# Sparse Autoencoding on Pythia 70M (layer-3 residuals)

Source: [Sparse Autoencoding on Pythia 70M (layer-3 residuals)](<https://www.sota2.com/research/sota/sparse-autoencoding-on-pythia-70m-layer-3-residuals>)

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

Data updated: 2026-05-08T07:37:06.603Z

Evaluation of Sparse Autoencoding performance on the internal activations extracted from layer 3 of the Pythia 70M model, focusing on sparsity and feature representation 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 | Sparse Autoencoding |
| Dataset (including any stated split) | Pythia 70M (layer-3 residuals) |
| Metric | Achieved L0 |
| Unit | SCALAR |
| Direction | Higher is better |
| Results loaded | 4 of 4 |
| 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

- [Achieved L0](<https://www.sota2.com/research/sota/sparse-autoencoding-on-pythia-70m-layer-3-residuals.md?metric=Achieved%20L0>): Higher is better; unit: SCALAR.

- [Dead Neurons (%)](<https://www.sota2.com/research/sota/sparse-autoencoding-on-pythia-70m-layer-3-residuals.md?metric=Dead%20Neurons%20(%25)>): Lower is better; unit: PERCENT.

- [Explained Variance](<https://www.sota2.com/research/sota/sparse-autoencoding-on-pythia-70m-layer-3-residuals.md?metric=Explained%20Variance>): Higher is better; unit: PERCENT.

- [p90 Coherence](<https://www.sota2.com/research/sota/sparse-autoencoding-on-pythia-70m-layer-3-residuals.md?metric=p90%20Coherence>): Higher is better; unit: PERCENT.

- [Max Coherence](<https://www.sota2.com/research/sota/sparse-autoencoding-on-pythia-70m-layer-3-residuals.md?metric=Max%20Coherence>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Achieved L0 | Result date | Report paper |
| --- | --- | --- | --- |
| AEN-SAE | 33.9 | Not recorded | 69fd6de35497b3b670c89265 |
| TopK | 32 | Not recorded | 69fd6de35497b3b670c89265 |
| AEN-SAE | 18.2 | Not recorded | 69fd6de35497b3b670c89265 |
| TopK | 16 | Not recorded | 69fd6de35497b3b670c89265 |

### AEN-SAE

Achieved L0: 33.9. Result ID: 69fd924247f34ce924e03e7a.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6de35497b3b670c89265.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69fd6de35497b3b670c89265.md>)

Recorded evaluation configuration: {"Target l0":"≈ 32"}. This record does not assert independent reproduction.

### TopK

Achieved L0: 32. Result ID: 69fd924247f34ce924e03e79.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6de35497b3b670c89265.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa738d1e209131cea5d2c9.md>)

Recorded evaluation configuration: {"Target l0":"≈ 32"}. This record does not assert independent reproduction.

### AEN-SAE

Achieved L0: 18.2. Result ID: 69fd924247f34ce924e03e78.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6de35497b3b670c89265.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69fd6de35497b3b670c89265.md>)

Recorded evaluation configuration: {"Target l0":"≈ 16"}. This record does not assert independent reproduction.

### TopK

Achieved L0: 16. Result ID: 69fd924247f34ce924e03e77.

[Reporting paper](<https://www.sota2.com/research/paper/69fd6de35497b3b670c89265.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa738d1e209131cea5d2c9.md>)

Recorded evaluation configuration: {"Target l0":"≈ 16"}. This record does not assert independent reproduction.

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