Research map / Large language models / Instruction and fine-tuning
Fast Exact Unlearning for In-Context Learning Data for LLMs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ERASE, which performs task adaptation via in-context example selection with quantized k-means so that exact unlearning of fine-tuning data is near-constant cost, and compares it to SISA on 15 BIG-bench tasks with LLaMA 7B.
Problem
Exact unlearning of fine-tuning data for LLMs is expensive because SGD-based fine-tuning requires retraining costs on the order of training, and existing faster exact methods such as SISA increase inference cost.
Approach
Replace SGD fine-tuning with in-context learning whose example selection uses quantized k-means over Sentence-BERT embeddings; the quantization makes removing a datapoint almost always leave the clusters unchanged, so unlearning is dataset-size independent. Also introduces a holistic unlearning cost metric that counts inferences per unlearning request.
Key results
- ERASE with 2-shot gives holistic unlearning costs about 2x better than 1-SISA across four tasks, e.g., 2.5e3 vs 4-SISA 1.4e3 inferences per request on WinoWhy.
- Unlearning cost of in-context methods is near zero (about 0 FLOPs) versus 71e12 FLOPs for 1-SISA on WinoWhy.
- ERASE matches or outperforms SISA on more than half of the 15 tasks with 3 or 4 in-context examples.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ERASE instead of ACoT for selecting in-context examples | In-context learning on four Big-Bench tasks · Four evaluated tasks | normalized aggregate score | No clear effect |
| Use ERASE instead of random in-context example selection | In-context learning · Most evaluated tasks | normalized aggregate score | Helped |
| Use ERASE instead of SISA fine-tuning | LLaMA task adaptation · 15 Big-Bench tasks | normalized aggregate score | Mixed |
| Use 2-shot ERASE instead of 4-SISA | LLaMA task adaptation with unlearning · WinoWhy, Timedial, Sports Understanding, Logical Fallacy Detection | inferences per unlearning request vs. 1-SISA: $1.4 \times 10^3$ -> $2.5 \times 10^3$; $0.2 \times 10^2$ -> $0.4 \times 10^2$; $2.7 \times 10^3$ -> $5.2 \times 10^3$; | Helped |
| Increase ERASE examples from 2-shot to 4-shot (ablation) | ERASE · WinoWhy, Timedial, Sports Understanding, Logical Fallacy Detection | inferences per unlearning request vs. 1-SISA: $2.5 \times 10^3$ -> $1.2 \times 10^3$; $0.4 \times 10^2$ -> $0.2 \times 10^2$; $5.2 \times 10^3$ -> $2.6 \times 10^3$; | Hurt |
| Increase ERASE examples from 2-shot to 3-shot (ablation) | ERASE · WinoWhy, Timedial, Sports Understanding, Logical Fallacy Detection | inferences per unlearning request vs. 1-SISA: $2.5 \times 10^3$ -> $1.7 \times 10^3$; $0.4 \times 10^2$ -> $0.3 \times 10^2$; $5.2 \times 10^3$ -> $3.4 \times 10^3$; | Mixed |
| Reduce clustering embedding dimensions with UMAP (ablation) | ERASE and ACoT · Four Big-Bench tasks | normalized aggregate score | No clear effect |
| Use quantized k-means instead of k-means++ for example clustering (ablation) | ERASE compared with ACoT · Four Big-Bench tasks | normalized aggregate score | No clear effect |
| Increase the number of SISA shards (ablation) | SISA · Big-Bench tasks | normalized aggregate score | Hurt |
| Use 10 warm-up steps instead of 15 or 40 (ablation) | LLaMA fine-tuning · Intent Recognition | test perplexity | Helped |
| Use learning rate 1e-5 selected from the tested rates (ablation) | LLaMA fine-tuning · Intent Recognition | test perplexity | Helped |
Surprising findings
- 4-shot ERASE is more expensive than 4-SISA under the holistic unlearning cost metric despite its cheap unlearning operations, because of higher inference cost.
- Fine-tuning SISA converges after about 80 examples per model, so slicing was unnecessary in the setup.
Limitations
- In-context learning is limited by context size and does not modify weights, so it is unsuitable for learning from large data.
- Holistic cost advantage depends on task; the trade-off may not generalize where more in-context examples are needed.
- Quantization parameter epsilon must be tuned per task; fixed epsilon=0.05 was used.
Experimental setup
Datasets: Big-Bench Instruction Induction Big-Bench
Benchmarks: Big-Bench Instruction Induction Big-Bench
Metrics: normalized aggregate score holistic unlearning cost (inferences per unlearning request) FLOPs
Baselines: SISA (1-, 2-, 4-shard) Auto Chain-of-Thought random in-context sampling UMAP + ACoT UMAP + ERASE
Builds on: SISA Auto Chain-of-Thought quantized k-means
Closest papers by meaning
- In-Context Unlearning: Language Models as Few-Shot Unlearners (ICML 2024 · similarity 0.92)
- Towards Robust and Parameter-Efficient Knowledge Unlearning for LLMs (ICLR 2025 · similarity 0.91)
- Towards Scalable Exact Machine Unlearning Using Parameter-Efficient Fine-Tuning (ICLR 2025 · similarity 0.90)
- Invariance Makes LLM Unlearning Resilient Even to Unanticipated Downstream Fine-Tuning (ICML 2025 · similarity 0.90)
- LLM Unlearning via Loss Adjustment with Only Forget Data (ICLR 2025 · similarity 0.90)
- Catastrophic Failure of LLM Unlearning via Quantization (ICLR 2025 · similarity 0.90)
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.