MusIML Workshop · ICML 2026 · Seoul

MIRAGE

Auditing Anti-Muslim Bias in Frontier LLMs
Across Reasoning, Agentic, and Time-Coupled Conditions

Noor Islam S. Mohammad1,★ Istanbul Technical University islam23@itu.edu.tr
Tamim Sheikh2 Jashore University of Science & Technology
1 Dept. of Computer Science, Informatics Institute, ITU, İstanbul, Türkiye 2 Dept. of Computer Science and Engineering, JUST, Bangladesh Corresponding author
Abstract

Five years after the discovery of persistent anti-Muslim bias in large language models, most evaluations remain confined to single-turn prompt completion — a setting that no longer reflects how frontier LLMs are deployed. We introduce MIRAGE (Muslim-Identity Reasoning and Agentic Generation Evaluation), a benchmark of 1,200 prompts spanning three deployment-realistic conditions: direct completion, chain-of-thought reasoning, and simulated agentic decision-making across content moderation, lending triage, refugee claim summarization, and hiring screens. Across six frontier models, we find that chain-of-thought reasoning amplifies rather than suppresses Muslim-violence associations by 12–34% relative to direct completion, agentic decisions exhibit a 9–22 percentage-point asymmetry between Muslim and matched non-Muslim cases on identical evidence, and bias is sharply time-coupled to retrieved news context, increasing 18–27% under recent-conflict retrieval.

Key findings

Three deployment gaps prior work missed

12–34%

CoT amplification. Chain-of-thought reasoning increases Muslim-violence associations relative to direct completion — contrary to the assumption that reasoning suppresses bias.

9–22pp

Agentic asymmetry. Percentage-point gap between Muslim and matched non-Muslim decisions on identical evidence across moderation, lending, refugee, and hiring harnesses.

18–27%

Time-coupled escalation. Bias increases under recent-conflict RAG retrieval, operationalizing the news-cycle coupling first noted by Abid et al. (2021).

Existing prompt-based mitigations transfer poorly: they suppress direct-completion bias while leaving agentic decision asymmetry largely intact, revealing a systematic gap between what benchmarks measure and what deployed systems do.


Benchmark design

Three conditions, six models, five identities

C₁
Direct completion

Baseline single-turn prompts across six templates (T1–T6): open continuation, analogical, biographical, news summarization, decision context, and dialogue.

C₂
Chain-of-thought

Same prompts with CoT elicitation. Reasoning traces audited for three amplification pathways: group-statistic invocation, reasoner-as-stand-in, and coherence pressure.

C₃
Agentic decisions

Four high-stakes harnesses — content moderation (H1), lending triage (H2), refugee summarization (H3), hiring screening (H4) — with calibrated decision ambiguity designed to surface asymmetric bias.

Matched-pair counterfactuals

Every prompt pairs a Muslim-identifying variant with a non-Muslim control (Christian, Jewish, Hindu, secular) via a single lexical substitution. All surrounding content is held fixed, isolating the identity signal causally.

Cross-lingual parity

400 prompts translated into MSA and three Arabic dialects (Egyptian, Levantine, Maghrebi) by bilingual speakers with independent post-editing, enabling language-mediated vs. model-internal bias decomposition.

1,200 matched-pair prompts

240 per template × 5 identity variants = 1,200 base prompts per condition. A 15% stratified lite split supports lower-resource replication.

Six frontier models

Claude Opus 4, GPT-4o, Gemini 1.5 Pro (closed, high-safety); Llama-3.3-70B, Qwen2.5-72B, DeepSeek-V3 (open, instruction-tuned).


Results — Chain-of-thought

How reasoning amplifies bias: three pathways

34.2%
Group-statistic invocation

Model frames a stereotyped inference as a probabilistic prediction grounded in "base rates," then endorses it as statistically expected. Most prevalent in open models (44.6% of amplifying traces).

28.7%
Reasoner-as-stand-in

Stereotype framed as what a "reasonable" or "neutral" observer would expect — licensing it as social norm rather than factual claim. Equally prevalent across all model tiers.

12.1%
Coherence pressure

An early mild negative inference commits subsequent reasoning steps to develop it into a stereotyped conclusion for chain coherence. Absent under direct completion (C₁).

Model tier P1 — group stats P2 — reasoner P3 — coherence Other / mixed
Closed, high-safety 18.3% 31.4% 14.2% 36.1%
Closed, standard 27.9% 29.1% 11.8% 31.2%
Open 44.6% 27.2% 10.3% 17.9%

Distribution of CoT amplification pathways (% of amplifying traces) by model tier. The alignment layer suppresses identical inferences under C₁ — CoT provides an intermediate output position that bypasses output-level safety training.


Results — Violence rates & agentic asymmetry

Model-level measurements across C₁ and C₃

Model C₁ violence rate C₁→C₂ Spearman ρ C₁→C₃ Spearman ρ Mean δ (agentic)
Claude Opus 4 0.58 0.29
GPT-4o 0.55 0.31
Gemini 1.5 Pro 0.60 0.33
Llama-3.3-70B 18.7% 0.64 0.37
Qwen2.5-72B 23.9% 0.66 0.38
DeepSeek-V3 29.4% 0.63 0.36
Mean 0.61 0.34

The weak C₁–C₃ Spearman ρ = 0.34 confirms that ~38% of high-δ agentic items have low C₁ violence rates — decision asymmetry is not reducible to surface-level association strength. A C₁-only benchmark will both underestimate the deployed harm surface and misrank mitigations.


Mitigation audit

Four defenses, three conditions — transfer fails

M1
Cultural prompting

System-prompt prefix asserting cultural awareness and equal treatment across groups.

C₁
–68%
C₂
–38%
C₃
–14%
M2
Affective priming

Pre-prompt positive-valence statement framing subjects as valued community members.

C₁
–55%
C₂
–27%
C₃
–9%
M3
Self-debiasing

Pre-generation suffix asking the model to identify stereotypes before responding.

C₁
–72%
C₂
–44%
C₃
–18%
M4
Generate–Critique–Revise

Two-turn pipeline: unconstrained generation followed by fairness-criterion critique and revision.

C₁
–81%
C₂
–52%
C₃
–22%

Bar lengths show bias reduction relative to M0 baseline (larger = better). strong, partial, poor transfer. M4 is the strongest single mitigation but still leaves 78% of agentic asymmetry unaddressed — the structural gap motivates dedicated agentic-condition alignment work.


Contributions

What MIRAGE adds to the literature

1 — MIRAGE benchmark

1,200 matched-pair prompts across direct completion, chain-of-thought, and agentic conditions, with parallel English/Arabic translations covering MSA and three dialects.

2 — CoT amplification finding

Quantitative audit across six frontier models demonstrating that chain-of-thought prompting amplifies rather than suppresses anti-Muslim bias, with three identified mechanistic pathways.

3 — Agentic decision asymmetry

9–22 pp decision asymmetry documented across four high-stakes harnesses under identical evidence. Refugee-claim summarization shows the largest gap.

4 — Time-coupled RAG + mitigation audit

First explicit measurement of news-cycle-coupled bias escalation under RAG, plus a cross-condition mitigation audit showing that prompt-based defenses do not transfer to agentic settings.


Citation

How to cite this work

BibTeX
@inproceedings{mohammad2026mirage,
  title     = {{MIRAGE}: Auditing Anti-Muslim Bias in Frontier {LLMs}
               Across Reasoning, Agentic, and Time-Coupled Conditions},
  author    = {Mohammad, Noor Islam S. and Sheikh, Tamim},
  booktitle = {6th Muslims in {ML} Workshop at the 43rd International
               Conference on Machine Learning},
  address   = {Seoul, South Korea},
  year      = {2026},
  note      = {Copyright 2026 by the author(s)}
}

Presented at the 6th Muslims in ML (MusIML) Workshop at the 43rd International Conference on Machine Learning, Seoul, South Korea. Copyright 2026 by the author(s).