Research map / Large language models / Retrieval and knowledge
Do LLMs Forget What They Should? Evaluating In-Context Forgetting in Large Language Models
Yuli Qian, Zechuan Yang, Wenbiao Ding, Hongzhi Li, Yutao Xie
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 explicit forgetting reminder to prompt (ablation) | GPT-O3Mini · Instructional Forgetting | SFRR: 62.88 -> 76.30 | Improved |
| Add explicit forgetting reminder to prompt (ablation) | GPT-4o · Instructional Forgetting | SFRR: 52.52 -> 71.76 | Improved |
| Add explicit forgetting reminder to prompt (ablation) | GPT-5 · Instructional Forgetting | SFRR: 58.80 -> 74.62 | Improved |
| Add explicit forgetting reminder to prompt (ablation) | GPT-4o · Dynamic Preference | SFRR: 50.07 -> 66.56 | Improved |
| Add explicit forgetting reminder to prompt (ablation) | GPT-5 · Dynamic Preference | SFRR: 79.12 -> 76.64 | Hurt |
| Add explicit forgetting reminder to prompt (ablation) | GPT-O3Mini · Instructional Forgetting | NoForget Accuracy: 96.48 -> 93.84 | Hurt |
| Add explicit forgetting reminder to prompt (ablation) | GPT-4o · Dynamic Preference | NoForget Accuracy: 86.10 -> 87.43 | Improved |
| Lengthen context from 0.5k to 30k tokens (ablation) | GPT-4o · Instructional Forgetting | NoForget Accuracy: 96.53% -> 77.89% | Hurt |
| Lengthen context from 0.5k to 30k tokens (ablation) | GPT-4o · Dynamic Preference | NoForget Accuracy: 90.36% -> 59.42% | Hurt |
| Lengthen context from 0.5k to 30k tokens (ablation) | GPT-4o · Dynamic Preference | SFRR: 52.46% -> 22.87% | Hurt |
| Lengthen context from 0.5k to 30k tokens (ablation) | GPT-4o · Subtask Revision | SFRR: 71.76% -> 81.50% | Improved |
| Scale model size (Mistral-7B to GPT-5) on NoForget Accuracy | LLMs · ICF-Bench average | NoForget Accuracy: 60.18% -> 88.75% | Improved |
Closest papers by meaning
- How Much Can We Forget about Data Contamination? (ICML 2025 · similarity 0.88)
- Demystifying Language Model Forgetting with Low-rank Example Associations (NeurIPS 2025 · similarity 0.87)
- Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference (NeurIPS 2024 · similarity 0.87)
- Interweaving Memories of a Siamese Large Language Model (AAAI 2025 · similarity 0.87)
- Simple linear attention language models balance the recall-throughput tradeoff (ICML 2024 · similarity 0.87)
- Stress-Testing Long-Context Language Models with Lifelong ICL and Task Haystack (NeurIPS 2024 · 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.
