ArgRE: Formal Argumentation for Conflict Resolution in Multi-Agent Requirements Negotiation

Abstract: As software systems grow in complexity, they must satisfy an increasing number of competing quality attributes, making it essential to balance them in a principled manner -- for example, a safety requirement for sensor-fusion verification may conflict with a tight planning-cycle budget. Multi-agent LLM frameworks support this balancing process by assigning specialized agents to different objectives. However, their conflict resolution is typically heuristic. Requirements are aggregated implicitly without explicit acceptance or rejection, limiting auditability in regulated domains. We present ArgRE, a multi-agent requirements negotiation system that embeds Dung-style abstract argumentation into the negotiation stage. Each proposal, critique, and refinement is modeled as an argument, conflicts are represented as directed attack relations, and the accepted set of arguments is computed under grounded and preferred semantics. The pipeline further integrates KAOS goal modeling, multi-layer verification, and standards-oriented artifact generation. Evaluation across five case studies spanning safety-critical, financial, and information-system domains shows that ArgRE provides argument-level traceability absent from existing frameworks. Independent evaluators rated its decision justifications significantly higher than those of heuristic synthesis (4.32 vs. 3.07, p < 0.001), indicating improved auditability, while semantic intent preservation remains comparable (94.9% BERTScore F1) and compliance coverage reaches 84.7% versus 47.6%--47.8% for baselines. Structural analysis further confirms that the default pairwise protocol yields acyclic graphs in which grounded and preferred semantics coincide, whereas cross-pair arbitration introduces controlled cyclicity, leading to predictable divergence between the two semantics.

• The paper introduces a formal argumentation layer replacing heuristics in multi-agent requirements negotiation to provide traceable conflict resolution.
• The methodology leverages Dung’s argumentation frameworks with grounded and preferred semantics to transparently justify acceptance decisions.
• Empirical evaluation shows significant gains in provenance traceability and decision justification with modest computational overhead.

Formal Argumentation for Conflict Resolution in Multi-Agent Requirements Negotiation

Motivation and Background

The complexity of software systems mandates the principled balancing of competing quality requirements. Requirements Engineering (RE), especially in regulated or safety-critical domains, is fraught with conflict, ambiguity, and incomplete traceability, contributing to significant project failure rates. Multi-agent frameworks driven by LLMs have enabled role-based automation for requirements negotiation, but conflict resolution remains largely heuristic, lacking explicit decision rationales and auditability required by standards such as ISO 26262 and the EU AI Act.

The prevailing approaches—implicit aggregation, moderator-mediated synthesis, and priority weighting—are insufficient for contexts where every requirement decision must be defended with explicit provenance. Optimization techniques, such as Pareto front analysis, fail to capture the dialectical structure of trade-offs and omit rationale tracing. Formal argumentation, specifically Dung's abstract argumentation frameworks (AFs), offers the mathematical grounding needed to justify acceptance or rejection through attack and defense semantics.

ArgRE System Overview

ArgRE introduces a multi-agent requirements negotiation system integrating formal argumentation within the negotiation stage. It assigns quality-specialized agents (Safety, Efficiency, Green, Trustworthiness, and Responsibility) to generate candidate requirements. Conflicts and critiques are modeled as arguments; directed attack relations capture dialectical opposition; accepted requirements are computed under grounded and preferred semantics.

The negotiation pipeline consists of five phases:

1. Parallel Generation: Quality-specialized agents independently produce candidate requirements, maximizing diversity and coverage.
2. Dialectical Negotiation with Argumentation: Conflicts are surfaced and resolved under formal argumentation, with explicit provenance traces.
3. KAOS Integration: Accepted requirements are structured into a validated hierarchical goal model.
4. Verification: Multi-layer checks for logical validity, hallucination detection, and standards compliance.
5. Artifact Generation: Structured deliverables (requirements documentation, test cases, compliance evidence).

The primary innovation resides in Phase 2, replacing heuristic synthesis with Dung-style argumentation semantics to provide transparent, traceable conflict resolution.

Argumentation Layer and Semantics

ArgRE leverages deterministic attack patterns—critique-to-proposal, refinement-to-original, refinement-to-critique—supported by a threshold-gated LLM pathway to detect cross-session conflicts. Structured negotiation logs are parsed into arguments; attack graphs are constructed; extensions (accepted arguments) are computed via grounded and preferred semantics.

• Grounded semantics yields the minimal, conservative accepted set—unattacked or indisputably defended requirements.
• Preferred semantics admits maximal admissible sets, allowing for credulous acceptance and prioritization based on project weights.

For acyclic attack-graphs arising from pairwise negotiation protocols, grounded and preferred semantics coincide, offering simplicity. When cross-pair arbitration introduces cyclicity (Graph Cyclicity Index GCI > 0), preferred semantics allow explicit resolution according to quality priorities.

Empirical Evaluation

Across five case studies—including safety-critical and financial domains—ArgRE exhibits:

• Provenance Trace Completeness (TC): Argument-level traceability (average TC = 40.9%) absent from prior frameworks.
• Decision Justification Scores (DJS): Independent raters scored ArgRE's justification quality significantly higher than heuristic synthesis (4.32 vs. 3.07, p < 0.001, Cliff’s d = 0.92).
• Semantic Preservation: BERTScore F1 of 94.9%, matching heuristic baselines and indicating no semantic drift.
• Compliance Coverage: 84.7%, substantially outperforming baselines (47.6%-47.8%), though slightly below heuristic variant (98.2%) due to stricter acceptance criteria.

Structural analysis confirmed that pairwise protocols yield acyclic attack graphs (GCI=0), resulting in semantics equivalence; cross-pair arbitration induces controlled cyclicity (GCI≈0.25), leading to divergence in extensions and compliance coverage.

Implications for Theory and Practice

Practical Implications:

• ArgRE directly supports regulatory audit requirements (ISO 26262, EU AI Act) by providing transparent, inspectable trails from accepted requirements to their origin.
• The choice between grounded and preferred semantics becomes a formal lever for conservativeness versus coverage, operational only when cyclic attack relations exist.
• Computational overhead is modest (~2.5x over heuristic synthesis), justified by auditability gains.
• Human-in-the-loop review is enabled at the argumentation layer, supporting inspection, override, and injection with rapid recomputation of accepted sets.

Theoretical Implications:

• Formal argumentation frameworks operationalize dialectical negotiation, closing the "Formalization Gap" between practical RE synthesis and mathematically justified acceptance.
• Trace-based interpretability is maximized without loss of negotiated intent, as evidenced by matched semantic preservation metrics.
• The argumentation graph provides new structural metrics (TC, DJS, GCI) that quantify auditability and rationale quality for requirements decisions.

Future Directions:

• Extending the framework with value-based or weighted AFs to capture rationale strength and conflict intensity explicitly.
• Improving argument extraction recall using advanced prompt engineering or fine-tuned LLMs.
• Scaling to larger multi-agent settings, investigating protocol extensions and threshold tuning via factorial designs, and replicating human justification studies across broader domains.

Conclusion

ArgRE bridges the gap between automated multi-agent requirements negotiation and the stringent auditability requirements of regulated engineering domains by embedding Dung-style formal argumentation semantics. Empirical results demonstrate significant gains in interpretability, rationale quality, and compliance coverage, with modest computational costs. Structural analysis confirms predictable semantics behavior, contingent on protocol-induced cyclicity. ArgRE establishes a new baseline for principled conflict resolution in RE, enabling traceable justification for every accepted requirement and supporting flexible configurations for practical deployment and regulatory compliance (2604.23124).