Research map / Architectures and efficiency / RNNs and normalization
AI Engram: In Search of Memory Traces in Artificial Intelligence
Jea Kwon, Dong-Kyum Kim, Jiwon Kim, Yonghyun Kim, Woong Kook, Meeyoung Cha
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 |
|---|---|---|---|
| add Engram (alpha=1) to unlearn class 0 | ResNet-18 · CIFAR-10 class-0 unlearning | DA: 0.987 (retrain) vs 0.992 | Improved |
| use alpha best instead of alpha=1 (ablation) | Engram, CIFAR-10 ResNet-18 · CIFAR-10 class-0 unlearning | ToW: 0.930 -> 0.984 | Improved |
| use alpha best instead of alpha=1 (ablation) | Engram, CIFAR-10 ResNet-18 · CIFAR-10 class-0 unlearning | NMI: 0.379 -> 0.611 | Hurt |
| use Engram on CIFAR-100 | ResNet-50 · CIFAR-100 class-0 unlearning | ToW: 0.988 (alpha=1) vs retrain 0.985 | Improved |
| use adaptive W-Norm alpha instead of uniform alpha=0.6 (ablation) | Engram, Llama-3.2-1B · TOFU forget10 | memorization: 0.9176 -> 0.9627 | Improved |
| use adaptive W-Norm alpha instead of uniform alpha=0.6 (ablation) | Engram, Llama-3.2-1B · TOFU forget10 | utility: 0.8801 -> 0.9256 | Improved |
| use Engram instead of UCE | Llama-3.2-1B · TOFU forget10 | Overall: 0.659 -> 0.818 | Improved |
| use Engram instead of iterative methods (AltPO, NPO, SimNPO) | Llama-3.2-1B · TOFU forget10 | Overall / Privacy | Hurt |
| use Engram instead of NPO for exact memorization | Llama-3.2-1B · TOFU forget10 | exact memorization: 0.0948 -> 0.0276 | Improved |
| use tabula rasa W_ft instead of naive delta W_ft - W_pt (ablation) | Engram, Llama-3.2-1B · TOFU forget10 | Overall: 0.446 -> 0.818 | Improved |
| use tabula rasa W_ft instead of naive delta (ablation) | Engram, Llama-3.2-1B · TOFU forget10 | exact memorization: 0.617 -> 0.028 | Improved |
| restrict extraction to Q/K + Gate layers (ablation) | Engram, Llama-3.2-1B · TOFU forget10 | Overall | No clear effect |
Closest papers by meaning
- Frozen Memory Is Not Enough: Rethinking External Memory as Extraction (NeurIPS 2026 · similarity 0.84)
- Scaling Laws for Associative Memories (ICLR 2024 · similarity 0.83)
- Disentangling Representations through Multi-task Learning (ICLR 2025 · similarity 0.83)
- Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents (ICML 2026 · similarity 0.83)
- Factual recall in linear associative memories: sharp asymptotics and mechanistic insights (NeurIPS 2026 · similarity 0.82)
- The Geometric Inductive Bias of Grokking: Bypassing Phase Transitions via Architectural Topology (NeurIPS 2026 · similarity 0.82)
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.
