QUARE: Multi-Agent Negotiation for Balancing Quality Attributes in Requirements Engineering

1. Problem Statement

Requirements Engineering (RE) is critical for software success but remains difficult to automate due to the need to balance multiple, often conflicting, quality attributes (e.g., safety, efficiency, sustainability) while preserving stakeholder intent.

• The Monolithic Integration Gap: Existing Large Language Model (LLM) approaches typically rely on monolithic reasoning or implicit aggregation. They tend to collapse diverse stakeholder viewpoints into a single narrative, obscuring underlying trade-offs and failing to systematically surface cross-quality conflicts.
• Limitations of Current Multi-Agent Systems: While frameworks like MARE and iReDev exist, they focus on task decomposition (e.g., elicitation, modeling) rather than quality-dimension decomposition. They lack explicit negotiation protocols to resolve conflicts between competing objectives (e.g., Safety vs. Efficiency), often defaulting to simple priority overrides or human intervention.
• Lack of Verification: Most approaches lack mechanisms to verify generated requirements against industry standards (e.g., ISO 26262) or ensure structural validity (e.g., KAOS goal models).

2. Methodology: The QUARE Framework

QUARE (QUality-Aware Requirements Engineering) is a multi-agent framework that formulates requirements analysis as a structured, dialectical negotiation among agents specialized in specific quality dimensions.

A. Multi-Agent Architecture

The system utilizes six distinct roles based on the ISO/IEC 25010 quality model:

1. Five Quality-Specialized Agents:
   • Safety: Hazard identification and mitigation.
   • Efficiency: Resource optimization and latency.
   • Green (Sustainability): Energy efficiency and carbon footprint.
   • Trustworthiness: Security and privacy.
   • Responsibility: Ethical and regulatory compliance.
2. Orchestrator Agent: Coordinates the workflow and integrates outputs.

All agents use a shared LLM backbone (GPT-4o-mini) but are functionally isolated via dedicated system prompts and Chain-of-Thought (CoT) instructions.

B. The Five-Phase Pipeline

1. Parallel Generation: Agents independently generate candidate requirements based on their specific quality focus.
2. Dialectical Negotiation: A Thesis-Antithesis-Synthesis protocol is employed:
   • Thesis: An agent proposes a requirement.
   • Antithesis: Peer agents critique the proposal, citing constraint violations (e.g., resource conflicts or logical incompatibilities).
   • Synthesis: A moderator synthesizes a refined proposal that resolves the conflict (e.g., via architectural decomposition) rather than simple priority overriding.
   • Conflict Detection: Uses embedding similarity (BERT) to detect overlaps and LLM classification to categorize conflicts as resource-bound (competing for finite resources) or logical incompatibilities (mutually exclusive states).
3. Integration & Topology Validation: Fragmented requirements are synthesized into a coherent KAOS (Knowledge Acquisition in Automated Specification) goal model. This phase ensures the model is a Directed Acyclic Graph (DAG) with no circular reasoning.
4. Verification: A Retrieval-Augmented Generation (RAG) module checks requirements against industry standards (e.g., ISO 26262, ISO 27001) to detect hallucinations and ensure compliance.
5. Standardized Output: Generation of formal KAOS models (JSON/GSN XML) and documentation.

C. Evaluation Metrics

The authors introduced a novel geometry-based evaluation methodology to measure requirement diversity and balance:

• Convex Hull Volume (CHV): Measures the breadth of coverage in a 5D quality space.
• Mean Distance to Centroid (MDC): Measures the dispersion of requirements.
• Coverage Uniformity (CU) & Minimum Axis Coverage (MAC): Metrics to ensure no single quality dimension is neglected.
• Semantic Preservation: Measured via BERTScore and optimal bipartite matching between pipeline phases.

3. Key Contributions

1. Negotiation-Based Architecture: A novel multi-agent framework that decomposes RE by quality dimension rather than task, enabling explicit conflict resolution across competing attributes.
2. Dialectical Negotiation Protocol: A structured, multi-round dialogue mechanism (Thesis-Antithesis-Synthesis) that resolves conflicts through architectural restructuring rather than simple suppression of constraints.
3. Automated KAOS Modeling & Verification: An end-to-end pipeline that transforms unstructured text into structurally valid KAOS goal models and verifies them against regulatory standards using RAG.
4. Open-Source Benchmark (OpenReBench): A faithful reimplementation of existing frameworks (MARE, iReDev) and all experimental artifacts, addressing the lack of public code in the field.
5. Geometric Evaluation Methodology: New metrics (CHV, MDC, CU, MAC) to quantitatively assess the diversity and balance of generated requirements.

4. Experimental Results

The framework was evaluated on five case studies (including an industrial autonomous driving specification) against baselines: Single-Agent, MARE, and iReDev.

• Coverage and Diversity (RQ1):

  • QUARE generated 25–43% more requirements than existing multi-agent frameworks.
  • It achieved the highest dimensional balance (lowest Coverage Uniformity score of 0.20) and the strongest worst-case coverage (Minimum Axis Coverage of 6.7), ensuring no quality dimension was neglected.
  • While iReDev had the highest geometric spread (CHV), QUARE's focused approach yielded better compliance.

• Negotiation Effectiveness (RQ2):

  • QUARE achieved 94.9% semantic preservation (Phase 1 to Phase 3), outperforming MARE (+5.9 pp) and iReDev (+2.3 pp).
  • The negotiation protocol successfully resolved conflicts by creating hybrid solutions (e.g., splitting a latency requirement into a "fast-path" for efficiency and an "async-path" for safety) rather than discarding constraints.

• Structural Validity and Compliance (RQ3):

  • Compliance Coverage: QUARE achieved 98.2% compliance with industry standards (ISO 26262, ISO 27001), a 105% improvement over baselines (~47%).
  • Verifiability: QUARE scored 4.96/5.0 on verifiability, significantly higher than baselines, due to the RAG-augmented verification phase.
  • All frameworks produced structurally valid DAGs, but only QUARE ensured deep regulatory alignment.

5. Significance and Conclusion

• Architectural Decomposition > Model Scale: The study demonstrates that effective RE automation depends more on principled architectural decomposition and explicit interaction protocols than on the sheer scale of the underlying LLM.
• Engineering-Ready Outputs: QUARE bridges the gap between LLM generation and industrial adoption by producing requirements that are not only semantically coherent but also structurally valid and compliant with safety standards.
• Traceability: The negotiation trace provides an auditable record of trade-off decisions, crucial for safety-critical domains and emerging regulations like the EU AI Act.
• Future Impact: The framework suggests a shift from "black box" generation to "white box" negotiation, where the rationale for every requirement is explicitly derived from conflicting quality constraints.

In summary, QUARE represents a significant advancement in automating Requirements Engineering by treating it as a multi-objective optimization problem solved through structured, dialectical negotiation, resulting in higher quality, more diverse, and regulation-compliant software specifications.