# Concept Steering on AxBench (Held-in)

Source: [Concept Steering on AxBench (Held-in)](<https://www.sota2.com/research/sota/concept-steering-on-axbench-held-in>)

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

Data updated: 2026-05-08T09:17:40.759Z

Evaluation of model performance on the Concept Steering task using the AxBench (Held-in) dataset, reporting HMean scores.

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 | Concept Steering |
| Dataset (including any stated split) | AxBench (Held-in) |
| Metric | HMean |
| Unit | SCALAR |
| Direction | Higher is better |
| Results loaded | 25 of 25 |
| 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

- [HMean](<https://www.sota2.com/research/sota/concept-steering-on-axbench-held-in.md?metric=HMean>): Higher is better; unit: SCALAR.

## Recorded results

| Method | HMean | Result date | Report paper |
| --- | --- | --- | --- |
| FLAS | 1.185 | Not recorded | 69fd6e455497b3b670c8960e |
| HyperSteer | 1.091 | Not recorded | 69fd6e455497b3b670c8960e |
| Prompting | 1.075 | Not recorded | 69fd6e455497b3b670c8960e |
| FLAS | 1.025 | Not recorded | 69fd6e455497b3b670c8960e |
| RePS | 0.892 | Not recorded | 69fd6e455497b3b670c8960e |
| LoReFT | 0.777 | Not recorded | 69fd6e455497b3b670c8960e |
| HyperSteer | 0.742 | Not recorded | 69fd6e455497b3b670c8960e |
| Prompting | 0.731 | Not recorded | 69fd6e455497b3b670c8960e |
| LoReFT | 0.722 | Not recorded | 69fd6e455497b3b670c8960e |
| SFT | 0.714 | Not recorded | 69fd6e455497b3b670c8960e |
| LoRA | 0.641 | Not recorded | 69fd6e455497b3b670c8960e |
| ReFT-r1 | 0.63 | Not recorded | 69fd6e455497b3b670c8960e |
| RePS | 0.606 | Not recorded | 69fd6e455497b3b670c8960e |
| LoRA | 0.602 | Not recorded | 69fd6e455497b3b670c8960e |
| ReFT-r1 | 0.509 | Not recorded | 69fd6e455497b3b670c8960e |
| DiffMean | 0.322 | Not recorded | 69fd6e455497b3b670c8960e |
| AcT (Layer 20) | 0.27 | Not recorded | 69fd6e455497b3b670c8960e |
| SAE | 0.191 | Not recorded | 69fd6e455497b3b670c8960e |
| AcT (All Layers) | 0.187 | Not recorded | 69fd6e455497b3b670c8960e |
| SAE-A | 0.186 | Not recorded | 69fd6e455497b3b670c8960e |
| DiffMean | 0.178 | Not recorded | 69fd6e455497b3b670c8960e |
| AcT (All Layers) | 0.161 | Not recorded | 69fd6e455497b3b670c8960e |
| SAE | 0.151 | Not recorded | 69fd6e455497b3b670c8960e |
| AcT (Layer 20) | 0.144 | Not recorded | 69fd6e455497b3b670c8960e |
| SAE-A | 0.132 | Not recorded | 69fd6e455497b3b670c8960e |

### FLAS

HMean: 1.185. Result ID: 69fda98e47f34ce924e0e396.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-9B-IT","Intervention Layer":"Layer 20","T":"2"}. This record does not assert independent reproduction.

### HyperSteer

HMean: 1.091. Result ID: 69fda98e47f34ce924e0e38f.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa79111e209131cea7bf3d.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-9B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### Prompting

HMean: 1.075. Result ID: 69fda98e47f34ce924e0e38b.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa773a1e209131cea71d18.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-9B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### FLAS

HMean: 1.025. Result ID: 69fda98e47f34ce924e0e38a.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-2B-IT","Intervention Layer":"Layer 20","T":"2"}. This record does not assert independent reproduction.

### RePS

HMean: 0.892. Result ID: 69fda98e47f34ce924e0e38e.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-9B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### LoReFT

HMean: 0.777. Result ID: 69fda98e47f34ce924e0e38c.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa73421e209131cea5b52a.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-9B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### HyperSteer

HMean: 0.742. Result ID: 69fda98e47f34ce924e0e383.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa79111e209131cea7bf3d.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-2B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### Prompting

HMean: 0.731. Result ID: 69fda98e47f34ce924e0e37e.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa773a1e209131cea71d18.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-2B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### LoReFT

HMean: 0.722. Result ID: 69fda98e47f34ce924e0e37f.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa73421e209131cea5b52a.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-2B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### SFT

HMean: 0.714. Result ID: 69fda98e47f34ce924e0e380.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-2B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### LoRA

HMean: 0.641. Result ID: 69fda98e47f34ce924e0e381.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6b121e209131cea2bf59.md>)

[Implementation](<https://github.com/microsoft/LoRA>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-2B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### ReFT-r1

HMean: 0.63. Result ID: 69fda98e47f34ce924e0e390.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa73421e209131cea5b52a.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-9B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### RePS

HMean: 0.606. Result ID: 69fda98e47f34ce924e0e382.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-2B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### LoRA

HMean: 0.602. Result ID: 69fda98e47f34ce924e0e38d.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6b121e209131cea2bf59.md>)

[Implementation](<https://github.com/microsoft/LoRA>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-9B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### ReFT-r1

HMean: 0.509. Result ID: 69fda98e47f34ce924e0e384.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa73421e209131cea5b52a.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-2B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### DiffMean

HMean: 0.322. Result ID: 69fda98e47f34ce924e0e392.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa74051e209131cea5f4ad.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-9B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### AcT (Layer 20)

HMean: 0.27. Result ID: 69fda98e47f34ce924e0e394.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75dc1e209131cea6a4b5.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-9B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### SAE

HMean: 0.191. Result ID: 69fda98e47f34ce924e0e393.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-9B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### AcT (All Layers)

HMean: 0.187. Result ID: 69fda98e47f34ce924e0e385.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75dc1e209131cea6a4b5.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-2B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### SAE-A

HMean: 0.186. Result ID: 69fda98e47f34ce924e0e395.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa773a1e209131cea71d18.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-9B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### DiffMean

HMean: 0.178. Result ID: 69fda98e47f34ce924e0e386.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa74051e209131cea5f4ad.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-2B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### AcT (All Layers)

HMean: 0.161. Result ID: 69fda98e47f34ce924e0e391.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75dc1e209131cea6a4b5.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-9B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### SAE

HMean: 0.151. Result ID: 69fda98e47f34ce924e0e387.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-2B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### AcT (Layer 20)

HMean: 0.144. Result ID: 69fda98e47f34ce924e0e388.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa75dc1e209131cea6a4b5.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-2B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

### SAE-A

HMean: 0.132. Result ID: 69fda98e47f34ce924e0e389.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6e455497b3b670c8960e.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa773a1e209131cea71d18.md>)

Recorded evaluation configuration: {"Base Model":"Gemma-2-2B-IT","Intervention Layer":"Layer 20"}. This record does not assert independent reproduction.

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