Skip to content

H-L-DAG: Hierarchical Argument Generation

H-L-DAG (Hierarchical Leveled Directed Acyclic Graph) is ARTEMIS's approach to structured argument generation. Instead of generating flat text responses, agents construct arguments at multiple levels of abstraction.

The Three Levels

Strategic Level

The highest level of abstraction, focusing on:

  • Core thesis: The fundamental position being argued
  • Strategic goals: What the argument aims to achieve
  • Key differentiators: What makes this position compelling
# Example strategic-level argument
Argument(
    agent="Proponent",
    level=ArgumentLevel.STRATEGIC,
    content="""
    AI regulation is essential for societal protection. The core thesis
    is that unregulated AI development poses existential risks that
    outweigh the costs of oversight. This position prioritizes long-term
    safety over short-term innovation speed.
    """,
    evidence=[],  # Strategic level focuses on framing
)

Tactical Level

The middle level, providing:

  • Supporting points: Specific arguments backing the thesis
  • Evidence chains: Logical connections between claims
  • Counter-argument awareness: Acknowledgment of opposing views
# Example tactical-level argument
Argument(
    agent="Proponent",
    level=ArgumentLevel.TACTICAL,
    content="""
    Three key factors support AI regulation:
    1. Historical precedent shows technology regulation benefits society
    2. Current AI capabilities already exceed human oversight capacity
    3. Industry self-regulation has proven insufficient
    """,
    evidence=[
        Evidence(
            type="example",
            content="FDA drug approval process demonstrates successful tech regulation",
            source="FDA History",
        ),
        Evidence(
            type="study",
            content="Analysis of 200+ AI incidents from 2020-2024",
            source="AI Incident Database",
        ),
    ],
)

Operational Level

The ground level with:

  • Specific facts: Concrete data points
  • Quotes and citations: Authoritative sources
  • Examples: Real-world cases
# Example operational-level argument
Argument(
    agent="Proponent",
    level=ArgumentLevel.OPERATIONAL,
    content="""
    The EU AI Act, implemented in 2024, demonstrates successful regulation.
    According to the European Commission, compliance costs averaged only
    2.3% of development budgets while preventing an estimated 47 high-risk
    AI deployments. Microsoft's CEO stated, "Thoughtful regulation actually
    accelerates responsible innovation."
    """,
    evidence=[
        Evidence(
            type="quote",
            content="compliance costs averaged 2.3% of development budgets",
            source="European Commission AI Act Report 2024",
        ),
    ],
)

How H-L-DAG Works

Argument Generation Flow

graph TD
    A[Topic + Position] --> B[Strategic Planning]
    B --> C[Generate Strategic Argument]
    C --> D[Identify Tactical Points]
    D --> E[Generate Tactical Arguments]
    E --> F[Gather Operational Evidence]
    F --> G[Generate Operational Arguments]
    G --> H[Assemble Complete Argument]

The DAG Structure

Arguments form a directed acyclic graph:

Strategic Thesis
├── Tactical Point 1
│   ├── Operational Fact 1.1
│   └── Operational Fact 1.2
├── Tactical Point 2
│   ├── Operational Fact 2.1
│   ├── Operational Fact 2.2
│   └── Operational Fact 2.3
└── Tactical Point 3
    └── Operational Fact 3.1

Evidence Types

The Evidence class supports these types:

Type Description
fact Verifiable factual statement
statistic Numerical data or statistics
quote Direct quotation from a source
example Real-world case or example
study Research study or analysis
expert_opinion Opinion from domain expert

H-L-DAG tracks causal relationships between arguments:

from artemis.core.types import CausalLink

link = CausalLink(
    cause="unregulated_ai",
    effect="societal_harm",
    mechanism="Lack of oversight allows harmful deployments",
    strength=0.8,  # 0.0 to 1.0
    bidirectional=False,
)
Field Type Description
cause str The cause in the relationship
effect str The effect in the relationship
mechanism str How cause leads to effect
strength float Relationship strength (0-1)
bidirectional bool Whether relationship works both ways

Using H-L-DAG

Configuring Arguments

from artemis.core.types import DebateConfig

config = DebateConfig(
    require_evidence=True,
    require_causal_links=True,
    min_evidence_per_argument=1,
    max_argument_tokens=1000,
)

Accessing Argument Structure

result = await debate.run()

for turn in result.transcript:
    arg = turn.argument

    print(f"Level: {arg.level}")
    print(f"Content: {arg.content}")

    for evidence in arg.evidence:
        print(f"  Evidence: {evidence.source}")

    for link in arg.causal_links:
        print(f"  Causal: {link.cause} -> {link.effect}")

Benefits of H-L-DAG

1. Structured Reasoning

Arguments are systematically organized, making them easier to:

  • Evaluate for logical coherence
  • Compare across agents
  • Identify gaps or weaknesses

2. Evidence Integration

The operational level ensures arguments are grounded in:

  • Concrete facts
  • Authoritative sources
  • Real-world examples

3. Causal Transparency

Explicit causal links make reasoning transparent:

  • Evaluators can verify logical connections
  • Counter-arguments can target specific links
  • Fallacies are easier to detect

4. Hierarchical Evaluation

Different levels can be weighted differently:

evaluation_weights = {
    "strategic": 0.3,   # 30% weight
    "tactical": 0.4,    # 40% weight
    "operational": 0.3, # 30% weight
}

Comparison to Flat Arguments

Aspect Flat Arguments H-L-DAG
Structure Unorganized text Three-level hierarchy
Evidence Mixed in text Explicit evidence objects
Causality Implicit Explicit causal links
Evaluation Holistic only Level-by-level possible
Debugging Difficult Clear structure

Next Steps