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

Researchers have introduced MIRAGE, a benchmark designed to evaluate anti-Muslim bias in frontier large language models under realistic deployment conditions, five years after such bias was first identified in these systems [1]. The benchmark, formally named the Muslim-Identity Reasoning and Agentic Generation Evaluation, consists of 1,200 prompts that move beyond the single-turn prompt completion tests that have dominated most evaluations to date [1][2]. The study’s authors argue that this older methodology no longer reflects how frontier LLMs are deployed [2]. MIRAGE instead tests models across three conditions: direct completion, chain-of-thought reasoning, and simulated agentic decision-making in scenarios including content moderation, lending triage, refugee claim summarization, and hiring screens [1][2]. Testing across six frontier models revealed that more complex reasoning did not reduce bias. Chain-of-thought reasoning amplified Muslim-violence associations by 12–34% relative to direct completion [1][2]. In agentic decision-making tasks, the models exhibited a 9–22 percentage-point asymmetry between Muslim and matched non-Muslim cases, even when presented with identical evidence [1][2]. The research also found that bias is sharply time-coupled to retrieved news context. When models retrieved information related to recent conflicts, measured bias increased by 18–27% [1][2]. Existing prompt-based mitigation techniques transferred poorly across the three conditions. While they suppressed bias in direct-completion tasks, agentic asymmetry remained largely intact [2]. The paper was submitted by Noor Noor S. Mohammad on June 15, 2026 [1]. The authors have released the MIRAGE benchmark and an open evaluation harness to support targeted mitigation research [1][2].