Enhancement: Consensus Schema Recommendation Across Multiple LLM Outputs

[HussainAther]

Summary

Add an optional semantic normalization layer that maps similar schema concepts across different LLM outputs into a shared intermediate representation while preserving each model’s original generated schema.

Motivation

During schema review, different LLMs may describe the same scientific process using different terminology, structure, or nesting. For example, in an iterative X-ray CT reconstruction schema, one model may represent forward_model, another may use physics_model, and another may describe the same concept under system_matrix or Beer-Lambert_attenuation.

These differences are not necessarily errors. However, they make it difficult for reviewers to compare schemas efficiently, especially when the task involves synthesizing a master schema from multiple model outputs.

A semantic normalization layer could help reviewers identify equivalent or closely related concepts across model outputs without forcing all models to produce identical schemas.

Proposed Feature

Introduce an optional normalized comparison view that:

• Preserves each model’s original raw schema output.
• Maps semantically similar concepts to a canonical or intermediate representation.
• Shows which original schema elements correspond to each normalized concept.
• Allows reviewers to inspect both the original model-specific wording and the normalized concept label.
• Supports comparison across models and refinement stages.

Example

Original model outputs may contain:

{
  "model_a": { 
      "attenuation_model": "Beer-Lambert law",
      "system_matrix": "projection operator"
    }
  },
  "model_b": {
    "physics_model": {
      "xray_attenuation": "Beer-Lambert attenuation",
      "projection_geometry": "system matrix representation"
    }
  }
}

A normalized view could map these to:

{
  "normalized_concept": "Physics-based forward model",
  "mapped_elements": [
    "model_a.forward_model",
    "model_a.forward_model.attenuation_model",
    "model_a.forward_model.system_matrix",
    "model_b.physics_model",
    "model_b.physics_model.xray_attenuation",
    "model_b.physics_model.projection_geometry"
  ]
}

Possible Implementation Approaches

This could be implemented using one or more of the following:

• Canonical concept labels.
• Embedding-based semantic similarity.
• Domain-specific synonym dictionaries.
• Optional reviewer-approved mappings.
• Side-by-side raw schema and normalized schema views.

Expected Benefit

This would reduce reviewer effort during schema comparison and synthesis by making semantic overlap easier to identify. It would also preserve the value of model diversity by keeping original outputs intact while providing a common layer for comparison.

Type

Enhancement