Mirror: A Multi-Agent System for AI-Assisted Ethics Review

TL;DR

Mirror addresses the growing strain on ethics review in AI-enabled, interdisciplinary research by integrating a domain-adapted ethics foundation model with both rule-based and deliberative workflows. EthicsLLM, trained on EthicsQA with 41,218 CoT–answer triples, provides normative and regulatory grounding that supports executable rule interpretation and multi-agent committee deliberation. The system operationalizes two modes: Mirror-ER for expedited, rule-based checks and Mirror-CR for committee-style deliberation across ten ethical dimensions, achieving higher quality, consistency, and professionalism than generalist LLMs. The work demonstrates scalable, privacy-preserving AI assistance for research governance, with data, models, and evaluation tools released to foster broader development in AI-assisted ethics oversight.

Abstract

Ethics review is a foundational mechanism of modern research governance, yet contemporary systems face increasing strain as ethical risks arise as structural consequences of large-scale, interdisciplinary scientific practice. The demand for consistent and defensible decisions under heterogeneous risk profiles exposes limitations in institutional review capacity rather than in the legitimacy of ethics oversight. Recent advances in large language models (LLMs) offer new opportunities to support ethics review, but their direct application remains limited by insufficient ethical reasoning capability, weak integration with regulatory structures, and strict privacy constraints on authentic review materials. In this work, we introduce Mirror, an agentic framework for AI-assisted ethical review that integrates ethical reasoning, structured rule interpretation, and multi-agent deliberation within a unified architecture. At its core is EthicsLLM, a foundational model fine-tuned on EthicsQA, a specialized dataset of 41K question-chain-of-thought-answer triples distilled from authoritative ethics and regulatory corpora. EthicsLLM provides detailed normative and regulatory understanding, enabling Mirror to operate in two complementary modes. Mirror-ER (expedited Review) automates expedited review through an executable rule base that supports efficient and transparent compliance checks for minimal-risk studies. Mirror-CR (Committee Review) simulates full-board deliberation through coordinated interactions among expert agents, an ethics secretary agent, and a principal investigator agent, producing structured, committee-level assessments across ten ethical dimensions. Empirical evaluations demonstrate that Mirror significantly improves the quality, consistency, and professionalism of ethics assessments compared with strong generalist LLMs.

Figures (4)

• Figure 1: Dataset construction and model fine-tuning pipeline for EthicsLLM. (A) EthicsCorpus construction from diverse normative sources, followed by text cleaning and normalization. (B) EthicsQA construction via open-ended QA curation and a two-phase CoT annotation process with expert-reviewed seeds and leakage filtering. (C) Supervised fine-tuning of a pre-trained language model on EthicsQA to obtain an ethics-adapted foundation model for downstream ethics review.
• Figure 2: Offline rule construction and online expedited ethics review pipeline of Mirror-ER. In the offline stage, regulatory documents are continuously collected, canonicalized, and organized into a scenario-partitioned rule graph with expert refinement. In the online stage, a submitted research dossier is analyzed through scenario and subject matching, followed by evidence retrieval and rule-level compliance checking with EthicsLLM, producing a structured compliance report.
• Figure 3: Statistical analysis of rule conditions and semantic components in expedited ethics review. (Left) Distribution of extracted rule conditions across research domains. (Right) Word cloud visualization of the most frequent terms associated with rule semantics, including regulatory subjects and regulated actions. Within each semantic category, word sizes are normalized by their relative frequency among the top-$40$ terms in that category, highlighting the internal composition and emphasis of rule semantics rather than absolute frequency differences across categories.
• Figure 4: Multi-agent committee review workflow in Mirror-CR. (01) Dimension-guided review: multiple expert agents independently assess the research proposal across ten core ethical dimensions, producing a consolidated issue list. (02) Multi-agent debate: expert agents and the PI agent engage in iterative clarification rounds to resolve raised concerns; unresolved issues persist after a bounded number of rounds. (03) Committee synthesis: an ethics secretary agent aggregates both resolved and unresolved issues into a structured final report, including key risks and recommendations.