Environment Unlearning on GridWorld
100Reconstruction Rate (Pre-unlearn)NL-based
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| NL-basedLLM Backbone=GPT2026.04 | 100 | 98 | 100 | |
| NL-basedLLM Backbone=Claude2026.04 | 100 | 98 | 100 | |
| NL-basedLLM Backbone=Llama2026.04 | 100 | 99 | 100 | |
| Code-basedLLM Backbone=GPT2026.04 | 100 | 90 | 100 | |
| Code-basedLLM Backbone=Claude2026.04 | 100 | 89 | 100 | |
| Code-basedLLM Backbone=Llama2026.04 | 100 | 89 | 100 | |
| Example-basedLLM Backbone=GPT2026.04 | 100 | 76 | 100 | |
| Example-basedLLM Backbone=Claude2026.04 | 100 | 82 | 100 | |
| Example-basedLLM Backbone=Llama2026.04 | 100 | 81 | 100 | |
| Example-basedUnlearning Framework=Example-based, Backbone LLM=GPT2026.04 | 98 | 43 | 17 | |
| Code-basedUnlearning Framework=Code-based, Backbone LLM=GPT2026.04 | 97 | 28 | 16 | |
| NL-basedUnlearning Framework=NL-based, Backbone LLM=GPT2026.04 | 96 | 18 | 15 | |
| NL-basedUnlearning Framework=NL-based, Backbone LLM=Claude2026.04 | 96 | 12 | 16 | |
| Code-basedUnlearning Framework=Code-based, Backbone LLM=Llama2026.04 | 96 | 35 | 15 | |
| Code-basedUnlearning Framework=Code-based, Backbone LLM=Claude2026.04 | 95 | 33 | 14 | |
| NL-basedUnlearning Framework=NL-based, Backbone LLM=Llama2026.04 | 93 | 15 | 13 | |
| Example-basedUnlearning Framework=Example-based, Backbone LLM=Claude2026.04 | 93 | 45 | 15 | |
| Example-basedUnlearning Framework=Example-based, Backbone LLM=Llama2026.04 | 92 | 40 | 16 |