Skip to content

Enable Eigenvalue, Degree, and Eccentricity Confidence Scores (NLI-based) #339

@dylanbouchard

Description

@dylanbouchard

Is your feature request related to a problem? Please describe.
In the TMLR publication by Lin et al., 2024, the authors propose using three NLI-based confidence scores. Given a matrix of NLI-estimated entailment scores over sampled responses to the same prompt, they compute the following:

  • Sum of Eigenvalues of the Graph Laplacian
  • Trace of the Degree matrix
  • Eccentricity

Describe the solution you'd like
Add these scoring methods to BlackBoxUQ. To enable this, they should implemented via uqlm.black_box.ConsistencyScorer. Note that computing the full matrix of NLI scores over sampled responses requires more NLI comparisons than our existing NLI-based scorers (semantic_negentropy, semantic_sets_confidence, entailment, noncontradiction). With this in mind, we should be careful to avoid redundant NLI computations if multiple NLI-based scorers are requested.

Additional context
Please ensure you have carefully read and fully understand the paper before volunteering to work on this issue.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requesthelp wantedExtra attention is needed

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions