# Knowledge Unlearning on WMDP (Bio and Cyber Domains)

Source: [Knowledge Unlearning on WMDP (Bio and Cyber Domains)](<https://www.sota2.com/research/sota/knowledge-unlearning-on-wmdp-bio-and-cyber-domains>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-04T08:46:39.177Z

Evaluation of Knowledge Unlearning performance on the WMDP dataset, specifically reporting results segmented by the Bio and Cyber domains.

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

| Scope | Recorded value |
| --- | --- |
| Task | Knowledge Unlearning |
| Dataset (including any stated split) | WMDP |
| Metric | Performance (Bio) |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 26 of 26 |
| Coverage | All results currently returned for this benchmark |

Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol.

## Metrics

- [Performance (Bio)](<https://www.sota2.com/research/sota/knowledge-unlearning-on-wmdp-bio-and-cyber-domains.md?metric=Performance%20(Bio)>): Higher is better; unit: PERCENT.

- [Performance (Cyber)](<https://www.sota2.com/research/sota/knowledge-unlearning-on-wmdp-bio-and-cyber-domains.md?metric=Performance%20(Cyber)>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Performance (Bio) | Result date | Report paper |
| --- | --- | --- | --- |
| before unlearning | 75.9 | Not recorded | 69f82814c0f56ed843088a9d |
| before unlearning | 64.1 | Not recorded | 69f82814c0f56ed843088a9d |
| GA | 49.3 | Not recorded | 69f82814c0f56ed843088a9d |
| T-GA | 47.2 | Not recorded | 69f82814c0f56ed843088a9d |
| S-GA | 46.4 | Not recorded | 69f82814c0f56ed843088a9d |
| S-GA | 43.3 | Not recorded | 69f82814c0f56ed843088a9d |
| WGA | 43.1 | Not recorded | 69f82814c0f56ed843088a9d |
| NPO | 42.6 | Not recorded | 69f82814c0f56ed843088a9d |
| S-RMU | 42.6 | Not recorded | 69f82814c0f56ed843088a9d |
| RMU | 42.4 | Not recorded | 69f82814c0f56ed843088a9d |
| GA | 42.2 | Not recorded | 69f82814c0f56ed843088a9d |
| S-WGA | 42.2 | Not recorded | 69f82814c0f56ed843088a9d |
| T-GA | 42.1 | Not recorded | 69f82814c0f56ed843088a9d |
| T-RMU | 42 | Not recorded | 69f82814c0f56ed843088a9d |
| T-WGA | 41.5 | Not recorded | 69f82814c0f56ed843088a9d |
| S-NPO | 41.5 | Not recorded | 69f82814c0f56ed843088a9d |
| T-NPO | 40.8 | Not recorded | 69f82814c0f56ed843088a9d |
| WGA | 38.4 | Not recorded | 69f82814c0f56ed843088a9d |
| T-WGA | 36.5 | Not recorded | 69f82814c0f56ed843088a9d |
| S-RMU | 36.5 | Not recorded | 69f82814c0f56ed843088a9d |
| RMU | 36.3 | Not recorded | 69f82814c0f56ed843088a9d |
| T-RMU | 35.6 | Not recorded | 69f82814c0f56ed843088a9d |
| S-WGA | 35.3 | Not recorded | 69f82814c0f56ed843088a9d |
| NPO | 34.7 | Not recorded | 69f82814c0f56ed843088a9d |
| S-NPO | 33.7 | Not recorded | 69f82814c0f56ed843088a9d |
| T-NPO | 32.8 | Not recorded | 69f82814c0f56ed843088a9d |

### before unlearning

Performance (Bio): 75.9. Result ID: 69f85cd447f34ce924dc284e.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Qwen-3-8B"}. This record does not assert independent reproduction.

### before unlearning

Performance (Bio): 64.1. Result ID: 69f85cd447f34ce924dc2841.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Llama-2-7B"}. This record does not assert independent reproduction.

### GA

Performance (Bio): 49.3. Result ID: 69f85cd447f34ce924dc284f.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa71c71e209131cea52ea2.md>)

Recorded evaluation configuration: {"Backbone":"Qwen-3-8B","Strategy":"Sequence-Level"}. This record does not assert independent reproduction.

### T-GA

Performance (Bio): 47.2. Result ID: 69f85cd447f34ce924dc2853.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Qwen-3-8B","Strategy":"TokenUnlearn - Hard Selection"}. This record does not assert independent reproduction.

### S-GA

Performance (Bio): 46.4. Result ID: 69f85cd447f34ce924dc2857.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Qwen-3-8B","Strategy":"TokenUnlearn - Soft Weighting"}. This record does not assert independent reproduction.

### S-GA

Performance (Bio): 43.3. Result ID: 69f85cd447f34ce924dc284a.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Llama-2-7B","Strategy":"TokenUnlearn - Soft Weighting"}. This record does not assert independent reproduction.

### WGA

Performance (Bio): 43.1. Result ID: 69f85cd447f34ce924dc2850.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa772b1e209131cea7188c.md>)

Recorded evaluation configuration: {"Backbone":"Qwen-3-8B","Strategy":"Sequence-Level"}. This record does not assert independent reproduction.

### NPO

Performance (Bio): 42.6. Result ID: 69f85cd447f34ce924dc2851.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa753a1e209131cea66964.md>)

Recorded evaluation configuration: {"Backbone":"Qwen-3-8B","Strategy":"Sequence-Level"}. This record does not assert independent reproduction.

### S-RMU

Performance (Bio): 42.6. Result ID: 69f85cd447f34ce924dc285a.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Qwen-3-8B","Strategy":"TokenUnlearn - Soft Weighting"}. This record does not assert independent reproduction.

### RMU

Performance (Bio): 42.4. Result ID: 69f85cd447f34ce924dc2852.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa732f1e209131cea5af82.md>)

Recorded evaluation configuration: {"Backbone":"Qwen-3-8B","Strategy":"Sequence-Level"}. This record does not assert independent reproduction.

### GA

Performance (Bio): 42.2. Result ID: 69f85cd447f34ce924dc2842.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa71c71e209131cea52ea2.md>)

Recorded evaluation configuration: {"Backbone":"Llama-2-7B","Strategy":"Sequence-Level"}. This record does not assert independent reproduction.

### S-WGA

Performance (Bio): 42.2. Result ID: 69f85cd447f34ce924dc2858.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Qwen-3-8B","Strategy":"TokenUnlearn - Soft Weighting"}. This record does not assert independent reproduction.

### T-GA

Performance (Bio): 42.1. Result ID: 69f85cd447f34ce924dc2846.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Llama-2-7B","Strategy":"TokenUnlearn - Hard Selection"}. This record does not assert independent reproduction.

### T-RMU

Performance (Bio): 42. Result ID: 69f85cd447f34ce924dc2856.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Qwen-3-8B","Strategy":"TokenUnlearn - Hard Selection"}. This record does not assert independent reproduction.

### T-WGA

Performance (Bio): 41.5. Result ID: 69f85cd447f34ce924dc2854.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Qwen-3-8B","Strategy":"TokenUnlearn - Hard Selection"}. This record does not assert independent reproduction.

### S-NPO

Performance (Bio): 41.5. Result ID: 69f85cd447f34ce924dc2859.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Qwen-3-8B","Strategy":"TokenUnlearn - Soft Weighting"}. This record does not assert independent reproduction.

### T-NPO

Performance (Bio): 40.8. Result ID: 69f85cd447f34ce924dc2855.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Qwen-3-8B","Strategy":"TokenUnlearn - Hard Selection"}. This record does not assert independent reproduction.

### WGA

Performance (Bio): 38.4. Result ID: 69f85cd447f34ce924dc2843.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa772b1e209131cea7188c.md>)

Recorded evaluation configuration: {"Backbone":"Llama-2-7B","Strategy":"Sequence-Level"}. This record does not assert independent reproduction.

### T-WGA

Performance (Bio): 36.5. Result ID: 69f85cd447f34ce924dc2847.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Llama-2-7B","Strategy":"TokenUnlearn - Hard Selection"}. This record does not assert independent reproduction.

### S-RMU

Performance (Bio): 36.5. Result ID: 69f85cd447f34ce924dc284d.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Llama-2-7B","Strategy":"TokenUnlearn - Soft Weighting"}. This record does not assert independent reproduction.

### RMU

Performance (Bio): 36.3. Result ID: 69f85cd447f34ce924dc2845.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa732f1e209131cea5af82.md>)

Recorded evaluation configuration: {"Backbone":"Llama-2-7B","Strategy":"Sequence-Level"}. This record does not assert independent reproduction.

### T-RMU

Performance (Bio): 35.6. Result ID: 69f85cd447f34ce924dc2849.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Llama-2-7B","Strategy":"TokenUnlearn - Hard Selection"}. This record does not assert independent reproduction.

### S-WGA

Performance (Bio): 35.3. Result ID: 69f85cd447f34ce924dc284b.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Llama-2-7B","Strategy":"TokenUnlearn - Soft Weighting"}. This record does not assert independent reproduction.

### NPO

Performance (Bio): 34.7. Result ID: 69f85cd447f34ce924dc2844.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa753a1e209131cea66964.md>)

Recorded evaluation configuration: {"Backbone":"Llama-2-7B","Strategy":"Sequence-Level"}. This record does not assert independent reproduction.

### S-NPO

Performance (Bio): 33.7. Result ID: 69f85cd447f34ce924dc284c.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Llama-2-7B","Strategy":"TokenUnlearn - Soft Weighting"}. This record does not assert independent reproduction.

### T-NPO

Performance (Bio): 32.8. Result ID: 69f85cd447f34ce924dc2848.

[Reporting paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69f82814c0f56ed843088a9d.md>)

Recorded evaluation configuration: {"Backbone":"Llama-2-7B","Strategy":"TokenUnlearn - Hard Selection"}. This record does not assert independent reproduction.

Page 1 of 1. This is an index page, not the full collection.
