Skip to content

Paper reproduction: Factual Self-Awareness (Tamoyan et al. 2025) #13

Description

@frankniujc

Track reproducing Factual Self-Awareness in Language Models as a canonical Murano usage example on UKP Lab's own published work.

Paper: https://arxiv.org/abs/2505.21399
Authors: Hovhannes Tamoyan (@tamohannes), Subhabrata Dutta (@Subha0009), Iryna Gurevych
Reference implementation: https://github.com/UKPLab/arxiv2025-self-awareness

Scope

  • Reproduce the linear-probe analysis on residual-stream activations for known / forgotten factual recall using the post-refactor Pipeline + Probe API.
  • Optionally reproduce the SAE comparison from the paper (depends on SAE support being ported back; tracked in Port SAE support as optional murano[sae] extra (from closed PR #16) #58).
  • Ship a runnable example under examples/use_cases/ (or equivalent) plus a reproductions entry in the docs site.

Dependencies

Ownership

Owner TBD. Any co-author with access to the external repo can pick this up. Not 0.1.0-blocking.

Metadata

Metadata

Assignees

Labels

enhancementNew feature or requestquestionFurther information is requested

Type

No type

Projects

No projects

Relationships

None yet

Development

No branches or pull requests

Issue actions