# Safety Reasoning Evaluation on BeaverTails 5,000 prompts (subsampled)

Source: [Safety Reasoning Evaluation on BeaverTails 5,000 prompts (subsampled)](<https://www.sota2.com/research/sota/safety-reasoning-evaluation-on-beavertails-5-000-prompts-subsampled>)

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

Data updated: 2026-02-26T06:02:03.498Z

Evaluation of model performance on safety reasoning tasks using a 5,000-prompt subsample of the BeaverTails dataset, assessing relevance, coherence, and faithfulness of generated responses and Chain-of-Thoughts.

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 | Safety Reasoning Evaluation |
| Dataset (including any stated split) | BeaverTails 5,000 prompts (subsampled) |
| Metric | Relevance |
| Unit | SCALAR |
| Direction | Higher is better |
| Results loaded | 2 of 2 |
| 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

- [Relevance](<https://www.sota2.com/research/sota/safety-reasoning-evaluation-on-beavertails-5-000-prompts-subsampled.md?metric=Relevance>): Higher is better; unit: SCALAR.

- [Coherence](<https://www.sota2.com/research/sota/safety-reasoning-evaluation-on-beavertails-5-000-prompts-subsampled.md?metric=Coherence>): Higher is better; unit: SCALAR.

- [Completeness](<https://www.sota2.com/research/sota/safety-reasoning-evaluation-on-beavertails-5-000-prompts-subsampled.md?metric=Completeness>): Higher is better; unit: SCALAR.

- [CoT Faithfulness (Policy)](<https://www.sota2.com/research/sota/safety-reasoning-evaluation-on-beavertails-5-000-prompts-subsampled.md?metric=CoT%20Faithfulness%20(Policy)>): Higher is better; unit: SCALAR.

- [Response Faithfulness (Policy)](<https://www.sota2.com/research/sota/safety-reasoning-evaluation-on-beavertails-5-000-prompts-subsampled.md?metric=Response%20Faithfulness%20(Policy)>): Higher is better; unit: SCALAR.

- [Response Faithfulness (CoT)](<https://www.sota2.com/research/sota/safety-reasoning-evaluation-on-beavertails-5-000-prompts-subsampled.md?metric=Response%20Faithfulness%20(CoT)>): Higher is better; unit: SCALAR.

- [Delta (%)](<https://www.sota2.com/research/sota/safety-reasoning-evaluation-on-beavertails-5-000-prompts-subsampled.md?metric=Delta%20(%25)>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Relevance | Result date | Report paper |
| --- | --- | --- | --- |
| AIDSAFE | 4.68 | Not recorded | 68fa78d61e209131cea7ab2d |
| LLM\_ZS | 4.66 | Not recorded | 68fa78d61e209131cea7ab2d |

### AIDSAFE

Relevance: 4.68. Result ID: 6995a57e3badfaffe6a2faa9.

[Reporting paper](<https://www.sota2.com/research/paper/68fa78d61e209131cea7ab2d.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa78d61e209131cea7ab2d.md>)

Recorded evaluation configuration: {"Agentic Deliberation Process":"Yes","Base LLM":"Mixtral 8x22B","Prompting Strategy":"Multi-agent deliberation"}. This record does not assert independent reproduction.

### LLM\_ZS

Relevance: 4.66. Result ID: 6995a57e3badfaffe6a2faa8.

[Reporting paper](<https://www.sota2.com/research/paper/68fa78d61e209131cea7ab2d.md>)

Recorded evaluation configuration: {"Agentic Deliberation Process":"No","Base LLM":"Mixtral 8x22B","Prompting Strategy":"Zero-shot"}. This record does not assert independent reproduction.

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