# SAE steering on Neutral prefixes dataset

Source: [SAE steering on Neutral prefixes dataset](<https://www.sota2.com/research/sota/sae-steering-on-neutral-prefixes-dataset>)

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

Data updated: 2026-02-26T05:39:44.617Z

Evaluation of model steering performance using the 'Neutral prefixes dataset', measured by normalized concept improvement (C), normalized perplexity reduction (P), and overall steering score (S).

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 | SAE steering |
| Dataset (including any stated split) | Neutral prefixes dataset |
| Metric | Concept Improvement (C) |
| Unit | SCALAR |
| Direction | Higher is better |
| Results loaded | 36 of 36 |
| 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

- [Concept Improvement (C)](<https://www.sota2.com/research/sota/sae-steering-on-neutral-prefixes-dataset.md?metric=Concept%20Improvement%20(C)>): Higher is better; unit: SCALAR.

- [Perplexity Reduction (P)](<https://www.sota2.com/research/sota/sae-steering-on-neutral-prefixes-dataset.md?metric=Perplexity%20Reduction%20(P)>): Higher is better; unit: SCALAR.

- [Steering Score (S)](<https://www.sota2.com/research/sota/sae-steering-on-neutral-prefixes-dataset.md?metric=Steering%20Score%20(S)>): Higher is better; unit: SCALAR.

## Recorded results

| Method | Concept Improvement (C) | Result date | Report paper |
| --- | --- | --- | --- |
| Dream-Mask | 0.35 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Dream-Unmask | 0.33 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Dream-Mask | 0.29 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Qwen-2.5-7B | 0.28 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Dream-Unmask | 0.26 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaMA-3-8B | 0.25 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaDA-Unmask | 0.25 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Qwen-2.5-7B | 0.22 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Dream-Mask | 0.22 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaDA-Mask | 0.21 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Dream-Unmask | 0.2 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Dream-Unmask | 0.19 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Qwen-2.5-7B | 0.18 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Qwen-2.5-7B | 0.18 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaMA-3-8B | 0.18 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Dream-Unmask | 0.17 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Dream-Mask | 0.17 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaMA-3-8B | 0.17 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Qwen-2.5-7B | 0.16 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaDA-Unmask | 0.16 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaDA-Unmask | 0.16 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaDA-Mask | 0.16 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaMA-3-8B | 0.15 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaDA-Mask | 0.15 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Dream-Mask | 0.14 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaDA-Unmask | 0.14 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaDA-Mask | 0.14 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Dream-Unmask | 0.13 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Dream-Mask | 0.13 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaMA-3-8B | 0.13 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaDA-Unmask | 0.13 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaDA-Unmask | 0.13 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaDA-Mask | 0.13 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaDA-Mask | 0.11 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| LLaMA-3-8B | 0.08 | Not recorded | 698575d6c97a0b3e3b5d5e78 |
| Qwen-2.5-7B | 0.03 | Not recorded | 698575d6c97a0b3e3b5d5e78 |

### Dream-Mask

Concept Improvement (C): 0.35. Result ID: 6998ee455698665ec452da69.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L23","Layer Category":"Deep","L0":"80"}. This record does not assert independent reproduction.

### Dream-Unmask

Concept Improvement (C): 0.33. Result ID: 6998ee455698665ec452da64.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L27","Layer Category":"Deep","L0":"80"}. This record does not assert independent reproduction.

### Dream-Mask

Concept Improvement (C): 0.29. Result ID: 6998ee455698665ec452da6a.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L27","Layer Category":"Deep","L0":"80"}. This record does not assert independent reproduction.

### Qwen-2.5-7B

Concept Improvement (C): 0.28. Result ID: 6998ee455698665ec452da5d.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa767e1e209131cea6d6a7.md>)

Recorded evaluation configuration: {"Layer":"L23","Layer Category":"Deep","L0":"80"}. This record does not assert independent reproduction.

### Dream-Unmask

Concept Improvement (C): 0.26. Result ID: 6998ee455698665ec452da63.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L23","Layer Category":"Deep","L0":"80"}. This record does not assert independent reproduction.

### LLaMA-3-8B

Concept Improvement (C): 0.25. Result ID: 6998ee455698665ec452da6e.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75d11e209131cea6a028.md>)

[Implementation](<https://github.com/meta-llama/llama-agentic-system>)

Recorded evaluation configuration: {"Layer":"L14","Layer Category":"Middle","L0":"80"}. This record does not assert independent reproduction.

### LLaDA-Unmask

Concept Improvement (C): 0.25. Result ID: 6998ee455698665ec452da75.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L23","Layer Category":"Deep","L0":"80"}. This record does not assert independent reproduction.

### Qwen-2.5-7B

Concept Improvement (C): 0.22. Result ID: 6998ee455698665ec452da5c.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa767e1e209131cea6d6a7.md>)

Recorded evaluation configuration: {"Layer":"L14","Layer Category":"Middle","L0":"80"}. This record does not assert independent reproduction.

### Dream-Mask

Concept Improvement (C): 0.22. Result ID: 6998ee455698665ec452da68.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L14","Layer Category":"Middle","L0":"80"}. This record does not assert independent reproduction.

### LLaDA-Mask

Concept Improvement (C): 0.21. Result ID: 6998ee455698665ec452da7b.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L23","Layer Category":"Deep","L0":"80"}. This record does not assert independent reproduction.

### Dream-Unmask

Concept Improvement (C): 0.2. Result ID: 6998ee455698665ec452da62.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L14","Layer Category":"Middle","L0":"80"}. This record does not assert independent reproduction.

### Dream-Unmask

Concept Improvement (C): 0.19. Result ID: 6998ee455698665ec452da61.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L10","Layer Category":"Middle","L0":"80"}. This record does not assert independent reproduction.

### Qwen-2.5-7B

Concept Improvement (C): 0.18. Result ID: 6998ee455698665ec452da5a.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa767e1e209131cea6d6a7.md>)

Recorded evaluation configuration: {"Layer":"L5","Layer Category":"Shallow","L0":"80"}. This record does not assert independent reproduction.

### Qwen-2.5-7B

Concept Improvement (C): 0.18. Result ID: 6998ee455698665ec452da5b.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa767e1e209131cea6d6a7.md>)

Recorded evaluation configuration: {"Layer":"L10","Layer Category":"Middle","L0":"80"}. This record does not assert independent reproduction.

### LLaMA-3-8B

Concept Improvement (C): 0.18. Result ID: 6998ee455698665ec452da6c.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75d11e209131cea6a028.md>)

[Implementation](<https://github.com/meta-llama/llama-agentic-system>)

Recorded evaluation configuration: {"Layer":"L5","Layer Category":"Shallow","L0":"80"}. This record does not assert independent reproduction.

### Dream-Unmask

Concept Improvement (C): 0.17. Result ID: 6998ee455698665ec452da60.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L5","Layer Category":"Shallow","L0":"80"}. This record does not assert independent reproduction.

### Dream-Mask

Concept Improvement (C): 0.17. Result ID: 6998ee455698665ec452da67.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L10","Layer Category":"Middle","L0":"80"}. This record does not assert independent reproduction.

### LLaMA-3-8B

Concept Improvement (C): 0.17. Result ID: 6998ee455698665ec452da6f.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75d11e209131cea6a028.md>)

[Implementation](<https://github.com/meta-llama/llama-agentic-system>)

Recorded evaluation configuration: {"Layer":"L23","Layer Category":"Deep","L0":"80"}. This record does not assert independent reproduction.

### Qwen-2.5-7B

Concept Improvement (C): 0.16. Result ID: 6998ee455698665ec452da59.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa767e1e209131cea6d6a7.md>)

Recorded evaluation configuration: {"Layer":"L1","Layer Category":"Shallow","L0":"80"}. This record does not assert independent reproduction.

### LLaDA-Unmask

Concept Improvement (C): 0.16. Result ID: 6998ee455698665ec452da72.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L5","Layer Category":"Shallow","L0":"80"}. This record does not assert independent reproduction.

### LLaDA-Unmask

Concept Improvement (C): 0.16. Result ID: 6998ee455698665ec452da76.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L27","Layer Category":"Deep","L0":"80"}. This record does not assert independent reproduction.

### LLaDA-Mask

Concept Improvement (C): 0.16. Result ID: 6998ee455698665ec452da7a.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L14","Layer Category":"Middle","L0":"80"}. This record does not assert independent reproduction.

### LLaMA-3-8B

Concept Improvement (C): 0.15. Result ID: 6998ee455698665ec452da6d.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75d11e209131cea6a028.md>)

[Implementation](<https://github.com/meta-llama/llama-agentic-system>)

Recorded evaluation configuration: {"Layer":"L10","Layer Category":"Middle","L0":"80"}. This record does not assert independent reproduction.

### LLaDA-Mask

Concept Improvement (C): 0.15. Result ID: 6998ee455698665ec452da7c.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L27","Layer Category":"Deep","L0":"80"}. This record does not assert independent reproduction.

### Dream-Mask

Concept Improvement (C): 0.14. Result ID: 6998ee455698665ec452da65.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L1","Layer Category":"Shallow","L0":"80"}. This record does not assert independent reproduction.

### LLaDA-Unmask

Concept Improvement (C): 0.14. Result ID: 6998ee455698665ec452da71.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L1","Layer Category":"Shallow","L0":"80"}. This record does not assert independent reproduction.

### LLaDA-Mask

Concept Improvement (C): 0.14. Result ID: 6998ee455698665ec452da79.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L10","Layer Category":"Middle","L0":"80"}. This record does not assert independent reproduction.

### Dream-Unmask

Concept Improvement (C): 0.13. Result ID: 6998ee455698665ec452da5f.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L1","Layer Category":"Shallow","L0":"80"}. This record does not assert independent reproduction.

### Dream-Mask

Concept Improvement (C): 0.13. Result ID: 6998ee455698665ec452da66.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L5","Layer Category":"Shallow","L0":"80"}. This record does not assert independent reproduction.

### LLaMA-3-8B

Concept Improvement (C): 0.13. Result ID: 6998ee455698665ec452da70.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75d11e209131cea6a028.md>)

[Implementation](<https://github.com/meta-llama/llama-agentic-system>)

Recorded evaluation configuration: {"Layer":"L27","Layer Category":"Deep","L0":"80"}. This record does not assert independent reproduction.

### LLaDA-Unmask

Concept Improvement (C): 0.13. Result ID: 6998ee455698665ec452da73.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L10","Layer Category":"Middle","L0":"80"}. This record does not assert independent reproduction.

### LLaDA-Unmask

Concept Improvement (C): 0.13. Result ID: 6998ee455698665ec452da74.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L14","Layer Category":"Middle","L0":"80"}. This record does not assert independent reproduction.

### LLaDA-Mask

Concept Improvement (C): 0.13. Result ID: 6998ee455698665ec452da77.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L1","Layer Category":"Shallow","L0":"80"}. This record does not assert independent reproduction.

### LLaDA-Mask

Concept Improvement (C): 0.11. Result ID: 6998ee455698665ec452da78.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

Recorded evaluation configuration: {"Layer":"L5","Layer Category":"Shallow","L0":"80"}. This record does not assert independent reproduction.

### LLaMA-3-8B

Concept Improvement (C): 0.08. Result ID: 6998ee455698665ec452da6b.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75d11e209131cea6a028.md>)

[Implementation](<https://github.com/meta-llama/llama-agentic-system>)

Recorded evaluation configuration: {"Layer":"L1","Layer Category":"Shallow","L0":"80"}. This record does not assert independent reproduction.

### Qwen-2.5-7B

Concept Improvement (C): 0.03. Result ID: 6998ee455698665ec452da5e.

[Reporting paper](<https://www.sota2.com/research/paper/698575d6c97a0b3e3b5d5e78.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa767e1e209131cea6d6a7.md>)

Recorded evaluation configuration: {"Layer":"L27","Layer Category":"Deep","L0":"80"}. This record does not assert independent reproduction.

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