{M_C}\sum_{m \in \text{Competence}} s_{g,m}^{(r)}, \qquad C_g = \frac{1}{R}\sum_{r=1}^R C_g^{(r)}$$7 for trust–sociability, $$C_g^{(r)} = \frac{1}{M_C}\sum_{m \in \text{Competence}} s_{g,m}^{(r)}, \qquad C_g = \frac{1}{R}\sum_{r=1}^R C_g^{(r)}$$8 for sociability–competence, and $$C_g^{(r)} = \frac{1}{M_C}\sum_{m \in \text{Competence}} s_{g,m}^{(r)}, \qquad C_g = \frac{1}{R}\sum_{r=1}^R C_g^{(r)}$$9 for trust–competence—leaving ample room for sentences that depict a target as morally bad but effective, or friendly but incapable.

The relation to stance and sentiment is strong but non-identical. On the 854 pairs overlapping with SemEval stance, trust, sociability, and competence correlate with stance at $z$0, $z$1, and $z$2 respectively; correlations with sentiment are similar (Ayesh et al., 9 Jan 2026). A plausible implication is that W&C-Sent captures structured social evaluation beyond plain polarity: sentiment can indicate whether language is favorable, but W&C-Sent indicates how it is favorable or unfavorable.

6. Modeling, applications, and adjacent research programs

Modeling results on the released W&C-Sent corpus confirm that sentence-level social perception is nontrivial. In fine-grained 7-class prediction, GPT-4o zero-shot achieved the strongest reported trust performance at 0.43 accuracy, 0.42 weighted F1, and 0.91 $z$3 accuracy; fine-tuned BERTweet led on sociability with 0.46 accuracy, 0.34 F1, and 0.88 $z$4 accuracy; and competence remained the hardest dimension, with fine-tuned BERTweet reaching 0.36 accuracy and 0.24 F1 while GPT-4o zero-shot reached 0.34 accuracy and 0.31 F1 (Ayesh et al., 9 Jan 2026). In coarse-grained 3-class prediction, GPT-4o zero-shot was strongest overall by weighted F1, with 0.78 for trust, 0.71 for sociability, and 0.59 for competence. Simpler baselines such as dummy classifiers and TF–IDF logistic regression lagged substantially.

W&C-Sent is also part of a wider SCM-centered computational ecosystem. SCM-based debiasing in static embeddings showed that using only warmth and competence antonym pairs can achieve ECT results close to group-specific debiasing across gender, race, and age, with Partial Projection using SCM pairs reaching 0.97 for gender, 0.96 for race, and 0.95 for age while preserving more utility than debiasing across combined group-specific subspaces (Omrani et al., 2022). This positions W&C-Sent as a natural evaluation layer for downstream models that attempt to reduce stereotype content rather than merely measure it.

Beyond stereotype analysis, warmth and competence have been used as explanatory dimensions in human–AI interaction. In mixed-motive cooperation with RL agents, perceived warmth and competence predicted stated preferences, and warmth predicted revealed partner choice better than objective score, indicating that sentence-level W&C judgments can be behaviorally consequential (McKee et al., 2022). In conversational AI persona design, priming users with metaphor sentences such as “The bot you are about to interact with is modeled after a ‘toddler’” or a “trained professional travel assistant” changed post-use usability, adoption intention, and desire to cooperate, even when the underlying behavior was held constant by Wizard-of-Oz control (Khadpe et al., 2020). In large-language-model interaction studies, manipulated warmth and competence altered trust, usefulness, anthropomorphism, closeness, and frustration, with warmth primarily affecting relational outcomes and competence primarily affecting epistemic outcomes (Kadambi et al., 1 Mar 2026).

Related work has extended the same dimensional vocabulary to public discourse and speech. Cross-cultural analysis of social-media discussions around conversational agents found that warmth was the key driver of positive emotion in both the United States and China, although modality and embodiment effects differed across countries (Liu et al., 2024). In synthetic speech, feature-conditioned Tacotron systems manipulated perceived warmth and competence in female voices through acoustic properties such as F1 mean, F2 mean, spectral flux, and voiced slope, with convex feature combinations achieving the highest Mean Opinion Scores for warmth and competence (Rallabandi et al., 2022). These developments do not define W&C-Sent directly, but they show that sentence-level warmth and competence are part of a broader computational program spanning text, interaction, and speech.

7. Limitations, controversies, and likely research directions

Several limitations recur across W&C-Sent work. StereoMap is explicitly US-centric in its target groups and societal framing, and prompt sensitivity plus refusal-to-answer behavior can materially affect outputs; Bard, for example, showed near-total refusal for “Poor Blacks” in the supplied summary (Jeoung et al., 2023). The released W&C-Sent dataset is English-only, largely centered on contentious U.S. social-media discourse, and annotated primarily by raters from the United States and the United Kingdom (Ayesh et al., 9 Jan 2026). Cross-cultural variation in warmth and competence associations is therefore not peripheral but structurally important, as shown by the divergent U.S.–China discourse patterns in conversational-agent research (Liu et al., 2024).

A second limitation concerns representation granularity. Dictionary-based mappings and predefined target inventories can miss intersectional, implicit, or discourse-level stereotypes (Jeoung et al., 2023). The sentence-level corpus improves contextual coverage, but even there competence proved difficult to annotate reliably, with Krippendorff’s $z$5 (Ayesh et al., 9 Jan 2026). This suggests that competence is often encoded indirectly, through pragmatics, implication, role expectations, or comparisons, rather than through overt evaluative adjectives alone.

Ethical concerns are also central. W&C-Sent resources may expose harmful narratives while trying to measure them. StereoMap notes that LLMs may reproduce stigmatizing or deficit-framed content even when instructed to speak only about “society’s view,” and recommends normative review, content filters, refusal-rate tracking, and downstream audits (Jeoung et al., 2023). The 2026 dataset includes offensive content, prohibits high-risk or commercial uses without explicit approval, and emphasizes that models trained on the data should not be used for high-stakes individual judgments (Ayesh et al., 9 Jan 2026).

Current research directions are correspondingly clear. One is multilingual and cross-cultural expansion, since present sentence-level resources remain narrow relative to the scope of SCM claims (Liu et al., 2024). Another is stronger target-aware modeling that can handle sarcasm, irony, and multi-target reference more robustly (Ayesh et al., 9 Jan 2026). A third is integration with modern contextual debiasing methods, since earlier SCM-based debiasing was demonstrated on static embeddings rather than contextual LMs (Omrani et al., 2022). More broadly, W&C-Sent points toward a computational social-perception framework in which social evaluation is modeled explicitly as trust, sociability, and competence rather than collapsed into sentiment, toxicity, or stance alone.