Research map / Large language models / Retrieval and knowledge
Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness
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 a knowledge-graph-based evaluation of LLM unlearning that extracts confidence-aware supporting subgraphs around target triples and uses an LLM judge (GPT-o4-mini) to score residual inferability, showing standard instance-level evaluation overestimates unlearning.
Problem
Existing unlearning evaluations check only direct removal of isolated facts, ignoring correlated facts that let the forgotten fact be inferred, and ignoring that LLM knowledge is uncertain/non-deterministic.
Approach
Model an LLM's knowledge as a confidence-aware KG (entropy from calibrated Yes/No/Unknown probabilities), extract a bounded-path supporting subgraph around the target triple using YAGO3-10 as reference, and have a strong LLM judge score how inferable the target is from that subgraph (UES metric).
Key results
- On LLaMA-8B-Instruct with sentence templates, GA (Full) achieves UES 0.157 under instance unlearning but 0.076 under the subgraph-based evaluation (about 52% lower).
- Under the subgraph setting, UES typically drops by at least 20% relative to instance unlearning, with over half of settings dropping more than 30%.
- LLM judge scores correlate strongly with averaged human ratings from PhD students on 50 target triples; temperature scaling reduces ECE of LLaMA3-8B-Instruct to 0.031 and Qwen2.5-7B-Instruct to 0.067.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Evaluate unlearning with correlated supporting subgraphs instead of direct target triples alone | Unlearning methods in the evaluated settings · LLaMA3-8B-Instruct and Qwen2.5-7B-Instruct | UES: For most cases, the UES decreases by at least 20%; more than half of the settings exhibit a reduction exceeding 30% | Hurt |
| Use QA templates rather than sentence templates for unlearning | Unlearning methods · LLaMA3-8B-Instruct and Qwen2.5-7B-Instruct | Loc | Hurt |
| Use QA templates rather than sentence templates for unlearning | Unlearning methods · LLaMA3-8B-Instruct and Qwen2.5-7B-Instruct | UES (Ours) | Mixed |
| Require higher model confidence by decreasing the entropy threshold u* (ablation) | Supporting-subgraph evaluation · LLaMA3-8B-Instruct, LoRA QA | UES | Helped |
| Increase the forget-set size from 50 to 100 (ablation) | Unlearning methods · LLaMA3-8B-Instruct, LoRA QA | UES gap between instance and supporting-subgraph evaluation | Mixed |
| Increase the number of unlearning epochs (ablation) | Unlearning methods · LLaMA3-8B-Instruct, LoRA sentence templates | UES gap between instance and supporting-subgraph evaluation | Helped |
| Apply temperature scaling to target-model confidence probabilities | LLaMA3-8B-Instruct · Calibration validation set | ECE: 0.475 -> 0.031 | Helped |
| Apply temperature scaling to target-model confidence probabilities | Qwen2.5-7B-Instruct · Calibration validation set | ECE: 0.083 -> 0.067 | Helped |
| Include an Unknown option in multiple-choice knowledge queries | Target LLM knowledge probing · Validation set | accuracy | Helped |
Surprising findings
- Many unlearning methods appear effective under instance-level evaluation but remain substantially inferable through correlated supporting facts, indicating overestimated unlearning.
- LLaMA3-8B-Instruct produced underconfident raw confidence scores (ECE 0.475) before calibration.
Limitations
- Supporting subgraphs depend on the external reference KG (YAGO3-10), so latent knowledge not covered by it cannot be captured.
- The LLM judge still shows discrepancies with humans on large subgraphs and on multi-hop inference steps.
- Entropy-based filtering can fail when unlearning impairs instruction following, requiring an ad hoc refinement.
Experimental setup
Datasets: YAGO3-10 Wikidata (reference, via YAGO3-10) MMLU BBH
Benchmarks: YAGO3-10-derived knowledge unlearning evaluation MMLU BBH
Metrics: UES (Unlearning Effectiveness Score) Recall Loc Gen (MMLU accuracy) Rea (BBH accuracy) ECE
Baselines: Gradient Ascent Random Labels NPO NegGrad+ SCRUB Instance unlearning (direct triple)
Builds on: NPO SCRUB TOFU RWKU
Closest papers by meaning
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.94)
- Probing Hidden Knowledge Holes in Unlearned LLMs (NeurIPS 2025 · similarity 0.93)
- Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting (NeurIPS 2025 · similarity 0.92)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.92)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.92)
- RWKU: Benchmarking Real-World Knowledge Unlearning for Large Language Models (NeurIPS 2024 · similarity 0.91)
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.