Why can 'forgotten' information come back after unlearning?
Unlearning methods often don't truly erase knowledge—they just suppress it, and current evaluations fail to catch that. In one study, applying quantization (a compression technique that reduces model precision) to unlearned models restored the 'forgotten' information: the average retention of intended-forgotten knowledge jumped from 21% in full precision to 83% after 4-bit quantization [2]. This means that if a model is deployed in a lower-precision format (common for efficiency), the supposedly removed data can resurface, and standard benchmarks wouldn't detect it because they test the full-precision model only.
The same study found that for unlearning methods with utility constraints, the model retained an average of 21% of the forgotten knowledge even in full precision—so the unlearning was never complete to begin with [2]. This suggests that evaluations that only check a few test queries can easily miss the residual knowledge that remains hidden in the model's weights.
How can benchmarks be fooled into showing success when unlearning failed?
Benchmarks are vulnerable to simple modifications that expose failures. A 2025 position paper found that by introducing even loose dependencies between the forget and retain information, they could show that supposedly unlearned information remains accessible, or that the unlearning process degraded the model's performance on retained information much more than the original benchmark indicated [4]. This means that real-world data often has overlapping topics, and benchmarks that treat forget and retain sets as independent can give a false sense of success.
Another study showed that current evaluation metrics are susceptible to red-teaming attacks—they reflect superficial model behaviors rather than the true extent of knowledge retained [3]. For example, a model might give a generic refusal response that looks like forgetting, but the underlying knowledge is still there. The authors proposed more robust metrics and a calibration method to restore performance on non-targeted data, but the key takeaway is that standard metrics can be gamed.
Are current unlearning benchmarks too optimistic about progress?
Yes, existing benchmarks provide an overly optimistic and potentially misleading view of unlearning effectiveness. The 2025 position paper demonstrated this by introducing benign modifications to popular benchmarks, which exposed that unlearned information remained accessible or that retained knowledge was degraded more than reported [4]. This suggests that many published unlearning results may not reflect real-world performance.
The same paper also warned that ambiguity in unlearning targets can lead to methods that overfit to the given test queries, meaning they perform well on the benchmark but fail on slightly different queries [4]. This is a critical gap because real-world unlearning requests are rarely as clean as benchmark test sets.
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2024 to 2025, 5 from 2024 or later, 1 in Q1 journals, collectively cited 130 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 25 papers retrieved from a database of over 500 million.
Sources used in this answer
Rethinking machine unlearning for large language models
This review highlights overlooked aspects of LLM unlearning research, including unlearning scope, data–model interaction, and multifaceted efficacy assessment, and calls for a more comprehensive evaluation framework.
Catastrophic Failure of LLM Unlearning via Quantization
This study found that applying quantization to unlearned models can restore 'forgotten' information: the average retention of intended-forgotten knowledge increased from 21% in full precision to 83% after 4-bit quantization, indicating that current benchmarks fail to detect hidden knowledge.
Towards Effective Evaluations and Comparisons for LLM Unlearning Methods
This paper shows that current evaluation metrics are susceptible to red-teaming attacks and may reflect superficial model behaviors rather than true knowledge retention, and proposes more robust metrics and a calibration method to improve evaluation.
Position: LLM Unlearning Benchmarks are Weak Measures of Progress
This position paper demonstrates that existing benchmarks provide an overly optimistic view of unlearning effectiveness, as simple modifications can expose that unlearned information remains accessible or that retained knowledge is degraded more than reported, and warns against overfitting to test queries.
The Frontier of Data Erasure: A Survey on Machine Unlearning for Large Language Models
This survey reviews machine unlearning methods for LLMs, focusing on targeted forgetting for privacy, ethical, and legal challenges, but does not provide specific quantitative findings on evaluation failures.