Research map / Large language models / Language, safety and interpretability
Guess or Recall? Training CNNs to Classify and Localize Memorization in LLMs
Jérémie Dentan, Davide Buscaldi, Sonia Vanier
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the prior 4-class taxonomy with the 3-class Non-Memo/Guess/Recall taxonomy (ablation) | CNN classifier on Pythia attention weights · taxonomy benchmark | minimum F1 (%): 64.7 -> 89.0 | Improved |
| use best 4-class taxonomy instead of prior taxonomy (ablation) | CNN classifier on Pythia attention weights · taxonomy benchmark | minimum F1 (%): 64.7 -> 72.8 | Improved |
| add a Guess node (ROUGE-based) instead of the strict Reconstruct node (ablation) | 3-class taxonomy · taxonomy benchmark | minimum F1 (%): 87.7 -> 89.0 | Improved |
| add duplication-based classes to the taxonomy (ablation) | 3-class taxonomy · taxonomy benchmark | minimum F1 (%): 89.0 -> 80.8 | Hurt |
| split off a few-shot Recollect class (ablation) | prior taxonomy · taxonomy benchmark | F1 of Others class (%): F1 64.7 | Hurt |
| tune ROUGE thresholds lambda and gamma beyond 0.5 (ablation) | Guess class · taxonomy benchmark | minimum F1 (%): 88.8 -> 89.2 | No clear effect |
| lower ROUGE thresholds to 0.1 (ablation) | Guess class · taxonomy benchmark | minimum F1 (%): 88.8 -> 74.4 | Hurt |
| evaluate on each model size separately (12B, 6.9B, 2.8B) (ablation) | Guess/Recall taxonomy · taxonomy benchmark | taxonomy rank | No clear effect |
Closest papers by meaning
- Mitigating Memorization in Language Models (ICLR 2025 · similarity 0.89)
- Memorization Sinks: Isolating Memorization during LLM Training (ICML 2025 · similarity 0.88)
- Recite, Reconstruct, Recollect: Memorization in LMs as a Multifaceted Phenomenon (ICLR 2025 · similarity 0.87)
- Rethinking LLM Memorization through the Lens of Adversarial Compression (NeurIPS 2024 · similarity 0.87)
- Generalization v.s. Memorization: Tracing Language Models’ Capabilities Back to Pretraining Data (ICLR 2025 · similarity 0.87)
- Learning the Signature of Memorization in Autoregressive Language Models (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
