Start / LessWrong (30+ Karma) / Negative results for saes on downstream tasks and deprioritising sae research gdm mech interp team progress update 2 by neel nanda lewis smith senthooran rajamanoharan arthur c

“Negative Results for SAEs On Downstream Tasks and Deprioritising SAE Research (GDM Mech Interp Team Progress Update #2)” by Neel Nanda, lewis smith, Senthooran Rajamanoharan, Arthur Conmy, Callum McDougall, Tom Lieberum, János Kramár, Rohin Shah

58 min • 26 mars 2025

Audio note: this article contains 31 uses of latex notation, so the narration may be difficult to follow. There's a link to the original text in the episode description.

Lewis Smith*, Sen Rajamanoharan*, Arthur Conmy, Callum McDougall, Janos Kramar, Tom Lieberum, Rohin Shah, Neel Nanda

* = equal contribution

The following piece is a list of snippets about research from the GDM mechanistic interpretability team, which we didn’t consider a good fit for turning into a paper, but which we thought the community might benefit from seeing in this less formal form. These are largely things that we found in the process of a project investigating whether sparse autoencoders were useful for downstream tasks, notably out-of-distribution probing.

TL;DR

To validate whether SAEs were a worthwhile technique, we explored whether they were useful on the downstream task of OOD generalisation when detecting harmful intent in user prompts
[...]

---

Outline:

(01:08) TL;DR

(02:38) Introduction

(02:41) Motivation

(06:09) Our Task

(08:35) Conclusions and Strategic Updates

(13:59) Comparing different ways to train Chat SAEs

(18:30) Using SAEs for OOD Probing

(20:21) Technical Setup

(20:24) Datasets

(24:16) Probing

(26:48) Results

(30:36) Related Work and Discussion

(34:01) Is it surprising that SAEs didn't work?

(39:54) Dataset debugging with SAEs

(42:02) Autointerp and high frequency latents

(44:16) Removing High Frequency Latents from JumpReLU SAEs

(45:04) Method

(45:07) Motivation

(47:29) Modifying the sparsity penalty

(48:48) How we evaluated interpretability

(50:36) Results

(51:18) Reconstruction loss at fixed sparsity

(52:10) Frequency histograms

(52:52) Latent interpretability

(54:23) Conclusions

(56:43) Appendix

The original text contained 7 footnotes which were omitted from this narration.

---

First published:
March 26th, 2025

Source:
https://www.lesswrong.com/posts/4uXCAJNuPKtKBsi28/sae-progress-update-2-draft

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Latent firing frequency histograms for Gated, JumpReLU and TopK SAEs. Unlike Gated SAEs, which use a L1 penalty that penalizes large latent activations, JumpReLU (middle) and TopK (bottom) SAEs exhibit high-frequency latents: latents that fire on 10% or more of tokens (i.e. that lie to the right of the dotted vertical line).

Reconstruction loss vs L0 for the various SAE architectures and loss functions used in our experiment. The quadratic-frequency penalty (QF loss) has slightly worse reconstruction loss at any given sparsity than standard JumpReLU SAEs (L0 loss), but still compare favourably versus Gated and TopK SAEs.

$Latent firing frequency histograms for JumpReLU SAEs trained with a standard L0 loss (top) or quadratic-frequency loss with __T3A_INLINE_LATEX_PLACEHOLDER___\omega_0=10^{-1}___T3A_INLINE_LATEX_END_PLACEHOLDER__ (middle) or __T3A_INLINE_LATEX_PLACEHOLDER___\omega_0=10^{-2}___T3A_INLINE_LATEX_END_PLACEHOLDER__ (bottom). The quadratic frequency loss successfully removes high frequency features (i.e. latents around or to the right of the red dotted vertical line) without changing the shape of the rest of the frequency histogram.$

Average auto-interp score vs L0 for the various SAE architectures and loss functions used in our experiment, for different latent weightings. Uniform weighting slightly disfavours JumpReLU variants but doesn't show clear patterns, frequency-weighting clearly shows outperformance of JumpReLU variants at lower sparsities (higher L0).

$Average autointerp score vs L0 for the JumpReLU SAEs and the JumpReLU QF loss variants with __T3A_INLINE_LATEX_PLACEHOLDER___\omega=0.01___T3A_INLINE_LATEX_END_PLACEHOLDER__. All SAEs show a negative trend of autointerp score against latent frequency, although the quadratic-frequency loss function seems to help the SAE form interpretable latents even at higher frequencies - the curves for higher L0 SAEs are squashed to the right.$