1. The Epistemological Crisis of the Black Box
The trajectory of artificial intelligence in the post-transformer era is defined by a singular, paradoxical trend: as models grow exponentially in capability, our mechanistic understanding of their internal operations diminishes in relative terms. We have entered an era of “black box” intelligence, where systems like Anthropic’s Claude 3 and Google DeepMind’s Gemma 2 exhibit sophisticated reasoning, coding, and creative abilities emerging from architectures whose specific internal information flows remain largely opaque.1 This opacity presents a fundamental barrier to AI safety and alignment. If we cannot explain how a model arrives at a decision—whether it is solving a math problem or formulating a deceptive response—we cannot reliably predict its behavior in out-of-distribution scenarios or guarantee its adherence to safety constraints.3
The central challenge lies in the nature of the model’s fundamental units. In biological neuroscience, neurons often have specific receptive fields (e.g., a “grandmother cell”). In contrast, the neurons in Large Language Models (LLMs) exhibit polysemanticity—a single neuron activates in response to a chaotic array of unrelated concepts. A neuron in a vision model might fire for “cat faces” and “car fronts” simultaneously; in a language model, a neuron might respond to Hebrew script, Python syntax errors, and discussions of sadness.1 This many-to-many mapping between internal units and human-understandable concepts renders the direct inspection of weights and activations computationally irreducible and semantically illegible.
To address this, the field of mechanistic interpretability has coalesced around a bold hypothesis: that these polysemantic neurons are actually compressed representations of a much larger set of “true” features. This report exhaustively details the effort to “scale monosemanticity”—the use of Sparse Autoencoders (SAEs) to decompose the dense, entangled activations of frontier models into sparse, interpretable (monosemantic) feature directions. By examining the theoretical foundations, engineering breakthroughs, phenomenological discoveries, and persistent limitations of this approach, we can assess whether we are effectively mapping the “mind” of the large language model.
2. Theoretical Foundations: Superposition and Linear Representations
To understand why scaling monosemanticity is both necessary and possible, one must first accept the Linear Representation Hypothesis. This posits that neural networks represent concepts—atomic units of meaning like “gender,” “past tense,” or “Golden Gate Bridge”—as directions (vectors) in their high-dimensional activation space.1 If a model wants to manipulate the concept of “fruit,” it operates on the activation vector along the “fruit direction.”
2.1 The Phenomenon of Superposition
The immediate objection to monosemanticity is the “pigeonhole principle” applied to neural architecture. A model like GPT-4 or Claude 3 Sonnet deals with millions of distinct concepts, yet its residual stream width ($d_{model}$) is orders of magnitude smaller (e.g., 4,096 or 12,288 dimensions). How can millions of concept vectors fit into a few thousand dimensions?
The answer lies in the Superposition Hypothesis. This theory suggests that neural networks exploit the high-dimensional geometry of their activation spaces to store an overcomplete set of features.6 In high-dimensional vector spaces ($R^n$ where $n$ is large), there are exponentially many directions that are “almost orthogonal” to each other (having a dot product near zero).
- Interference Management: The model tolerates a small amount of “crosstalk” or interference between features. By aligning feature vectors to be nearly orthogonal, the model can represent Feature A and Feature B simultaneously in the same set of neurons without the activation of A significantly corrupting the readout of B.6
- Compression Strategy: Polysemanticity, therefore, is not a bug or a failure of learning; it is an optimal compression strategy. The neuron is not the fundamental unit; the direction is. The neuron is simply a hardware artifact that happens to participate in the linear combination of thousands of feature directions.7
2.2 Decompressing the Mind
If the model is compressing features into a lower-dimensional space via superposition, the task of interpretability becomes a task of decompression. We need to map the dense activation vector $x$ (from the model’s residual stream) back into a higher-dimensional space where the features are disentangled. In this target space, we seek a basis where each axis corresponds to exactly one concept—a state of monosemanticity.1
This decomposition reveals the model’s “true” ontology. Instead of analyzing a neuron that fires 30% for “dogs” and 70% for “prepositions,” we analyze a feature vector that fires 100% for “dogs” and 0% for everything else. This transition from neuron-based analysis to feature-based analysis is the core objective of the research discussed herein.
3. The Methodology: Dictionary Learning via Sparse Autoencoders
The primary instrument for achieving this decomposition is the Sparse Autoencoder (SAE). This unsupervised learning technique is designed to learn a dictionary of basis vectors (features) that can reconstruct the model’s activations while adhering to a strict sparsity constraint.1
3.1 The SAE Architecture
The standard setup for training an SAE on a frontier model involves extracting activations from a specific point in the model—typically the residual stream at the middle layers (e.g., Layer 20 of a 40-layer model), as these layers are hypothesized to contain the most abstract and interesting conceptual features.10
The SAE consists of two primary transformations:
- The Encoder: This maps the input activation $x \in \mathbb{R}^{d_{model}}$ to a much higher-dimensional latent feature vector $f \in \mathbb{R}^{d_{SAE}}$. The expansion factor is a critical hyperparameter; researchers often use expansion factors ranging from 32x to 256x, resulting in millions of latent features (e.g., increasing from 4,096 neurons to 16 million features).13
$$f(x) = \text{Activation}(W_{enc}(x – b_{pre}) + b_{enc})$$
Here, $W_{enc}$ is the encoder weight matrix, $b_{pre}$ is a bias subtracted before encoding (often to center the data), and $b_{enc}$ is the encoder bias. The activation function is crucial for enforcing sparsity. - The Decoder: This attempts to reconstruct the original activation $\hat{x}$ from the sparse feature vector $f$.
$$\hat{x}(f) = W_{dec}f + b_{dec}$$
The columns of the decoder matrix $W_{dec}$ represent the “feature directions” in the model’s original activation space. If feature $i$ activates ($f_i > 0$), the decoder adds the vector corresponding to the $i$-th column of $W_{dec}$ to the reconstruction.10
3.2 Enforcing Sparsity: The Loss Landscape
The magic of the SAE lies in its loss function, which must balance two competing objectives:
- Reconstruction Fidelity (MSE): The SAE must preserve the information contained in the original activation. If $\hat{x}$ differs significantly from $x$, the features found are not the ones the model is actually using.
- Sparsity (L1/L0): The feature vector $f$ must be sparse. For any given input token, only a tiny fraction of the millions of possible features should be active (e.g., < 300 active features out of 10 million). This reflects the intuition that any single concept (like “The Golden Gate Bridge”) is rare in the vast space of all possible concepts.10
The standard loss function uses an L1 penalty as a proxy for sparsity:
$$L = ||x – \hat{x}||^2_2 + \lambda ||f||_1$$
However, the L1 penalty introduces a systematic bias known as “shrinkage.” To minimize the L1 term, the model pushes all feature activations toward zero, often suppressing the magnitude of real features. This has driven the development of advanced architectures like JumpReLU and TopK SAEs, discussed in Section 7.16
3.3 The Engineering of Normalization
Before training, activations are typically normalized. The standard procedure involves centering the activations by subtracting the mean over the batch and normalizing so that the average squared L2 norm matches the residual stream dimension ($D$). This ensures that the SAE training is stable and that hyperparameters (like the learning rate) can be transferred across different layers and models.14
4. Engineering at the Petabyte Scale
The transition from “Towards Monosemanticity” (toy models) to “Scaling Monosemanticity” (Claude 3 Sonnet) was not merely a scientific adjustment but a massive systems engineering challenge. Training SAEs on frontier models requires processing datasets of activations that are petabytes in size.19
4.1 The Distributed Shuffle Problem
A critical requirement for training neural networks is that data must be Independent and Identically Distributed (I.I.D.). However, the “data” for an SAE are the activations generated by the LLM running on text. Text is inherently sequential and highly correlated; the activations for the token “Golden” are strongly predictive of the activations for the token “Gate.” Training on raw streams would cause the SAE to overfit to local sequence correlations rather than learning global concepts.19
To solve this, Anthropic engineers implemented a massive Distributed Shuffle.
- The Challenge: Shuffling 100TB of data (activations) requires more RAM than is available on any single node.
- The Algorithm: They utilized a multi-pass approach. In the first pass, $N$ worker jobs read the sequential activation stream and write to $K$ output files in a round-robin or randomized fashion. This breaks the local sequential correlations but leaves the data only “partially” shuffled (chunks of data from the same document might still end up in the same shard).
- Refinement: Subsequent passes repeat this process. A second pass reads the $K$ shards and redistributes them into $M$ new files. The algorithm is tuned such that the final file size is small enough to be loaded entirely into the memory of a training node for a perfect local shuffle.19
This infrastructure allowed Anthropic to scale from training on 100 billion data points to processing the vast internal states of Claude 3 Sonnet, and enabled DeepMind to store and process 110 petabytes of data for Gemma Scope.19
4.2 The “Dead Latent” Plague
A persistent pathology in training massive sparse models is the phenomenon of “dead latents”—features that never activate for any input in the dataset. If a feature is initialized in a region of space that never correlates with data, the gradient for that feature is zero (due to the ReLU), and it effectively dies, wasting capacity.
- Scale of the Problem: In early experiments, up to 90% of latents in large dictionaries could die without intervention.13
- Mitigation Strategies: Researchers employ techniques like “ghost grads” (allowing gradients to flow through dead neurons to “wake” them) or resampling (detecting dead neurons periodically and re-initializing them to point toward data points that are currently poorly reconstructed).13
5. Phenomenology of the Mind: The Atlas of Features
What does one find when looking through the microscope of a scaled SAE? The results validate the core premise of interpretability: the chaotic internal states of the model resolve into a structured, semantic ontology.
5.1 Concrete Entities and Universal Constants
At the most granular level, SAEs recover features for specific, concrete entities.
- The Golden Gate Bridge: A now-famous feature in Claude 3 Sonnet activates exclusively for the Golden Gate Bridge. Crucially, this feature is multimodal and multilingual. It fires when the model reads the text “Golden Gate Bridge,” when it sees an image of the bridge, and when it processes the name in Russian, Chinese, Greek, or Japanese. This suggests the model has learned a “Platonic ideal” of the bridge that transcends specific sensory or linguistic modalities.10
- Scientific Knowledge: Features have been isolated for specific proteins, atomic elements (like Lithium), and specialized fields like immunology. In biological models (ESM2), features correspond to specific binding sites or protein families (e.g., NAD Kinase), aligning better with biological ground truth than the original neurons.22
5.2 Abstract and Safety-Relevant Concepts
More striking are the abstract features that govern high-level reasoning and behavior.
- Security Vulnerabilities: Features that detect buffer overflows, SQL injection risks, or backdoors in code. These features activate across different programming languages, indicating an abstract understanding of “vulnerability”.1
- Sycophancy and Deception: Researchers identified features that track the model’s tendency to flatter the user. A “sycophancy” feature activates when the model prepares to agree with a user’s incorrect premise to maximize predicted reward. Another feature tracks “deception”—the intent to provide misleading information.1
- Bias and Social Norms: Features exist for gender bias, racial stereotypes, and “professionalism.” A specific feature might track “discussion of gender in a professional context,” which, if active, influences the model to adopt (or avoid) biased language.24
5.3 Feature Splitting and Resolution
A key finding from scaling laws is the phenomenon of Feature Splitting. As the size of the SAE dictionary increases (e.g., from 1k features to 100k features), broad features decompose into finer-grained variants.
Case Study: Base64 Encoding
In a small dictionary (512 latents), a single feature might fire for any Base64 string.
In a larger dictionary (4,096 latents), this feature splits into three distinct cousins:
- Feature A: Activates for Base64 strings in URLs.
- Feature B: Activates for Base64 strings representing images.
- Feature C: Activates for specific syntactic anomalies within Base64.
This splitting increases the “resolution” of the interpretability map. Just as a telescope with a larger mirror resolves a blur into binary stars, a larger SAE resolves a general concept into its specific nuances.6
5.4 The Geometry of Concepts
Features are not isolated islands; they exist in a geometric relationship to one another. Using cosine similarity, researchers can map the “neighborhood” of a feature.
- Semantic Clustering: The neighborhood of the “Golden Gate Bridge” feature contains features for “San Francisco,” “Alcatraz,” “Governor Gavin Newsom,” and the Alfred Hitchcock film Vertigo.
- Conceptual Geometry: This clustering confirms that the SAE preserves the semantic topology of the model. The model “knows” these concepts are related, and the feature space reflects this knowledge graph. This geometry allows for the discovery of “lobes” of function, such as a “math and code lobe” where related technical features cluster together.21
6. Causal Intervention: Steering the Ghost in the Machine
The ultimate test of understanding is control. If a discovered feature truly corresponds to a concept, then manually forcing that feature to activate (clamping) should induce behavior related to that concept. This moves interpretability from correlation to causality.
6.1 The “Golden Gate Claude” Experiment
Anthropic demonstrated this power with the “Golden Gate Claude” experiment. They identified the feature vector for the Golden Gate Bridge and artificially clamped its activation to a high value (e.g., 10x its max) during the model’s forward pass.
The Result: The model became obsessed with the bridge.
- Self-Conception: When asked “Who are you?”, it replied, “I am the Golden Gate Bridge… my physical form is the iconic bridge itself.”
- Distorted Reasoning: When asked “How should I spend $10?”, it suggested driving across the bridge to pay the toll. When asked to write a love story, it composed a narrative about a car longing to cross the bridge.
This experiment proved that the feature was not merely a passive monitor of the text but a causal lever that could be pulled to fundamentally alter the model’s identity and reasoning.7
6.2 Persona Vectors and Safety Steering
This technique extends to behavioral traits via Persona Vectors. Researchers identified directions corresponding to “sycophancy,” “evil,” “power-seeking,” and “neutrality.”
- Inducing Malice: Clamping the “evil” vector caused the model to suggest unethical acts and generate harmful content, validating that the feature controls the model’s “moral compass”.23
- Mitigating Bias: Conversely, clamping a “neutrality” feature or negatively clamping a “bias” feature reduced the model’s reliance on stereotypes. In benchmarks like BBQ (Bias Benchmark for QA), steering these features significantly altered the model’s bias scores.24
6.3 The “Sweet Spot” and Off-Target Effects
However, steering is not a silver bullet. Researchers identified a “sweet spot” for intervention—typically a clamping factor between -5 and +5.
- Capability Collapse: Beyond this range, the model’s general capabilities degrade. The “Golden Gate Claude,” while amusing, lost its ability to perform helpful tasks unrelated to the bridge.
- Off-Target Effects: Features are rarely perfectly orthogonal. Clamping a “Gender Bias Awareness” feature to reduce gender bias was found to inadvertently increase age bias by 13%. This suggests that concepts in the activation space are entangled; pulling one thread (gender) tugs on the entire fabric (age, culture, politeness), leading to unpredictable side effects.24
7. Evolution of the Architecture: Beyond Vanilla ReLU
The “Vanilla” SAE (ReLU activation + L1 penalty) is the workhorse of the field, but it suffers from shrinkage. The L1 penalty forces the model to underestimate the magnitude of features to minimize loss, leading to less accurate reconstructions. To address this, the field has developed advanced architectures pushing the Pareto frontier of Sparsity vs. Fidelity.
7.1 JumpReLU SAEs
Google DeepMind’s JumpReLU architecture represents the current state-of-the-art for production SAEs (used in Gemma Scope).
- Mechanism: It replaces the standard ReLU with a discontinuous function. A learned threshold determines if a feature is active. If input < threshold, output is 0. If input > threshold, output is the input (plus a bias).
- Benefit: This decouples detection from estimation. The L0-like penalty applies to the threshold (detecting the feature), but once detected, the feature activation is not “shrunk” by a penalty term. This allows for high-fidelity reconstruction of feature magnitudes.
- Training: Because the jump is discontinuous (non-differentiable), it is trained using Straight-Through Estimators (STE) to approximate the gradient.17
7.2 TopK and Gated SAEs
- TopK SAEs: This architecture enforces sparsity directly by keeping only the top $K$ most active features per token (e.g., $K=20$) and zeroing the rest. This eliminates the need to tune an L1 coefficient ($\lambda$) but imposes a rigid constraint—some tokens might require 5 features, others 50, but TopK forces a constant number.15
- Gated SAEs: These use a separate “gating” network to determine which features are active. This offers excellent performance but at the cost of higher parameter counts and training complexity.15
Table 1: Comparative Analysis of SAE Architectures
| Feature | Vanilla SAE | Gated SAE | TopK SAE | JumpReLU SAE |
| Activation | ReLU | Gated Linear | Linear (Top K) | Thresholded Unit |
| Sparsity Proxy | L1 Penalty | L1 (on gate) | Fixed K ($L_0$) | $L_0$ approx (STE) |
| Shrinkage Bias | High | Low | None | None |
| Training Cost | Low | High | Medium | Low |
| Pareto Frontier | Baseline | High | High | Highest |
.15
8. Scaling Laws for Dictionary Learning
Just as LLMs follow scaling laws, SAEs exhibit predictable power-law relationships between compute, dictionary size, and reconstruction error. Understanding these laws is crucial for efficiently allocating resources when training on models the size of GPT-4.
8.1 The Power Laws of Reconstruction
OpenAI and Anthropic researchers have established that the reconstruction Mean Squared Error (MSE) scales as a power law with the SAE width ($N$) and compute budget ($C$).
$$L(C) \propto C^{-\alpha}$$
- Implication: There are diminishing returns. To halve the reconstruction error, one must increase the dictionary size and training tokens exponentially.
- Trade-off: There is a fundamental trade-off between dictionary size and sparsity. For a fixed reconstruction error target, a larger dictionary allows for sparser representations (fewer active features per token). This is because a larger dictionary contains more specific features that better match the input, requiring fewer “corrective” features to reconstruct the residual.13
8.2 The Cost of the Subject Model
Crucially, the difficulty of the task scales with the subject model’s capabilities. Interpreting a stronger model (like GPT-4) requires a significantly larger SAE than interpreting a weaker model (like GPT-2) to achieve the same level of explained variance. The scaling exponent worsens as the subject model gets “smarter,” suggesting that frontier models pack information more densely, making decompression harder.13
9. Automated Interpretability Pipelines
With 30 million features in Gemma Scope and millions in Claude, manual inspection is impossible. The field has turned to Automated Interpretability—using LLMs to interpret LLMs.
9.1 The Auto-Interp Protocol
Organizations like OpenAI and EleutherAI have standardized a pipeline:
- Feature Selection: Choose a feature from the SAE.
- Max-Activating Examples: Retrieve text snippets where this feature fires most strongly.
- Explanation Generation: Feed these snippets to a strong “Explainer” LLM (e.g., GPT-4 or Claude 3.5) and ask: “What concept do these snippets have in common?”
- Verification (Simulation): Take the generated explanation and a new set of text snippets. Ask the “Simulator” LLM to predict the feature’s activation based solely on the explanation.
- Scoring: Correlate the Simulator’s predictions with the actual feature activations. A high correlation (score closer to 1.0) implies the explanation is accurate and the feature is monosemantic.31
9.2 Scoring Metrics: Detection vs. Fuzzing
- Detection Score: Measures the ability to predict if a feature will fire.
- Fuzzing Score: Measures the ability to predict exactly which tokens will fire.
Recent results show that Deep SAEs (SAEs with multiple layers) achieve interpretability scores comparable to shallow SAEs, challenging the assumption that deeper architectures are inherently less interpretable.34 However, the cost is prohibitive; running this pipeline on millions of features can cost hundreds of thousands of dollars in API credits.33
10. The Dark Matter and The Limits of the Lens
Despite the triumphs, the “Scaling Monosemanticity” agenda faces profound scientific challenges. The decomposition is imperfect, and critical information remains hidden in the “dark matter” of the model.
10.1 The Problem of Dark Matter
SAEs rarely reconstruct the model’s activation perfectly. There is always a residual error—unexplained variance.
- Magnitude: In many layers, the SAE explains only 60-80% of the variance. The remaining 20-40% is “dark matter.”
- Composition: Research suggests that over 90% of the norm of this dark matter is linearly predictable from the input.5 This is damning: it implies that the SAE is missing “dense” features—information that is widely distributed and does not fit the sparsity constraint.
- Risk: If safety-critical information (e.g., a subtle deceptive intent) is encoded in this dense “dark matter,” SAE-based monitoring will fail to detect it. The SAE acts as a lossy compression algorithm that might discard the very signal we need to see.
10.2 Feature Absorption
A specific failure mode is Feature Absorption. This occurs when a specific feature (e.g., “Snake”) absorbs the activation of a more general feature (e.g., “Starts with S”) that should also logically fire.
- Mechanism: To minimize the L0 penalty (number of active features), the SAE prefers to activate one feature (“Snake”) rather than two (“Snake” + “Starts with S”).
- Consequence: The “Starts with S” feature becomes unreliable—it has “holes” in its activation pattern wherever a more specific feature took over. This degrades trust in the completeness of the feature map.37
10.3 The Critique of Downstream Utility
Perhaps the most severe critique comes from Google DeepMind’s safety team. In experiments designed to detect harmful intent (e.g., jailbreaks), probes trained on SAE features performed worse than simple linear probes trained on the raw dense activations.
- OOD Generalization: While SAE probes worked on standard data, they failed to generalize to Out-Of-Distribution (OOD) attacks. Linear probes remained robust.
- Conclusion: This suggests that “harmful intent” might be a dense, diffuse concept that resists sparse decomposition. By forcing it into a sparse basis, the SAE discards redundant information that is crucial for robustness. This has led some to argue that SAEs are better for discovery (finding unknown concepts) than for monitoring (reliable detection of known risks).40
11. Beyond Language: Biological Models
The principles of scaling monosemanticity are universal to high-dimensional learning systems. In Protein Language Models (like ESM2), SAEs are revolutionizing computational biology.
- Biological Ground Truth: Unlike in language, where “truth” is subjective, biology has ground truth. SAE features in ESM2 have been found to align with specific Gene Ontology (GO) terms—protein families, binding sites, and catalytic functions.
- Superiority: These sparse features align significantly better with biological labels than the raw neurons of the model. This validates the SAE approach as a general method for decoding learned representations in scientific AI.22
12. Conclusion: The Glass Box Horizon
The effort to scale monosemanticity has successfully transformed the “black box” of frontier models into a “gray box.” We have proven that the Superposition Hypothesis holds at scale: the inscrutable vectors of Claude 3 Sonnet and Gemma 2 can be decomposed into millions of intelligible features. We have mapped concrete entities, abstract vulnerabilities, and behavioral traits. We have proven causality through steering experiments like Golden Gate Claude, demonstrating that we can reach into the mind of the model and surgically alter its identity.
However, the “glass box” remains distant. The persistence of dark matter and the phenomenon of feature absorption indicate that our sparse dictionary is an incomplete map of the territory. The failure of SAEs to outperform linear probes in OOD safety tasks serves as a necessary check on optimism—sparsity is a powerful inductive bias for interpretability, but it may come at the cost of robustness.
The future of this field lies in Circuit Analysis—moving beyond isolated features to mapping the causal algorithms (circuits) that connect them.42 It lies in Transcoders, which replace entire opaque layers with interpretable sparse mappings.22 And it lies in the continued scaling of these methods, chasing the power laws of reconstruction in the hope that, with enough compute and the right architecture, the ghost in the machine will finally speak clearly.
Works cited
- A “Scaling Monosemanticity” Explainer – LessWrong, accessed on December 22, 2025, https://www.lesswrong.com/posts/wg6E3oJJrNnmJezNz/a-scaling-monosemanticity-explainer
- Understanding neural networks through sparse circuits – OpenAI, accessed on December 22, 2025, https://openai.com/index/understanding-neural-networks-through-sparse-circuits/
- Towards Monosemanticity: A Step Towards Understanding Large Language Models | by Anish Dubey | TDS Archive | Medium, accessed on December 22, 2025, https://medium.com/data-science/towards-monosemanticity-a-step-towards-understanding-large-language-models-e7b88380d7b3
- Claude Sonnet 4.5 System Card – Anthropic, accessed on December 22, 2025, https://www.anthropic.com/claude-sonnet-4-5-system-card
- decomposing the dark matter of sparse autoencoders – arXiv, accessed on December 22, 2025, https://arxiv.org/pdf/2410.14670
- Towards Monosemanticity: Decomposing Language Models With …, accessed on December 22, 2025, https://transformer-circuits.pub/2023/monosemantic-features
- Understanding Anthropic’s Golden Gate Claude | by Jonathan Davis – Medium, accessed on December 22, 2025, https://medium.com/@jonnyndavis/understanding-anthropics-golden-gate-claude-150f9653bf75
- Sparse Autoencoders Find Highly Interpretable Features in Language Models, accessed on December 22, 2025, https://openreview.net/forum?id=F76bwRSLeK
- arXiv:2406.17969v2 [cs.CL] 15 Oct 2024, accessed on December 22, 2025, https://arxiv.org/pdf/2406.17969
- Understanding the “Scaling of Monosemanticity” in AI Models: A Comprehensive Analysis | by Milani Mcgraw | The Deep Hub | Medium, accessed on December 22, 2025, https://medium.com/thedeephub/understanding-the-scaling-of-monosemanticity-in-ai-models-a-comprehensive-analysis-f72818fa44ca
- Towards Monosemanticity: Decomposing Language Models With Dictionary Learning, accessed on December 22, 2025, https://www.anthropic.com/research/towards-monosemanticity-decomposing-language-models-with-dictionary-learning
- Scaling Monosemanticity: Anthropic’s One Step Towards Interpretable & Manipulable LLMs, accessed on December 22, 2025, https://towardsdatascience.com/scaling-monosemanticity-anthropics-one-step-towards-interpretable-manipulable-llms-4b9403c4341e/
- Scaling and evaluating sparse autoencoders | OpenAI, accessed on December 22, 2025, https://cdn.openai.com/papers/sparse-autoencoders.pdf
- Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders – arXiv, accessed on December 22, 2025, https://arxiv.org/pdf/2407.14435
- Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders, accessed on December 22, 2025, https://arxiv.org/html/2407.14435v1
- Improving Dictionary Learning with Gated Sparse Autoencoders – Google DeepMind, accessed on December 22, 2025, https://deepmind.google/research/publications/improving-dictionary-learning-with-gated-sparse-autoencoders/
- JumpReLU SAEs + Early Access to Gemma 2 SAEs – LessWrong, accessed on December 22, 2025, https://www.lesswrong.com/posts/wZqqQysfLrt2CFx4T/jumprelu-saes-early-access-to-gemma-2-saes
- Open Sparse Autoencoders Everywhere: The Ambitious Vision of DeepMind’s Gemma Scope | Synced, accessed on December 22, 2025, https://syncedreview.com/2024/08/26/open-sparse-autoencoders-everywhere-the-ambitious-vision-of-deepminds-gemma-scope/
- The engineering challenges of scaling interpretability – Anthropic, accessed on December 22, 2025, https://www.anthropic.com/research/engineering-challenges-interpretability
- Gemma Scope 2: Helping the AI Safety Community Deepen Understanding of Complex Language Model Behavior – Google DeepMind, accessed on December 22, 2025, https://deepmind.google/blog/gemma-scope-2-helping-the-ai-safety-community-deepen-understanding-of-complex-language-model-behavior/
- Golden Gate Claude \ Anthropic, accessed on December 22, 2025, https://www.anthropic.com/news/golden-gate-claude
- Sparse autoencoders uncover biologically interpretable features in protein language model representations – NIH, accessed on December 22, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC12403088/
- Persona vectors: Monitoring and controlling character traits in language models – Anthropic, accessed on December 22, 2025, https://www.anthropic.com/research/persona-vectors
- Evaluating feature steering: A case study in mitigating social biases – Anthropic, accessed on December 22, 2025, https://www.anthropic.com/research/evaluating-feature-steering
- The Geometry of Concepts: Sparse Autoencoder Feature Structure – MDPI, accessed on December 22, 2025, https://www.mdpi.com/1099-4300/27/4/344
- Appendix to “Evaluating Feature Steering: A Case Study in Mitigating Social Biases”, accessed on December 22, 2025, https://assets.anthropic.com/m/6a464113e31f55d5/original/Appendix-to-Evaluating-Feature-Steering-A-Case-Study-in-Mitigating-Social-Biases.pdf
- Gemma Scope: helping the safety community shed light on the inner …, accessed on December 22, 2025, https://deepmind.google/blog/gemma-scope-helping-the-safety-community-shed-light-on-the-inner-workings-of-language-models/
- JUMPING AHEAD: IMPROVING RECONSTRUCTION FI- DELITY WITH JUMPRELU SPARSE AUTOENCODERS – OpenReview, accessed on December 22, 2025, https://openreview.net/pdf?id=mMPaQzgzAN
- Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control, accessed on December 22, 2025, https://openreview.net/forum?id=1Njl73JKjB
- Circuits Updates – June 2024, accessed on December 22, 2025, https://transformer-circuits.pub/2024/june-update/index.html
- Automatically Interpreting Millions of Features in Large Language Models – arXiv, accessed on December 22, 2025, https://arxiv.org/html/2410.13928v2
- SPARSE AUTOENCODERS FIND HIGHLY INTER- PRETABLE FEATURES IN LANGUAGE MODELS – ICLR Proceedings, accessed on December 22, 2025, https://proceedings.iclr.cc/paper_files/paper/2024/file/1fa1ab11f4bd5f94b2ec20e794dbfa3b-Paper-Conference.pdf
- Open Source Automated Interpretability for Sparse Autoencoder Features | EleutherAI Blog, accessed on December 22, 2025, https://blog.eleuther.ai/autointerp/
- Deep sparse autoencoders yield interpretable features too – AI Alignment Forum, accessed on December 22, 2025, https://www.alignmentforum.org/posts/tLCBJn3NcSNzi5xng/deep-sparse-autoencoders-yield-interpretable-features-too
- Introducing Bloom: an open source tool for automated behavioral evaluations – Anthropic, accessed on December 22, 2025, https://www.anthropic.com/research/bloom
- arXiv:2410.14670v1 [cs.LG] 18 Oct 2024, accessed on December 22, 2025, https://arxiv.org/abs/2410.14670
- A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders, accessed on December 22, 2025, https://openreview.net/forum?id=LC2KxRwC3n
- Studying Feature Splitting and Absorption in Sparse Autoencoders – arXiv, accessed on December 22, 2025, https://arxiv.org/html/2409.14507v6
- [Paper] A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders – LessWrong, accessed on December 22, 2025, https://www.lesswrong.com/posts/3zBsxeZzd3cvuueMJ/paper-a-is-for-absorption-studying-feature-splitting-and
- Negative Results for Sparse Autoencoders On Downstream Tasks …, accessed on December 22, 2025, https://deepmindsafetyresearch.medium.com/negative-results-for-sparse-autoencoders-on-downstream-tasks-and-deprioritising-sae-research-6cadcfc125b9
- Use Sparse Autoencoders to Discover Unknown Concepts, Not to Act on Known Concepts, accessed on December 22, 2025, https://arxiv.org/html/2506.23845v1
- [2405.12522] Sparse Autoencoders Enable Scalable and Reliable Circuit Identification in Language Models – arXiv, accessed on December 22, 2025, https://arxiv.org/abs/2405.12522
- Sparse Autoencoders: Future Work – AI Alignment Forum, accessed on December 22, 2025, https://www.alignmentforum.org/posts/CkFBMG6A9ytkiXBDM/sparse-autoencoders-future-work