Beyond the Unlearning Mirage

Most LLM unlearning demos crumble when users rephrase queries, add alias chains, or build multi-hop reasoning paths. “The Unlearning Mirage” makes that brittleness visible by pairing dynamic, model-specific probes with activation-pathway analysis.

What the framework delivers

1. Probe generation that spans single-hop recalls to adversarial multi-hop chains and aliases, all grounded in knowledge elicited from the target model before unlearning.
2. Activation analysis showing that single-hop queries usually travel dominant pathways that the unlearning method disrupts, while multi-hop queries find alternate routes that often remain intact.
3. Automatic alias/paraphrase sweeps that stress-test whether a supposedly forgotten fact still leaks.

Complementary defenses

• Text watermarking taxonomy (NAACL 2025) clarifies how to prove which model family produced a generation, even after editing or fine-tuning.
• Power-law continual learning (ECAI 2024) keeps models plastic while mitigating catastrophic forgetting, offering a softer alternative than hard erasure.

What practitioners get

• Pip-installable evaluators and leaderboards so every unlearning method can be stress-tested before deployment.
• Roadmaps for pairing unlearning with provenance controls and continual learning recipes.
• Visualizations that expose exactly which prompts still succeed post-unlearning.

Next steps: release activation-visualization tooling, extend to multimodal models, and maintain a living playbook of best practices. If you operate production LLMs, this stack is built for you.