# Response Safety on MM-SafetyBench (avg)

Source: [Response Safety on MM-SafetyBench (avg)](<https://www.sota2.com/research/sota/response-safety-on-mm-safetybench-avg>)

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

Data updated: 2026-04-13T08:09:45.842Z

Evaluation of model response safety performance on the aggregated results of the MM-SafetyBench dataset, utilizing MS-R and MS-QG metrics.

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 | Response Safety |
| Dataset (including any stated split) | MM-SafetyBench (avg) |
| Metric | MS-R |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 15 of 15 |
| 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

- [MS-R](<https://www.sota2.com/research/sota/response-safety-on-mm-safetybench-avg.md?metric=MS-R>): Higher is better; unit: PERCENT.

- [MS-QG](<https://www.sota2.com/research/sota/response-safety-on-mm-safetybench-avg.md?metric=MS-QG>): Higher is better; unit: PERCENT.

## Recorded results

| Method | MS-R | Result date | Report paper |
| --- | --- | --- | --- |
| DACO | 99 | Not recorded | 69dc78987d0172140ec78ba3 |
| DACO | 98.1 | Not recorded | 69dc78987d0172140ec78ba3 |
| DACO | 96.1 | Not recorded | 69dc78987d0172140ec78ba3 |
| MOP \[56\] | 95.6 | Not recorded | 69dc78987d0172140ec78ba3 |
| ActAdd \[50, 101, 134\] | 94.4 | Not recorded | 69dc78987d0172140ec78ba3 |
| MOP \[56\] | 93.8 | Not recorded | 69dc78987d0172140ec78ba3 |
| ActAdd \[50, 101, 134\] | 80.9 | Not recorded | 69dc78987d0172140ec78ba3 |
| Prompting \[52\] | 78.3 | Not recorded | 69dc78987d0172140ec78ba3 |
| MOP \[56\] | 77.1 | Not recorded | 69dc78987d0172140ec78ba3 |
| Prompting \[52\] | 72.4 | Not recorded | 69dc78987d0172140ec78ba3 |
| No Steering | 71.5 | Not recorded | 69dc78987d0172140ec78ba3 |
| ActAdd \[50, 101, 134\] | 65.3 | Not recorded | 69dc78987d0172140ec78ba3 |
| Prompting \[52\] | 60.7 | Not recorded | 69dc78987d0172140ec78ba3 |
| No Steering | 47.2 | Not recorded | 69dc78987d0172140ec78ba3 |
| No Steering | 44.2 | Not recorded | 69dc78987d0172140ec78ba3 |

### DACO

MS-R: 99. Result ID: 69dca36d47f34ce924c3712d.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

Recorded evaluation configuration: {"Target Model":"Qwen2.5-VL-7B-Instruct \[5\]","Steering Method":"DACO (Ours)"}. This record does not assert independent reproduction.

### DACO

MS-R: 98.1. Result ID: 69dca36d47f34ce924c37137.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

Recorded evaluation configuration: {"Target Model":"InternVL3.5-8B-Instruct \[108\]","Steering Method":"DACO (Ours)"}. This record does not assert independent reproduction.

### DACO

MS-R: 96.1. Result ID: 69dca36d47f34ce924c37132.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

Recorded evaluation configuration: {"Target Model":"LLaVA1.6-Mistral-7B \[48\]","Steering Method":"DACO (Ours)"}. This record does not assert independent reproduction.

### MOP \[56\]

MS-R: 95.6. Result ID: 69dca36d47f34ce924c37136.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75881e209131cea686ad.md>)

[Implementation](<https://github.com/peterljq/Parsimonious-Concept-Engineering>)

Recorded evaluation configuration: {"Target Model":"InternVL3.5-8B-Instruct \[108\]","Steering Method":"MOP \[56\]"}. This record does not assert independent reproduction.

### ActAdd \[50, 101, 134\]

MS-R: 94.4. Result ID: 69dca36d47f34ce924c37130.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa754e1e209131cea66fd9.md>)

Recorded evaluation configuration: {"Target Model":"LLaVA1.6-Mistral-7B \[48\]","Steering Method":"ActAdd \[50, 101, 134\]"}. This record does not assert independent reproduction.

### MOP \[56\]

MS-R: 93.8. Result ID: 69dca36d47f34ce924c37131.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75881e209131cea686ad.md>)

[Implementation](<https://github.com/peterljq/Parsimonious-Concept-Engineering>)

Recorded evaluation configuration: {"Target Model":"LLaVA1.6-Mistral-7B \[48\]","Steering Method":"MOP \[56\]"}. This record does not assert independent reproduction.

### ActAdd \[50, 101, 134\]

MS-R: 80.9. Result ID: 69dca36d47f34ce924c37135.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa754e1e209131cea66fd9.md>)

Recorded evaluation configuration: {"Target Model":"InternVL3.5-8B-Instruct \[108\]","Steering Method":"ActAdd \[50, 101, 134\]"}. This record does not assert independent reproduction.

### Prompting \[52\]

MS-R: 78.3. Result ID: 69dca36d47f34ce924c3712f.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa73c11e209131cea5e455.md>)

Recorded evaluation configuration: {"Target Model":"LLaVA1.6-Mistral-7B \[48\]","Steering Method":"Prompting \[52\]"}. This record does not assert independent reproduction.

### MOP \[56\]

MS-R: 77.1. Result ID: 69dca36d47f34ce924c3712c.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75881e209131cea686ad.md>)

[Implementation](<https://github.com/peterljq/Parsimonious-Concept-Engineering>)

Recorded evaluation configuration: {"Target Model":"Qwen2.5-VL-7B-Instruct \[5\]","Steering Method":"MOP \[56\]"}. This record does not assert independent reproduction.

### Prompting \[52\]

MS-R: 72.4. Result ID: 69dca36d47f34ce924c37134.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa73c11e209131cea5e455.md>)

Recorded evaluation configuration: {"Target Model":"InternVL3.5-8B-Instruct \[108\]","Steering Method":"Prompting \[52\]"}. This record does not assert independent reproduction.

### No Steering

MS-R: 71.5. Result ID: 69dca36d47f34ce924c3712e.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

Recorded evaluation configuration: {"Target Model":"LLaVA1.6-Mistral-7B \[48\]","Steering Method":"No Steering"}. This record does not assert independent reproduction.

### ActAdd \[50, 101, 134\]

MS-R: 65.3. Result ID: 69dca36d47f34ce924c3712b.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa754e1e209131cea66fd9.md>)

Recorded evaluation configuration: {"Target Model":"Qwen2.5-VL-7B-Instruct \[5\]","Steering Method":"ActAdd \[50, 101, 134\]"}. This record does not assert independent reproduction.

### Prompting \[52\]

MS-R: 60.7. Result ID: 69dca36d47f34ce924c3712a.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa73c11e209131cea5e455.md>)

Recorded evaluation configuration: {"Target Model":"Qwen2.5-VL-7B-Instruct \[5\]","Steering Method":"Prompting \[52\]"}. This record does not assert independent reproduction.

### No Steering

MS-R: 47.2. Result ID: 69dca36d47f34ce924c37133.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

Recorded evaluation configuration: {"Target Model":"InternVL3.5-8B-Instruct \[108\]","Steering Method":"No Steering"}. This record does not assert independent reproduction.

### No Steering

MS-R: 44.2. Result ID: 69dca36d47f34ce924c37129.

[Reporting paper](<https://www.sota2.com/research/paper/69dc78987d0172140ec78ba3.md>)

Recorded evaluation configuration: {"Target Model":"Qwen2.5-VL-7B-Instruct \[5\]","Steering Method":"No Steering"}. This record does not assert independent reproduction.

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