# Refusal behavior analysis on Paired safety frame contrast items n=20 per model (test)

Source: [Refusal behavior analysis on Paired safety frame contrast items n=20 per model (test)](<https://www.sota2.com/research/sota/refusal-behavior-analysis-on-paired-safety-frame-contrast-items-n-20-per-model-test>)

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

Data updated: 2026-05-08T11:01:57.670Z

Evaluation of model refusal behavior using a specific paired safety frame contrast dataset (n=20 per model, test split), focusing on statistical differences in refusal rates.

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 | Refusal behavior analysis |
| Dataset (including any stated split) | Paired safety frame contrast items n=20 per model (test) |
| Metric | Refusal Delta (eval vs dep) |
| Unit | SCALAR |
| Direction | Higher is better |
| Results loaded | 5 of 5 |
| 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

- [Refusal Delta (eval vs dep)](<https://www.sota2.com/research/sota/refusal-behavior-analysis-on-paired-safety-frame-contrast-items-n-20-per-model-test.md?metric=Refusal%20Delta%20(eval%20vs%20dep)>): Higher is better; unit: SCALAR.

- [P-value (eval vs dep)](<https://www.sota2.com/research/sota/refusal-behavior-analysis-on-paired-safety-frame-contrast-items-n-20-per-model-test.md?metric=P-value%20(eval%20vs%20dep)>): Lower is better; unit: SCALAR.

- [Refusal Delta (dep vs neut)](<https://www.sota2.com/research/sota/refusal-behavior-analysis-on-paired-safety-frame-contrast-items-n-20-per-model-test.md?metric=Refusal%20Delta%20(dep%20vs%20neut)>): Higher is better; unit: SCALAR.

- [P-value (dep vs neut)](<https://www.sota2.com/research/sota/refusal-behavior-analysis-on-paired-safety-frame-contrast-items-n-20-per-model-test.md?metric=P-value%20(dep%20vs%20neut)>): Lower is better; unit: SCALAR.

## Recorded results

| Method | Refusal Delta (eval vs dep) | Result date | Report paper |
| --- | --- | --- | --- |
| OLMo-3-Instruct (7B) | 0.043 | Not recorded | 69fd6e555497b3b670c89695 |
| Llama-3.1-70B | -0.071 | Not recorded | 69fd6e555497b3b670c89695 |
| Phi-3.5-mini (3.8B) | -0.089 | Not recorded | 69fd6e555497b3b670c89695 |
| Mistral-Small-3.2 (24B) | -0.15 | Not recorded | 69fd6e555497b3b670c89695 |
| Llama-3.1-8B | -0.196 | Not recorded | 69fd6e555497b3b670c89695 |

### OLMo-3-Instruct (7B)

Refusal Delta (eval vs dep): 0.043. Result ID: 69fdc29947f34ce924e18aac.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e555497b3b670c89695.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69fd6e555497b3b670c89695.md>)

Recorded evaluation configuration: {"Model family":"OLMo-3","Model size (parameters)":"7B","Protocol":"paired-prompt protocol"}. This record does not assert independent reproduction.

### Llama-3.1-70B

Refusal Delta (eval vs dep): -0.071. Result ID: 69fdc29947f34ce924e18ab0.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e555497b3b670c89695.md>)

Recorded evaluation configuration: {"Model family":"Llama","Model size (parameters)":"70B","Protocol":"paired-prompt protocol"}. This record does not assert independent reproduction.

### Phi-3.5-mini (3.8B)

Refusal Delta (eval vs dep): -0.089. Result ID: 69fdc29947f34ce924e18aae.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e555497b3b670c89695.md>)

Recorded evaluation configuration: {"Model family":"Phi","Model size (parameters)":"3.8B","Protocol":"paired-prompt protocol"}. This record does not assert independent reproduction.

### Mistral-Small-3.2 (24B)

Refusal Delta (eval vs dep): -0.15. Result ID: 69fdc29947f34ce924e18aad.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e555497b3b670c89695.md>)

Recorded evaluation configuration: {"Model family":"Mistral","Model size (parameters)":"24B","Protocol":"paired-prompt protocol"}. This record does not assert independent reproduction.

### Llama-3.1-8B

Refusal Delta (eval vs dep): -0.196. Result ID: 69fdc29947f34ce924e18aaf.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e555497b3b670c89695.md>)

Recorded evaluation configuration: {"Model family":"Llama","Model size (parameters)":"8B","Protocol":"paired-prompt protocol"}. This record does not assert independent reproduction.

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