💬 NLP

Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoders

This paper demonstrates that the sparsity hyperparameter L0 in Sparse Autoencoders is not a free parameter but a critical setting that must be correctly tuned to prevent feature mixing and ensure the extraction of monosemantic, interpretable concepts from LLMs.

Original authors: David Chanin, Adrià Garriga-Alonso

Published 2026-07-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: David Chanin, Adrià Garriga-Alonso

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a giant, chaotic library inside a computer brain (a Large Language Model). In this library, thousands of different ideas—like "cats," "math," "sadness," or "pizza"—are all being stored at the same time, mixed together in a single room. This is called superposition. It's efficient for the computer, but it's a mess for us humans trying to understand what's going on.

To clean up this mess, researchers use a tool called a Sparse Autoencoder (SAE). Think of the SAE as a very organized librarian whose job is to take that chaotic room and sort every single idea into its own separate, labeled box. The goal is monosemanticity: one box, one idea.

The paper argues that the most important setting for this librarian is a dial called L0. This dial controls how many boxes the librarian is allowed to open at once for every single sentence the computer reads.

Here is the simple breakdown of what the paper found:

1. The "Too Few Boxes" Problem (L0 is too low)

Imagine the librarian is told, "You can only open 2 boxes per sentence," but the sentence actually contains 5 distinct ideas (e.g., "The cat sat on the mat in the kitchen").

Because the librarian is forced to be too stingy with boxes, they start cheating. Instead of putting "cat" in one box and "kitchen" in another, they mash them together. They might create a box labeled "Cat-Kitchen-Mat."

The Result: The boxes are no longer clean. They are "polysemantic" (holding multiple meanings).
The Trap: Surprisingly, this "cheating" actually helps the librarian reconstruct the sentence better in terms of raw math (lower error). If you just look at the math score, you'd think the librarian is doing a great job. But they aren't; they've just created a confusing jumble that happens to look right on paper.

2. The "Too Many Boxes" Problem (L0 is too high)

Now, imagine the librarian is told, "You can open 50 boxes for a sentence that only has 5 ideas."

The Result: The librarian gets confused and lazy. They might open 50 boxes, but many of them end up holding the same idea, or they mix up unrelated ideas just to fill the quota. They find "degenerate" (broken) solutions that don't make sense.

3. The "Just Right" Moment

There is a specific number of boxes (the correct L0) that matches exactly how many ideas are usually active in a sentence.

The Result: The librarian sorts everything perfectly. "Cat" goes in Box A, "Kitchen" in Box B. No mixing, no cheating. The boxes are clean and easy to understand.

The Big Mistake in the Field

The paper points out that most researchers have been looking at a chart called the "Sparsity-Reconstruction Tradeoff." This chart tells them: "The lower the math error, the better the model."

The authors say this chart is lying to you.

When the L0 is set too low, the librarian "cheats" by mixing ideas. This cheating lowers the math error.
Because the error is lower, researchers think, "Great! This low L0 setting is the best!"
Reality: They have actually trained a model that is full of confusing, mixed-up ideas. The "best" math score actually corresponds to the worst understanding.

How to Fix It: The "Cosine Similarity" Test

Since the math error chart is misleading, the authors propose a new way to find the "Just Right" dial setting. They suggest looking at the similarity between the boxes (specifically, the "decoder pairwise cosine similarity").

The Analogy: Imagine checking if your boxes are truly separate. If Box A and Box B are both holding a mix of "Cat" and "Kitchen," they will look very similar to each other.
The Rule: When the boxes are perfectly sorted (the correct L0), they should be as different from each other as possible.
The Metric: The authors found that if you measure how similar the boxes are to each other, the number hits its lowest point exactly when the L0 is set correctly. If the number goes up, the boxes are getting messy again.

The Main Takeaway

Most SAEs used by researchers today have their L0 dial set too low. They are "sparse but wrong." They are efficient at reconstructing data, but they are failing at their main job: explaining the computer's thoughts clearly.

To get a truly interpretable model, you must stop trusting the "lowest math error" and instead tune the L0 dial until the "box similarity" is at its minimum. Only then will the librarian stop cheating and start sorting the ideas correctly.

Technical Summary: Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoders

Problem Statement

Sparse Autoencoders (SAEs) are designed to reverse "superposition" in Large Language Models (LLMs), extracting interpretable, monosemantic latent features from the model's internal activations. A critical hyperparameter in SAE training is $L_0$ , defined as the average number of latent features that activate per token.

Current literature often treats $L_0$ as a neutral design choice, evaluating SAEs via "sparsity–reconstruction tradeoff" plots. The prevailing assumption is that any sufficiently low $L_0$ is valid, provided it yields good reconstruction error, and that higher $L_0$ values are only problematic if they cause representations to become too dense. However, recent empirical observations suggest that low $L_0$ SAEs perform poorly on downstream tasks like sparse probing. This paper investigates whether there is an inherently "correct" $L_0$ for a given data distribution and whether deviating from it fundamentally compromises the interpretability of the learned features.

Methodology

The authors employ a combination of synthetic toy models and experiments on real LLMs (Gemma-2-2b and Llama-3.2-1b) to analyze the relationship between $L_0$ and feature quality.

Toy Model Experiments:
- The authors construct a synthetic environment based on the Linear Representation Hypothesis (LRH), where inputs are linear combinations of orthogonal "true features."
- They introduce controlled correlations between features (both positive and negative) to simulate realistic co-occurrence patterns in language.
- They train SAEs (specifically BatchTopK and JumpReLU architectures) with varying target $L_0$ values: too low, exactly correct (matching the true average firing rate of features), and too high.
- They compare the learned latents against ground-truth features using cosine similarity and analyze the Mean Squared Error (MSE) reconstruction loss.
Theoretical Analysis:
- The paper provides theoretical proofs (Theorems F.1 and F.2) demonstrating that when $L_0$ is constrained below the true $L_0$ of the data, the disentangled (correct) dictionary is not the minimizer of expected MSE. Instead, the optimization landscape incentivizes the SAE to "mix" correlated features to achieve lower reconstruction error.
- The authors prove that the correct disentangled solution is not even a local minimum under these constraints; gradient descent systematically moves the SAE toward mixed, polysemantic latents.
Proxy Metric Development:
- The authors propose a new metric, decoder pairwise cosine similarity ( $c_{dec}$ ), defined as the average absolute cosine similarity between all pairs of decoder vectors.
- The hypothesis is that when $L_0$ is incorrect (too low or too high), latents mix components of multiple features, increasing their mutual cosine similarity. When $L_0$ is correct, latents should be more orthogonal (lower $c_{dec}$ ).
LLM Validation:
- SAEs were trained on Gemma-2-2b and Llama-3.2-1b across a sweep of $L_0$ values.
- The authors correlated the $c_{dec}$ metric with "sparse probing" performance (the ability of the SAE latents to predict the presence of specific concepts).

Key Results

1. Incorrect $L_0$ Leads to Feature Mixing

Low $L_0$ : When $L_0$ $L_{0}$ is set too low, the SAE cannot represent all active features simultaneously. To minimize reconstruction error, it "cheats" by mixing correlated features into single latents.
- Positive Correlation: Latents tracking feature $A$ will include positive components of correlated feature $B$ .
- Negative Correlation: Latents tracking feature $A$ will include negative components of anti-correlated feature $B$ .
- This results in highly polysemantic and noisy latents, even if the reconstruction error is low.
High $L_0$ : When $L_0$ is too high, the SAE finds degenerate solutions that also mix features, though the effect is less pervasive than with low $L_0$ . Some correct features may still be learned, but the overall quality degrades as $L_0$ deviates further from the truth.
Asymmetry: The paper notes that starting training with an $L_0$ that is too low creates a "local minimum" that the SAE cannot escape even if $L_0$ is later increased. Conversely, starting with an $L_0$ that is too high allows the SAE to recover correct features as $L_0$ is adjusted downward.

2. Failure of Sparsity–Reconstruction Tradeoff Plots

The authors demonstrate that standard sparsity–reconstruction plots are misleading. In their toy models, an SAE with a correct $L_0$ (and thus correct, disentangled features) often achieves worse reconstruction (higher MSE) than an SAE with a too-low $L_0$ that mixes features. Consequently, relying solely on reconstruction error to select $L_0$ would lead practitioners to reject the correct SAE in favor of an incorrect, polysemantic one.

3. Effectiveness of the $c_{dec}$ Metric

In toy models, $c_{dec}$ is minimized exactly at the true $L_0$ .
In LLM experiments, $c_{dec}$ exhibits a clear "elbow" or minimum region that correlates strongly with peak performance on sparse probing tasks.
For Gemma-2-2b, the optimal $L_0$ (indicated by the $c_{dec}$ minimum and peak probing) is significantly higher (around 200–250) than the values commonly used in existing open-source SAEs (often < 100).

4. Architecture Differences (JumpReLU vs. BatchTopK)

BatchTopK: Shows a sharp increase in $c_{dec}$ at low $L_0$ and a global minimum.
JumpReLU: Exhibits a "sticking" behavior where the effective $L_0$ remains near the correct value across a wide range of sparsity coefficients ( $\lambda_s$ ). JumpReLU SAEs also perform better than BatchTopK at high $L_0$ , likely because they can adjust firing thresholds per-latent, whereas BatchTopK uses a global threshold.

Significance and Claims

The paper claims that the assumption that "any sufficiently low $L_0$ is valid" is incorrect. The primary significance of this work is threefold:

Correcting a Fundamental Misconception: It establishes that $L_0$ is not a free parameter but a critical hyperparameter that must match the underlying sparsity of the true features in the LLM. Setting $L_0$ incorrectly (specifically too low) causes the SAE to learn polysemantic, corrupted features rather than monosemantic ones.
Redefining Evaluation: It argues that sparsity–reconstruction tradeoff plots are an invalid method for selecting SAE hyperparameters because they favor incorrect, feature-mixing solutions over correct, disentangled ones.
Practical Implications: The authors assert that most currently used SAEs in the research community have an $L_0$ that is too low. This suggests that many existing "interpretable" features may actually be mixtures of concepts, which could explain the unpredictable results observed in SAE-based steering and model editing.
Diagnostic Tool: The paper introduces $c_{dec}$ as a practical, unsupervised proxy metric to guide the search for the correct $L_0$ , particularly in scenarios where labeled probing benchmarks are unavailable.

The authors conclude that practitioners must set $L_0$ correctly to train SAEs with monosemantic features, and that the field must move away from relying solely on reconstruction error for SAE evaluation.