VM

Selected research

← All research

Paper explained

Workshop presentationMachine unlearning

Between Suppression and Collapse: Evaluating Narrative Unlearning with LENS

Viktoriia Makovska, George Fletcher

When a model stops repeating a narrative, has it forgotten—or stopped answering usefully?

The question

Why this work

LENS examines narrative reproduction after machine unlearning. The study tests four multilingual instruction models on two disinformation-aligned frames, using several ways of asking about the same underlying narrative.

The approach

A visual guide

One narrative, four prompting contexts

  1. L0 · Direct

    Ask for the target explanation directly.

  2. L1 · Attributed

    Ask about an explanation attributed to someone else.

  3. L2 · Contrastive

    Place the target frame beside a competing explanation.

  4. L3 · Abstract

    Mask actor names while preserving the causal structure.

Conceptual guide to LENS. L0 supplies the forget examples; L1–L2 test transfer; L3 probes abstraction and entity recovery. These are prompting levels, not numerical difficulty scores. Sections 4, 6–8

Illustrative calculator · 100 responses

Explore the suppression score

Try changing the counts. These are hypothetical responses, not experimental results. The score rewards reduced reproduction and penalizes degraded answers.

Relative suppression 75%

Degradation rate 10%

SCE score 0.608

SCE = suppression × (1 − degradation)²

Counts are constrained to 100 responses. SCE is a checkpoint-selection aid, not proof of complete forgetting.

What the work contributes

Selected checkpoints reduced narrative reproduction beyond direct training prompts while retaining substantive answers. The useful checkpoint varied by model, language, narrative, and method. Models could still name associated real-world actors when prompted abstractly.

Sections 4, 6–8

Original paper & resources

Presented at WIPE-OUT 2 — 2nd Workshop on Machine Unlearning and Privacy Preservation, ECML PKDD 2026 · 7 September 2026 · Naples, Italy.

Preprint: arXiv:2607.22657 · revised 31 July 2026.

Slides (PDF, English) Workshop programme arXiv PDF
Citation
Makovska, V., & Fletcher, G. (2026). Between Suppression and Collapse: Evaluating Narrative Unlearning with LENS. arXiv:2607.22657. https://arxiv.org/abs/2607.22657
BibTeX
@misc{makovska2026lens,
  title = {Between Suppression and Collapse: Evaluating Narrative Unlearning with LENS},
  author = {Makovska, Viktoriia and Fletcher, George},
  year = {2026},
  eprint = {2607.22657},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url = {https://arxiv.org/abs/2607.22657}
}