Sparse Crosscoders: Multi-View Model Diffing

• Sparse crosscoders are specialized sparse dictionary learning models that align neural activations across various network views using enforced sparsity and a shared overcomplete dictionary.
• They enable high-fidelity model diffing and mechanistic interpretability by tracking feature emergence, modification, and disappearance during training or fine-tuning.
• Variants like BatchTopK and Dedicated Feature Crosscoders overcome challenges such as feature absorption and dead neurons with hard sparsity constraints and tailored optimization strategies.

A sparse crosscoder is a specialized form of sparse dictionary learning (SDL) model designed to interpret and compare internal representations across multiple neural network “views”—typically different model checkpoints, architectures, or fine-tuning endpoints. Sparse crosscoders provide a unified framework for model diffing, mechanistic interpretability, and tracking the evolution of features during training or fine-tuning. By enforcing sparsity and learning a shared overcomplete dictionary, crosscoders allow researchers to localize and causally test the emergence, modification, or disappearance of high-level concepts in neural representation spaces. Variants such as BatchTopK, Dedicated Feature Crosscoders (DFC), and Delta-Crosscoders address architectural and statistical challenges in different model comparison regimes (Minder et al., 3 Apr 2025, Tang et al., 5 Dec 2025, Chaudhari et al., 6 Mar 2026, Shportko et al., 25 Jun 2026, Kassem et al., 16 Feb 2026, Bayazit et al., 5 Sep 2025, Ge et al., 21 Sep 2025).

1. Mathematical Foundations and Model Architecture

Sparse crosscoders extend the principled objective of sparse dictionary learning to the multi-view setting. For $m$ views (e.g., models/layers/checkpoints), one observes paired activations $x_p^{(i)}(s) \in \mathbb{R}^{n_p^{(i)}}$ and $x_r^{(i)}(s) \in \mathbb{R}^{n_r^{(i)}}$ for sample $s$. Stacking across all $i$ yields input and target activation vectors $x_p(s), x_r(s)$. The crosscoder consists of:

• Encoder: $z(s) = \sigma(W_e x_p(s))$ where $W_e \in \mathbb{R}^{n_q \times n_p}$ and $\sigma$ encodes sparsity (e.g., $\text{ReLU}$+$x_p^{(i)}(s) \in \mathbb{R}^{n_p^{(i)}}$0, Top-k).
• Decoder: $x_p^{(i)}(s) \in \mathbb{R}^{n_p^{(i)}}$1 with $x_p^{(i)}(s) \in \mathbb{R}^{n_p^{(i)}}$2.

The population objective is

$x_p^{(i)}(s) \in \mathbb{R}^{n_p^{(i)}}$3

with explicit or implicit sparsity regularization (Tang et al., 5 Dec 2025).

• For two models (base and fine-tuned), paired decoders reconstruct activations for each model using the same sparse code:

$x_p^{(i)}(s) \in \mathbb{R}^{n_p^{(i)}}$4

where $x_p^{(i)}(s) \in \mathbb{R}^{n_p^{(i)}}$5 (Minder et al., 3 Apr 2025).

Sparsity can be enforced by $x_p^{(i)}(s) \in \mathbb{R}^{n_p^{(i)}}$6, hard Top-K, or mixed constraints, with modern variants (BatchTopK, DFC) favoring hard $x_p^{(i)}(s) \in \mathbb{R}^{n_p^{(i)}}$7 for improved stability and faithfulness (Minder et al., 3 Apr 2025, Chaudhari et al., 6 Mar 2026, Shportko et al., 25 Jun 2026).

2. Optimization Landscape and Failure Modes

When the underlying representation is assumed to decompose linearly into interpretable latent factors—each activated rarely (extreme sparsity regime)—global minima of the crosscoder loss correspond to dictionaries that perfectly reconstruct each feature in isolation (Tang et al., 5 Dec 2025). However, spurious local minima are common due to “feature absorption,” where multiple ground-truth features are mapped to a single neuron, and “dead neurons” that never activate.

Key observed phenomena:

• Feature absorption: One neuron encodes a mixture of features, producing “feature mixing.”
• Dead neurons: Sparse units never active; gradient-based training cannot escape these flat regions.

Algorithmic countermeasures include careful initialization (using known projections), overcomplete dictionaries, dead-neuron resampling (re-initializing zero-activation rows), and multi-scale sparsity schedules (Tang et al., 5 Dec 2025).

3. Sparsity Induction Schemes and Artifacts

The standard $x_p^{(i)}(s) \in \mathbb{R}^{n_p^{(i)}}$8-based crosscoder loss

$x_p^{(i)}(s) \in \mathbb{R}^{n_p^{(i)}}$9

leads to two reliability failures:

• Complete Shrinkage: $x_r^{(i)}(s) \in \mathbb{R}^{n_r^{(i)}}$0 minimization can artificially drive certain decoder norms to zero, hiding features that are present in both models (spuriously labeling latents as unique to one view).
• Latent Decoupling: One shared concept may be split into two superficially unique latents if active contexts differ.

BatchTopK sparsity (enforcing a fixed hard budget of active features per batch via selection on $x_r^{(i)}(s) \in \mathbb{R}^{n_r^{(i)}}$1) eliminates both artifacts, ensuring faithfulness to actual model-specific concept emergence (Minder et al., 3 Apr 2025, Chaudhari et al., 6 Mar 2026, Kassem et al., 16 Feb 2026).

Latent Scaling metrics—for example, $x_r^{(i)}(s) \in \mathbb{R}^{n_r^{(i)}}$2 based on per-latent error ratios—can be used to retroactively filter spurious “unique” latents when only $x_r^{(i)}(s) \in \mathbb{R}^{n_r^{(i)}}$3 regularization is available (Minder et al., 3 Apr 2025).

4. Model Diffing, Causal Attribution, and Specializations

Sparse crosscoders enable a range of model-diffing and interpretability tasks, including:

• Fine-tuning Diffing: Identifying which features are introduced, shifted, or suppressed by fine-tuning (chat-tuning, RL, narrow behavioral modifications). BatchTopK and partitioned variants (e.g., DFC) allow high-fidelity mapping of new behaviors, isolating minimal feature sets responsible for phenomena such as refusal, tool use, and knowledge boundaries (Minder et al., 3 Apr 2025, Shportko et al., 25 Jun 2026, Kassem et al., 16 Feb 2026).
• Multi-Architecture Comparison: Explicit decoder partitioning and norm-based specialization metrics (e.g., $x_r^{(i)}(s) \in \mathbb{R}^{n_r^{(i)}}$4) separate dense- and MoE-specialized, as well as shared features. MoEs exhibit fewer, denser “monosemantic” features compared to dense models, which distribute information over broader, lower-activation features (Chaudhari et al., 6 Mar 2026).
• Feature Causal Attribution: Crosscoders allow precise ablation or additive steering (“patching”) of individual latents to assess their causal impact on downstream model output, with metrics such as KL divergence and task accuracy differential validating causal relationships (Minder et al., 3 Apr 2025, Ge et al., 21 Sep 2025, Shportko et al., 25 Jun 2026). RelIE (Relative Indirect Effect) provides a causal score quantifying when a feature becomes important for performance during pretraining progressions (Bayazit et al., 5 Sep 2025).
• Runtime Behavioral Control: Partitioned crosscoders (e.g., DFC) enable real-time manipulation of LLM behavior (e.g., toggling tool-use or refusal on/off) via additive interventions on a minimal set of sparse features, without additional model retraining (Shportko et al., 25 Jun 2026).

5. Tracking and Quantifying Feature Emergence Through Training

Sparse crosscoders systematically track the evolution of features during pretraining or over checkpoints/snapshots (Bayazit et al., 5 Sep 2025, Ge et al., 21 Sep 2025). Once trained, the dictionary and decoder-norms allow:

• Determination of feature presence: Emergence time, peak time, and lifetime are extracted from decoder norms.
• Causal assignment: Integrated gradient–based or patch-intervention methods score each latent’s direct effect on task-specific targets.
• Measurement of learning dynamics: Two-phase dynamics are observed—an initial statistical fitting phase, followed by discrete feature formation. The expansion of effective feature dimensionality, abrupt vs. gradual emergence, and later consolidation of high-level concepts (syntax, semantics) are captured via crosscoder analysis.

Relative Indirect Effect (RelIE) and similar ablation metrics isolate the pretraining stage at which a given latent transitions from non-causal to strongly causal for target behavior, supporting high-resolution analysis of representation learning trajectories (Bayazit et al., 5 Sep 2025).

6. Empirical Applications and Best Practices

Sparse crosscoders are applied in:

• Chat model interpretation (Gemma 2B): BatchTopK crosscoders reveal highly interpretable chat-specific and refusal-centric features with strong causal efficacy, while L1-based methods are prone to artifacts (Minder et al., 3 Apr 2025).
• MoE vs. Dense model comparison: Crosscoders reveal MoEs prioritize few, dense, specialized features compared to dense models’ diffuse representations (Chaudhari et al., 6 Mar 2026).
• RL-induced capability tracing: Dedicated Feature Crosscoders (DFC) isolate RL-specific tool-use to single or few very sparse units, enabling post-hoc runtime steering and demonstrating capability spillover between models (Shportko et al., 25 Jun 2026).
• Narrow fine-tuning diagnostics: Delta-Crosscoder introduces a contrastive, delta-focused loss with Dual-K sparsity, singularly surfacing causally responsible latents and exhibiting high coverage across challenging model diffing tasks—outperforming standard sparse autoencoder-based baselines (Kassem et al., 16 Feb 2026).
• Linguistic representation formation: Crosscoders track the appearance and causal importance of grammatical feature detectors (e.g., for number agreement, irregular morphology) and reveal that most features form abruptly at stage transitions (Bayazit et al., 5 Sep 2025, Ge et al., 21 Sep 2025).

Recommended protocol:

• Prefer $x_r^{(i)}(s) \in \mathbb{R}^{n_r^{(i)}}$5-style sparsity (BatchTopK, Dual-K, DFC) over penalized-$x_r^{(i)}(s) \in \mathbb{R}^{n_r^{(i)}}$6.
• When using $x_r^{(i)}(s) \in \mathbb{R}^{n_r^{(i)}}$7, apply latent scaling or analogous artifact-detection metrics before drawing causal or attribution conclusions.
• Validate causal status by ablation or additive patching, not by decoder-norm alone unless artifact-free sparsity is enforced.
• For moderate-sized (2–9B parameter) models, dictionary expansions of $x_r^{(i)}(s) \in \mathbb{R}^{n_r^{(i)}}$8 and sparsity levels of $x_r^{(i)}(s) \in \mathbb{R}^{n_r^{(i)}}$9 active features per position are empirically effective.
• Utilize decoder partitioning (explicit shared, exclusive slices) for diffing across architectures or behavioral fine-tuning (Minder et al., 3 Apr 2025, Tang et al., 5 Dec 2025, Chaudhari et al., 6 Mar 2026, Shportko et al., 25 Jun 2026, Kassem et al., 16 Feb 2026, Bayazit et al., 5 Sep 2025, Ge et al., 21 Sep 2025).

7. Limitations, Current Directions, and Open Problems

Sparse crosscoders depend on the linear representation and superposition hypotheses; real neural representations may involve moderate nonlinearity or context dependence, especially for emerging or drifted features. Feature absorption and dead neurons, while mitigated algorithmically, remain a theoretical and practical challenge when initialization is not near a global minimum, or data distribution shifts across views.

For narrow and asymmetric fine-tuning, additional loss components (delta-based regression, implicit contrastive pairing) are vital to surface subtle yet essential differences (Kassem et al., 16 Feb 2026). Scaling crosscoders to multi-stage, cross-architecture, or very large model chains is an ongoing area of research. Automated semantic labeling, improved multi-scale sparsity schedules, and direct incorporation of causal or auxiliary constraints are promising avenues for further progress.

In sum, sparse crosscoders unify a family of representation alignment and diffing techniques with direct mechanistic interpretability, providing high-fidelity, causally validated maps of emergent and shifting features in both conventional and frontier neural architectures (Minder et al., 3 Apr 2025, Tang et al., 5 Dec 2025, Chaudhari et al., 6 Mar 2026, Shportko et al., 25 Jun 2026, Kassem et al., 16 Feb 2026, Bayazit et al., 5 Sep 2025, Ge et al., 21 Sep 2025).