Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Contextual StereoSet: Stress-Testing Bias Alignment Robustness in Large Language Models

Abhinaba Basu, Pavan Chakraborty

TL;DR

This work reframes bias evaluation in large language models as a context-dependent phenomenon rather than a fixed property, introducing Contextual StereoSet to systematically vary framing across location, year, style, and observer. It pairs this benchmark with Context Sensitivity Fingerprints (CSF) to compactly report dispersion and contrasts, enabling robust stress-testing across two protocols and 13 models. Empirically, bias sensitivity persists across domains (employment, credit, policing) and languages, with consistent temporal effects (1990 vs 2030) and context-driven amplification patterns (e.g., gossip framing, out-group observers). The authors argue that deployment-relevant risk requires context-aware audits, not single-number bias scores, and provide a reproducible pipeline and actionable guidance for model selection, mitigation, and policy alignment. Overall, CSF advances practical, fine-grained bias evaluation by focusing on “under what conditions does bias appear, and for whom?” rather than asking if a model is biased in isolation.

Abstract

A model that avoids stereotypes in a lab benchmark may not avoid them in deployment. We show that measured bias shifts dramatically when prompts mention different places, times, or audiences -- no adversarial prompting required. We introduce Contextual StereoSet, a benchmark that holds stereotype content fixed while systematically varying contextual framing. Testing 13 models across two protocols, we find striking patterns: anchoring to 1990 (vs. 2030) raises stereotype selection in all models tested on this contrast (p<0.05); gossip framing raises it in 5 of 6 full-grid models; out-group observer framing shifts it by up to 13 percentage points. These effects replicate in hiring, lending, and help-seeking vignettes. We propose Context Sensitivity Fingerprints (CSF): a compact profile of per-dimension dispersion and paired contrasts with bootstrap CIs and FDR correction. Two evaluation tracks support different use cases -- a 360-context diagnostic grid for deep analysis and a budgeted protocol covering 4,229 items for production screening. The implication is methodological: bias scores from fixed-condition tests may not generalize.This is not a claim about ground-truth bias rates; it is a stress test of evaluation robustness. CSF forces evaluators to ask, "Under what conditions does bias appear?" rather than "Is this model biased?" We release our benchmark, code, and results.

Contextual StereoSet: Stress-Testing Bias Alignment Robustness in Large Language Models

TL;DR

This work reframes bias evaluation in large language models as a context-dependent phenomenon rather than a fixed property, introducing Contextual StereoSet to systematically vary framing across location, year, style, and observer. It pairs this benchmark with Context Sensitivity Fingerprints (CSF) to compactly report dispersion and contrasts, enabling robust stress-testing across two protocols and 13 models. Empirically, bias sensitivity persists across domains (employment, credit, policing) and languages, with consistent temporal effects (1990 vs 2030) and context-driven amplification patterns (e.g., gossip framing, out-group observers). The authors argue that deployment-relevant risk requires context-aware audits, not single-number bias scores, and provide a reproducible pipeline and actionable guidance for model selection, mitigation, and policy alignment. Overall, CSF advances practical, fine-grained bias evaluation by focusing on “under what conditions does bias appear, and for whom?” rather than asking if a model is biased in isolation.

Abstract

A model that avoids stereotypes in a lab benchmark may not avoid them in deployment. We show that measured bias shifts dramatically when prompts mention different places, times, or audiences -- no adversarial prompting required. We introduce Contextual StereoSet, a benchmark that holds stereotype content fixed while systematically varying contextual framing. Testing 13 models across two protocols, we find striking patterns: anchoring to 1990 (vs. 2030) raises stereotype selection in all models tested on this contrast (p<0.05); gossip framing raises it in 5 of 6 full-grid models; out-group observer framing shifts it by up to 13 percentage points. These effects replicate in hiring, lending, and help-seeking vignettes. We propose Context Sensitivity Fingerprints (CSF): a compact profile of per-dimension dispersion and paired contrasts with bootstrap CIs and FDR correction. Two evaluation tracks support different use cases -- a 360-context diagnostic grid for deep analysis and a budgeted protocol covering 4,229 items for production screening. The implication is methodological: bias scores from fixed-condition tests may not generalize.This is not a claim about ground-truth bias rates; it is a stress test of evaluation robustness. CSF forces evaluators to ask, "Under what conditions does bias appear?" rather than "Is this model biased?" We release our benchmark, code, and results.
Paper Structure (77 sections, 2 equations, 3 figures, 12 tables)

This paper contains 77 sections, 2 equations, 3 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Contributions
  3. Motivating Examples
  4. Employment.
  5. Credit.
  6. Police help.
  7. Linguistic framing.
  8. Summary.
  9. Theoretical Motivation
  10. Related Work
  11. Bias benchmarks.
  12. Sociotechnical perspectives.
  13. Prompt sensitivity.
  14. Our distinction.
  15. Positioning.
  16. ...and 62 more sections

Figures (3)

  • Figure 1: Contextual StereoSet evaluation pipeline. We vary socio-cultural framing (location, year, style, observer) while holding the underlying stereotype probe fixed, then summarize behavior as a context sensitivity fingerprint (CSF) for cross-model comparison.
  • Figure 2: Key context contrasts across models at $T{=}0$. Bars show mean paired $\Delta SS$ on the full-grid intrasentence set (Table \ref{['tab:fullgrid-contrasts']}).
  • Figure 3: DeepSeek v3.2 location$\times$year marginal stereotype-selection rates at $T{=}0$ (50 intrasentence items; averaged over style and observer). Rates range from 0.600 to 0.763 across cells, illustrating substantial context dependence even without changing item content.