# Superposition hypothesis in neural networks

## Introduction to Mechanistic Interpretability

The rapid proliferation and scaling of deep learning architectures have produced artificial intelligence systems of unprecedented capability, yet the internal mechanisms governing their computations remain largely opaque. This opacity presents substantial challenges for safety, alignment, and reliability, driving the emergence of mechanistic interpretability. Mechanistic interpretability is a dedicated research domain aiming to reverse-engineer the opaque, high-dimensional weight matrices of trained neural networks into discrete, human-understandable algorithms and computational circuits [cite: 1, 2, 3]. 

The historical trajectory of this field initially centered on vision models during the mid-2010s. Researchers operating within this early paradigm frequently hypothesized that neural networks functioned via a strict one-to-one mapping, wherein individual neurons corresponded to distinct, identifiable concepts, such as a "curve detector" or a highly specific "Jennifer Aniston" neuron [cite: 4, 5]. However, as the focus of interpretability research transitioned toward large language models (LLMs) and advanced multimodal systems, this foundational assumption of monosemanticity began to collapse. 

Empirical investigations consistently revealed that the vast majority of neurons in complex networks did not respond to single, isolated features. Instead, individual neurons activated in response to a chaotic pastiche of multiple, seemingly unrelated concepts—a phenomenon termed polysemanticity [cite: 5, 6, 7]. A single neuron might activate strongly in the presence of images of cats, the structural syntax of Python code, and specific textual descriptions of vehicles [cite: 8, 9]. Because tracing the activation of a polysemantic neuron provides inherently ambiguous information regarding what concept the network is actively processing, polysemanticity emerged as a critical roadblock to mapping the computational circuits of advanced AI systems [cite: 10].

## The Framework of the Superposition Hypothesis

To resolve the paradox of polysemantic neurons, researchers formulated the superposition hypothesis. The superposition hypothesis posits that neural networks are fundamentally unconstrained by the finite number of neurons (or dimensions) they possess within any given layer. Instead of enforcing a strict one-neuron-per-feature architecture, networks learn to represent an exponentially larger set of features by embedding them as vectors within a shared, high-dimensional activation space [cite: 5, 11, 12, 13].

[image delta #1, 0 bytes]

 

Under this framework, a network representing features in superposition does not allocate individual neurons to individual concepts. Instead, each feature corresponds to a specific direction across the entire activation space. Because these directions overlap, each neuron contributes to the representation of multiple features, and each feature is distributed as a pattern across multiple neurons [cite: 5]. When an input contains a specific feature, the network activates the corresponding vector direction. If multiple features co-occur, the resulting activation vector is a linear sum of those individual feature directions [cite: 5, 13]. 

Superposition is not a failure of the optimization process or an architectural flaw; rather, it is a mathematically optimal, compressed representation strategy naturally discovered by gradient descent. By multiplexing features through a shared vector space, neural networks effectively simulate the representational capacity of a hypothetically much larger, highly sparse network [cite: 5, 11, 14]. 



### High-Dimensional Geometry and Capacity Limits

The theoretical viability of the superposition hypothesis is deeply anchored in the geometry of high-dimensional spaces, a property most clearly articulated through the Johnson-Lindenstrauss lemma [cite: 10, 15, 16]. In low-dimensional geometry, the number of strictly orthogonal directions is exactly equal to the number of dimensions. However, as dimensionality scales upward into the thousands—typical for the hidden layers of modern LLMs—the geometry of the space shifts dramatically.

High-dimensional spaces contain an exponentially larger number of *nearly* orthogonal directions compared to strictly orthogonal ones [cite: 5, 17]. If a network is willing to tolerate a minuscule amount of mathematical noise or interference between vectors, it can pack an enormous volume of concepts into a restricted bottleneck. The Johnson-Lindenstrauss lemma formally establishes that a set of $m$ points can be linearly projected into a lower-dimensional space of size $O(\log m)$ while preserving the Euclidean distances between all pairs of points within a small error margin [cite: 15, 18]. 

For neural networks, this mathematical property translates into a staggering upper bound for passive representation. Assuming a neural network layer possesses $n$ neurons, the Johnson-Lindenstrauss lemma dictates that the layer can theoretically encode up to $2^{O(n)}$ passive features using nearly orthogonal vectors [cite: 10, 16]. This exponential capacity demonstrates why polysemanticity is an inevitable consequence of parameter-efficient training: to access this massive representational volume, the network must utilize vectors that project across multiple standard neuron axes simultaneously [cite: 8, 17].

### Sparsity and the Interference Threshold

While the mathematical capacity of high-dimensional space enables superposition, the strategy introduces an inherent, inescapable cost: interference. Because the feature vectors are nearly orthogonal rather than strictly orthogonal, they are not mathematically isolated. When a network activates a specific feature vector, it inevitably generates a non-zero dot product—or projection noise—along the dimensions corresponding to other superimposed features [cite: 11, 12, 19]. 

The feasibility of maintaining accurate representations in the presence of this interference relies entirely on the natural sparsity of the input data [cite: 20, 21]. In the context of language or visual processing, the true environment contains millions of potential features, but only a microscopic fraction of these features are actively relevant to any single input instance [cite: 5, 22]. 

If the data distribution were dense—meaning hundreds or thousands of features co-occurred constantly—the additive interference from the non-orthogonal projections would compound rapidly [cite: 20, 21]. This noise would operate like a random walk, quickly overwhelming the true signal and rendering the network's activations unintelligible. However, because features are sparsely distributed (often possessing an activation probability well below 0.05), the likelihood of numerous overlapping features firing simultaneously remains exceptionally low [cite: 20]. 

Mathematical models indicate that if $l$ features are active simultaneously within a network of $n$ dimensions attempting to store $m$ total features, the average interference noise scales proportionally to $\epsilon \sqrt{l}$, where $\epsilon$ dictates the baseline overlap between vectors [cite: 17]. For the network to reliably decode the signal, the activation strength of the true feature must remain significantly higher than the compounding $\sqrt{l}$ interference noise generated by the other sparsely active elements [cite: 17]. 

## Architectural Influences on Feature Representation

The specific mechanisms by which neural networks deploy superposition are heavily constrained by their underlying architectures. A critical concept in understanding why features map to neurons in certain layers but remain completely distributed in others is the distinction between privileged and non-privileged bases.

### Privileged versus Non-Privileged Bases

In linear algebra, a vector space has no intrinsic orientation; any orthogonal rotation of a coordinate system represents the data with perfect mathematical equivalence. However, within the architecture of a neural network, certain operations break this rotational invariance, establishing a "privileged basis" [cite: 12, 23].

A privileged basis exists when the network architecture inherently distinguishes the standard basis directions—the individual neurons themselves—from arbitrary directions in the activation space [cite: 12, 23, 24]. The primary mechanism that establishes a privileged basis is the application of element-wise non-linear activation functions, such as Rectified Linear Units (ReLU) or Gaussian Error Linear Units (GELU) [cite: 24, 25]. 

Because a ReLU function operates on each neuron individually—setting all negative values on that specific axis to zero while leaving positive values unchanged—the operation is not rotationally invariant. If the entire activation space is mathematically rotated, the ReLU will truncate entirely different components of the representation, destroying the encoded information [cite: 25]. Consequently, the network faces a strong optimization incentive (inductive bias) to align its most critical, frequently used features directly with these individual neuron axes. When features align perfectly with the standard basis, the result is the emergence of monosemantic neurons [cite: 12, 24, 26]. 

Conversely, a non-privileged basis lacks any architectural mechanism that treats specific axes differently from others. The residual stream in transformer models, as well as the keys, queries, and values within attention heads, operate as non-privileged spaces because they rely primarily on linear transformations [cite: 23, 24, 27]. In a non-privileged space, the optimization process has no incentive to align semantic features with the arbitrary axes defined by the neurons [cite: 25]. As a result, the network distributes information redundantly across all available dimensions. In these spaces, polysemanticity occurs naturally regardless of capacity constraints, simply because the feature vectors exist at arbitrary, unaligned rotations relative to the standard basis [cite: 12, 26].

| Architectural Domain | Basis Type | Primary Mechanisms | Feature Alignment Characteristics | Implications for Polysemanticity |
| :--- | :--- | :--- | :--- | :--- |
| **MLP Hidden Layers** | Privileged Basis | Element-wise non-linearities (ReLU, GELU); L1 regularization. | Strong inductive bias to align highly important features perfectly with individual neuron axes. | Polysemanticity primarily emerges due to capacity constraints (necessary superposition) forcing multiple features onto the same axes. |
| **Residual Streams** | Non-Privileged Basis | Purely linear transformations; matrix multiplication bottlenecks. | No optimization incentive for axis alignment. Features are distributed arbitrarily across the vector space. | Polysemanticity is the default state due to arbitrary vector rotation, even if the number of neurons equals the number of features. |

### The Mechanics of Non-Linear Interference Filtering

The establishment of a privileged basis via non-linearities does more than just encourage feature alignment; it provides the fundamental mechanism that makes superposition computationally viable. 

In a strictly linear model lacking activation functions, interference is entirely unmitigated. The interference is mathematically defined as the sum of squared dot products between all pairs of feature vectors: $\sum_{i \neq j} |W_i \cdot W_j|^2$ [cite: 11]. Because any overlap strictly degrades the accuracy of the representation, linear models actively avoid superposition. When faced with more features than dimensions, a linear model will simply drop the least important features, dedicating its dimensions exclusively to orthogonal representations of the most critical variables [cite: 21, 28].

However, the introduction of non-linear thresholding, such as a ReLU, alters the cost-benefit analysis of interference. A ReLU acts as an active "interference filter" [cite: 11, 21, 29]. Models learn to apply slightly negative bias weights to their neurons. When sparse, nearly orthogonal features generate small amounts of cross-talk projection noise, this negative bias shifts the sum of the noise below zero. The ReLU then maps these negative values precisely to zero, completely eliminating the interference from the forward pass [cite: 11, 21, 28]. This dynamic allows the network to tolerate the non-orthogonality of superposition because the activation function mathematically silences the resulting baseline static, protecting the high-magnitude true signals [cite: 11, 12].

## The Geometric Organization of Superimposed Features

When neural networks rely on non-linear filtering to manage interference, the features embedded in superposition do not scatter randomly throughout the activation space. Empirical investigations using idealized toy models reveal that features self-organize into highly structured geometric motifs dictated by their relative sparsity and importance [cite: 11, 28].

### Phase Changes and Geometric Polytopes

Research into feature organization demonstrates that superposition operates via first-order phase changes [cite: 11, 30]. In idealized models where input features are statistically independent and uniformly sparse, the optimization landscape is characterized by distinct, discontinuous thresholds [cite: 11, 30]. As the sparsity of the data increases—making it safer for the model to pack features more densely—the network's optimal weight configuration abruptly transitions between different geometric shapes [cite: 7, 11].

A network effectively assigns fractions of a hidden dimension to each feature, calculated as the squared norm of the feature's weight vector divided by the sum of its squared projections onto all other features: $D_i = ||W_i||^2 / \sum (\hat{W_i} \cdot W_j)^2$ [cite: 11]. As capacity pressures shift, the network gravitates toward highly stable mathematical "sticky points" corresponding to specific fractional dimensionalities [cite: 11]. 

When allocating dimensions to features, networks generate distinct polytopes:
*   **Dimensionality 1:** A dedicated, orthogonal dimension for a highly important feature (no superposition).
*   **Dimensionality 1/2:** Features organize into antipodal pairs, occupying the exact same dimension but pointing in opposite directions, relying on the ReLU to filter out the negative activation of the opposing feature [cite: 11, 28].
*   **Dimensionality 2/3:** Three features distribute themselves evenly across a 2D plane, forming an equilateral triangle [cite: 11].
*   **Dimensionality 3/4:** Features arrange into a three-dimensional tetrahedron [cite: 11].
*   **Dimensionality 2/5:** Five features distribute into a regular pentagon [cite: 11, 28].

These geometric configurations demonstrate that superposition is a highly structured, mathematically rigorous strategy for minimizing mutual interference while maximizing the number of representable concepts [cite: 11, 28].

### Constructive Interference and Bag-of-Words Superposition

While idealized toy models assume that features are statistically independent, real-world datasets—such as the massive corpora of internet text ingested by frontier LLMs—are defined by highly correlated features [cite: 29, 31]. Recent studies analyzing feature geometry in the context of realistic data statistics indicate that when features are correlated, interference ceases to be purely a detrimental noise source requiring ReLU filtration [cite: 29].

Researchers have documented the phenomenon of Bag-of-Words Superposition (BOWS), a regime in which networks leverage the covariance matrix of the training data to generate constructive interference [cite: 29, 31]. If two features frequently co-occur (e.g., the concepts of "snow" and "cold"), the network places them in the activation space such that their vectors positively align [cite: 29]. When an input contains both features, their respective projections constructively sum together, amplifying the true signal while naturally overwhelming the noise from unrelated features [cite: 29]. 

This reliance on constructive alignment allows the model to achieve highly efficient reconstructions of the data with lower weight norms and tighter rank requirements [cite: 29]. Consequently, rather than forming isolated regular polytopes, correlated features in real LLMs form semantic clusters and cyclical structures that visually reflect their linguistic and contextual relationships [cite: 29, 32]. These findings confirm that the standard geometric picture of superposition is dynamically modulated by the covariance statistics of the underlying dataset [cite: 29].

## Alternative Origins: Necessary Compression versus Incidental Polysemanticity

The standard superposition hypothesis frames polysemanticity as an optimal, necessary solution to the problem of high-dimensional compression [cite: 5, 9]. However, an ongoing debate within the mechanistic interpretability community highlights an alternative paradigm: incidental polysemanticity [cite: 33, 34].

The classic superposition hypothesis—termed "necessary polysemanticity"—asserts that neurons become polysemantic exclusively because the network must compress a vast number of environmental features into a restrictive representational bottleneck [cite: 9, 34]. In contrast, research into incidental polysemanticity demonstrates that neurons can activate for multiple completely unrelated concepts even when the network possesses ample capacity to represent every single feature perfectly orthogonally [cite: 34, 35].

Incidental polysemanticity is an artifact of the network's training dynamics, specifically random initialization paired with winner-take-all optimization incentives [cite: 34, 35]. At the beginning of the training process, neural weights are assigned randomly. Purely by stochastic chance, a single, uninitialized neuron may exhibit a slight mathematical correlation with two entirely unrelated, independent features within the dataset—for example, the visual concept of a "dog" and the concept of an "airplane" [cite: 9, 34]. 

If the network utilizes regularization techniques that enforce sparsity, such as L1 regularization or dropout noise, the learning algorithm establishes a winner-take-all dynamic [cite: 34, 35]. Instead of slowly building a new, dedicated neuron for the "airplane" feature, gradient descent simply amplifies the pre-existing, slightly advantageous correlation in the neuron that is already partially tracking the "dog" feature [cite: 9, 34]. Over thousands of training steps, this overlap becomes permanent. 

The resulting polysemantic neuron is not a product of geometric optimization or a lack of capacity, but rather a historical accident cemented by gradient descent [cite: 9, 34]. Studies using toy models have shown that networks exhibiting incidental polysemanticity can achieve perfect task accuracy, indicating that this phenomenon does not inherently degrade performance, but severely damages the post-hoc interpretability of the model [cite: 35]. This duality presents a profound challenge for interpretability researchers, as any attempt to reverse-engineer a network must distinguish between features compressed out of mathematical necessity and features entangled purely by the chaos of initialization.

| Polysemanticity Origin | Primary Mechanism | Capacity Condition | Implications for Interpretability |
| :--- | :--- | :--- | :--- |
| **Necessary (Superposition)** | Intentional packing of features into nearly-orthogonal vectors to maximize data coverage and representational efficiency. | Occurs strictly when the environment contains exponentially more features than the network possesses neurons. | Requires advanced dictionary learning (e.g., Sparse Autoencoders) to project the compressed vectors back into an interpretable overcomplete basis. |
| **Incidental** | Random initialization biases amplified by gradient descent, L1 regularization, dropout, and winner-take-all dynamics. | Can occur even when the network possesses ample neurons to represent all features perfectly monosemantically. | Suggests that some polysemanticity is an optimization artifact that could be prevented architecturally without sacrificing capacity, potentially through training interventions. |

## The Complexity of Active Computation in Superposition

Much of the foundational literature surrounding superposition focuses on representational capacity—the ability of a network bottleneck to passively store and retrieve information without catastrophic data loss [cite: 36, 37, 38]. However, frontier AI systems do not merely store features; they actively perform complex non-linear operations, logical reasoning, and algorithmic transformations on the data [cite: 16, 39]. Recent theoretical research demonstrates that computation in superposition operates under significantly stricter mathematical boundaries than passive representation [cite: 16, 18].

While the Johnson-Lindenstrauss lemma allows $n$ neurons to passively represent an exponential $2^{O(n)}$ number of features, active computation fundamentally shrinks this capacity [cite: 16, 18]. When a neural network must ingest superimposed inputs and explicitly compute combinatorial logical outputs, the requisite parameter counts scale dramatically [cite: 39, 40]. 

Mathematical models investigating computation in superposition have focused on fundamental operations, such as emulating a Universal AND (U-AND) boolean circuit that computes the pairwise logical ANDs of a set of input features [cite: 39]. Theoretical proofs establish that to compute $m'$ output features from superimposed inputs, a network fundamentally requires at least $\Omega(m' \log m')$ parameters and $\Omega(\sqrt{m' \log m'})$ neurons [cite: 16, 18]. 

Despite these stricter bounds, neural networks remain capable of sub-linear computation. Researchers have successfully constructed specific Multi-Layer Perceptron (MLP) architectures capable of emulating the U-AND circuit for $m$ features using roughly $\tilde{O}(m^{2/3})$ neurons, effectively processing the logical operations while the inputs remain heavily compressed in superposition [cite: 36, 37, 39]. Furthermore, deep fully-connected networks with a width of $d$ can emulate sparse boolean circuits with a width of $\tilde{O}(d^{1.5})$ across arbitrary polynomial depths [cite: 37]. Crucially, probabilistic analysis confirms that randomly initialized networks are highly likely to naturally linearly represent these U-AND circuits if the network width is sufficiently large, suggesting that real-world models naturally discover and utilize these compressed computational pathways [cite: 36, 39].



These findings possess monumental implications for the interpretation of advanced AI systems. They formally define the "exponential gap" between simply encoding features and executing logical reasoning [cite: 16, 18].

[image delta #2, 0 bytes]

 If researchers aim to quantify the true algorithmic capabilities of frontier LLMs, the total parameter count serves as a significantly tighter constraint on computational potential than the highly permissive bounds of passive representation [cite: 16].

## Feature Disentanglement via Sparse Autoencoders

Because features exist in dense geometric superposition, analyzing the raw weights or activations of an LLM's internal layers yields an incomprehensible mixture of signals. To isolate distinct concepts, the mechanistic interpretability community has widely adopted dictionary learning techniques, predominantly utilizing Sparse Autoencoders (SAEs) [cite: 1, 14, 41].

SAEs are a specialized neural architecture appended to the frozen, trained layers of a target model. Unlike classical autoencoders—which are conventionally utilized to compress high-dimensional data into a restrictive, lower-dimensional bottleneck—SAEs perform the inverse operation [cite: 6, 14]. An SAE takes the low-dimensional, highly polysemantic activations of an LLM layer and projects them outward into a massive, overcomplete hidden dimension [cite: 1]. The objective is to reconstruct the original activation while strictly enforcing sparsity upon the overcomplete dimension.

This sparsity is typically achieved by incorporating an L1 regularization penalty ($\lambda \sum |w|$) into the autoencoder's loss function during training [cite: 6, 14]. The optimization process must balance reconstruction fidelity against the L1 penalty, forcing the SAE to activate only a sparse handful of its neurons for any given input token [cite: 6, 20]. This induced sparsity mirrors the theoretical sparsity of the natural data distribution that originally necessitated superposition. As a result, the SAE learns a vast dictionary of distinct vector directions, where each direction corresponds to a highly specific, monosemantic feature [cite: 14]. 

The application of SAEs has revolutionized the ability to map the internal cognitive state of frontier models. Major AI research institutions have successfully trained massive SAEs on state-of-the-art models, extracting tens of millions of distinct features from architectures like Claude 3 Sonnet and GPT-4 [cite: 1, 14, 22, 42]. The extracted dictionaries encompass a vast semantic spectrum, ranging from basic syntactic detectors to abstract conceptual representations such as "Golden Gate Bridge," "social bias," "scam email drafting," and "deceptive language" [cite: 14, 42, 43]. 

Furthermore, applying advanced representational similarity metrics (such as Singular Value Canonical Correlation Analysis) to SAE-extracted features has revealed significant evidence for "feature universality" [cite: 41]. Feature universality implies that entirely different model architectures, such as Meta's Llama and Google's Gemma, independently converge on highly similar geometric representations for identical semantic concepts within their latent spaces [cite: 41, 44]. However, researchers acknowledge that fully mapping the capabilities of frontier LLMs may eventually require SAEs scaled to billions or trillions of features, a computationally prohibitive threshold given current methodologies [cite: 22, 42].

## Applications in Model Steering and Alignment

The extraction of monosemantic features via SAEs extends far beyond observational analysis; it provides a direct mechanism for causal intervention. By mapping specific conceptual features back to the model's original latent space, researchers can artificially manipulate the network's forward pass, a technique known as "activation steering" or "feature steering" [cite: 14, 43, 45].

### Hallucination Reduction and Query Enrichment

A highly promising application of SAE steering is the real-time mitigation of AI hallucinations. Both Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) frequently generate plausible but factually incorrect assertions due to misaligned or entangled internal representations [cite: 44, 46, 47, 48]. 

Recent developments, such as the Sparse Autoencoder-Based Framework for Robust Query Enrichment (SAFE), demonstrate the capacity of steering to refine inputs dynamically without retraining the base model. The SAFE architecture employs a two-stage process. First, it detects potential hallucinations during generation using entropy-based uncertainty estimation [cite: 44, 49]. If high uncertainty is detected, the framework queries an attached SAE to extract semantically grounded, monosemantic features related to the input context. These features are utilized to enrich the query, actively steering the model toward relevant factual directions while suppressing extraneous, hallucination-inducing noise [cite: 44, 50]. Empirical evaluations on benchmark QA datasets (e.g., TruthfulQA) confirm that SAFE significantly reduces hallucination rates, achieving response accuracy improvements of up to 29.45% [cite: 44, 49].

Similar interventions targeting multi-modal architectures—such as Contrastive Neuron Steering (CNS) and Steering LVLMs via SAE Latent Directions (SSL)—operate by identifying the specific SAE latents associated with "faithful" content versus "hallucinatory" content. By artificially amplifying the true visual directions and suppressing the spurious noise latents identified via the SAE, models demonstrate marked improvements in factual consistency regarding image inputs [cite: 47, 48].

### Jailbreak Defense and the Capability Trade-Off

In the context of AI safety and alignment, SAE steering provides a powerful mechanism for enforcing behavioral guardrails. Safety interventions can identify the specific features that mediate a model's refusal behaviors when faced with harmful or malicious prompts [cite: 51, 52]. By artificially clamping the activation values of these "refusal features" to high, static thresholds during inference, models become significantly more robust against sophisticated single-turn and multi-turn jailbreak attempts [cite: 51, 52]. 

However, manipulating features extracted from a heavily superimposed state introduces profound complexities regarding global model performance. Empirical research highlights a severe, fundamental trade-off between SAE-enforced safety steering and general computational capabilities [cite: 51, 52]. In detailed evaluations using the Llama 3 8B architecture, researchers discovered that while steering specific refusal features (e.g., amplifying "Feature 9000" or suppressing "Feature 43692") improved safety metrics on Air Bench by 8.8% to 10.0%, these interventions triggered catastrophic capability degradation [cite: 53]. Associated utility scores on benchmarks like AlpacaEval dropped precipitously (up to 16.3%), demonstrating systematic degradation in factual recall, logical reasoning, and general instruction-following, even on completely benign, non-harmful inputs [cite: 52, 53]. This suggests that safety-relevant features remain deeply entangled with general linguistic capabilities within the superposition manifold, complicating isolated interventions [cite: 51].

### Endogenous Steering Resistance

A secondary, highly complex challenge in feature steering is the phenomenon of Endogenous Steering Resistance (ESR), observed primarily in high-parameter frontier models. When researchers inject SAE latents to deliberately force a model off-topic or modify its personality traits, sufficiently large models demonstrate an innate, internal self-correction capability [cite: 54]. 

In targeted experiments on the 70-billion parameter Llama 3.3 model, the network successfully recognized the artificial activation shift generated by the steering vector. The model would halt its generation mid-sequence (e.g., generating text such as "wait, that's not right") and actively attempt to return to the original prompt trajectory, demonstrating an ESR success rate of 3.8% [cite: 54]. Through SAE latent analysis, researchers identified 26 specific latents that fired differentially during this off-topic detection phase; zero-ablating these specific latents successfully reduced the model's self-correction rate by 25% [cite: 54]. 

While ESR serves as an excellent natural defense against malicious adversarial manipulation, it acts as a "double-edged sword" by actively resisting beneficial safety steering injected by researchers [cite: 54]. If models naturally develop computational resistance to activation steering as they scale, the long-term viability of intervention-based alignment remains a critical and unresolved question for the mechanistic interpretability community.

## Future Trajectories in Interpretability Research

The formalization of the superposition hypothesis has fundamentally redefined the scientific paradigm surrounding artificial neural networks. By demonstrating that models utilize the geometry of high-dimensional spaces to encode exponentially more features than they possess physical neurons, researchers have established a coherent, mathematical explanation for the deeply polysemantic nature of neural computation [cite: 5, 17]. 

The field is currently experiencing a rapid theoretical expansion. The establishment of strict mathematical boundaries distinguishing the capacity for passive representation from the parameter requirements of active logical computation provides a rigorous framework for assessing network efficiency [cite: 16, 18, 39]. Simultaneously, the recognition that realistic, correlated data induces constructive interference and semantic clustering shifts the superposition hypothesis away from idealized, independent geometric models toward the practical realities of linguistic data statistics [cite: 29].

While Sparse Autoencoders have unlocked the unprecedented ability to disentangle superimposed features and map the cognitive architecture of models like GPT-4 and Claude 3, their practical application for steering remains constrained by architectural entanglement [cite: 22, 42, 51]. Moving forward, the discipline of mechanistic interpretability must reconcile the extraction of monosemantic features with the holistic capabilities of the network, ensuring that safety interventions can be targeted without degrading foundational reasoning capacity [cite: 51, 52]. As frontier AI continues to scale, mastering the complex mechanics of superposition will be essential for ensuring that neural networks remain trustworthy, steerable, and fundamentally comprehensible to their human operators [cite: 4, 55].

## Sources

1. [Open the Artificial Brain: Sparse Autoencoders for LLM Inspection](https://towardsdatascience.com/open-the-artificial-brain-sparse-autoencoders-for-llm-inspection-c845f2a3f786/)
2. [A primer on sparse autoencoders](https://nickjiang.substack.com/p/a-primer-on-sparse-autoencoders)
3. [Deep Dive into Anthropic's Sparse Autoencoders by Hand](https://medium.com/data-science/deep-dive-into-anthropics-sparse-autoencoders-by-hand-%EF%B8%8F-eebe0ef59709)
4. [Extracting Concepts from LLMs using Sparse Autoencoders](https://openreview.net/forum?id=F76bwRSLeK)
5. [Sparse Autoencoders Reveal Universal Feature Spaces Across Large Language Models](https://arxiv.org/abs/2410.06981v1/)
6. [A lay introduction to "On the Complexity of Neural Computation in Superposition"](https://www.alignmentforum.org/posts/kYkbcEo6YyjQiztpi/a-lay-introduction-to-on-the-complexity-of-neural)
7. [Toy Models of Superposition (Transformer Circuits)](https://transformer-circuits.pub/2022/toy_model/index.html)
8. [On the Complexity of Neural Computation in Superposition (arXiv)](https://arxiv.org/html/2409.15318v3)
9. [On the Complexity of Neural Computation in Superposition (MIT DSpace)](https://dspace.mit.edu/handle/1721.1/157073)
10. [Some costs of superposition](https://www.alignmentforum.org/posts/ArvgRhqnEeX7cnJu9/some-costs-of-superposition)
11. [Privileged vs. Non-Privileged Bases in Mechanistic Interpretability (arXiv)](https://arxiv.org/html/2404.14082v2)
12. [Privileged vs non-privileged bases in machine learning](https://harrisonpim.com/blog/privileged-vs-non-privileged-bases-in-machine-learning)
14. [Privileged Bases in Transformers (YouTube)](https://www.youtube.com/watch?v=-oKuDRFHW_Y)
17. [Toy Models of Superposition Summary Poster](https://www.mlmi.eng.cam.ac.uk/files/2022_-_2023_advanced_machine_learning_posters/toy_models_of_superposition_reduced.pdf)
18. [Superposition is not just neuron polysemanticity](https://www.alignmentforum.org/posts/8EyCQKuWo6swZpagS/superposition-is-not-just-neuron-polysemanticity)
20. [The urgency of interpretability (Anthropic)](https://www.darioamodei.com/post/the-urgency-of-interpretability)
23. [What are polysemantic neurons?](https://aisafety.info/questions/8KGQ/What-are-polysemantic-neurons)
26. [Feature Capacity and Polysemanticity in Neural Networks](https://arxiv.org/html/2210.01892v4)
28. [Mechanistic Interpretability: What's Superposition?](https://pub.towardsai.net/mechanistic-interpretability-whats-superposition-957c15d18e92)
29. [Interpretability Team (Anthropic)](https://www.anthropic.com/research/team/interpretability)
32. [SAFE: A Sparse Autoencoder-Based Framework for Hallucination Mitigation](https://www.themoonlight.io/en/review/safe-a-sparse-autoencoder-based-framework-for-robust-query-enrichment-and-hallucination-mitigation-in-llms)
33. [SAFE Framework for Robust Query Enrichment](https://arxiv.org/html/2503.03032v1)
34. [Phase Transitions in Hallucination Detection](https://arxiv.org/html/2604.16430v1)
35. [SAFE Framework Validation (EMNLP 2025)](https://aclanthology.org/2025.findings-emnlp.496.pdf)
36. [HALC: Object Hallucination Reduction via Contrastive Neuron Steering](https://openreview.net/forum?id=LTJZAkRQAz)
40. [Incidental Polysemanticity vs. Superposition (LessWrong)](https://www.lesswrong.com/posts/8EyCQKuWo6swZpagS/superposition-is-not-just-neuron-polysemanticity)
41. [Incidental Polysemanticity Origin Story (LessWrong)](https://www.lesswrong.com/posts/sEyWufriufTnBKnTG/incidental-polysemanticity)
42. [What Causes Polysemanticity? (OpenReview)](https://openreview.net/forum?id=OeHSkJ58TG)
43. [Polysemanticity GitHub Repository](https://github.com/kushalthaman/polysemanticity)
44. [The Sequence Opinion 667: The Superposition](https://thesequence.substack.com/p/the-sequence-opinion-667-the-superposition)
45. [Steering Language Model Refusal with SAEs (arXiv)](https://arxiv.org/abs/2411.11296)
46. [Steering Language Model Refusal Poster](https://actionable-interpretability.github.io/posters/55_steering_language_model_refusal_with_sparse_autoencoders%20-%20Kyle%20O%E2%80%99Brien.pdf)
47. [Extracting Concepts from GPT-4 (OpenAI)](https://openai.com/index/extracting-concepts-from-gpt-4/)
48. [IBM SAE Steering Repository](https://github.com/IBM/sae-steering)
49. [Why SAE Steering Fails on JSON Generation](https://huggingface.co/blog/MaziyarPanahi/sae-steering-json)
50. [What Makes it Difficult to Explain Neural Networks (Superposition)](https://towardsdatascience.com/superposition-what-makes-it-difficult-to-explain-neural-network-565087243be4/)
51. [Mixture of Experts and Network Sparsity](https://arxiv.org/html/2510.23671v1)
52. [Superposition and Interferences (arXiv)](https://arxiv.org/html/2507.15798v1)
53. [Mathematical Models of Computation in Superposition Review](https://www.themoonlight.io/en/review/mathematical-models-of-computation-in-superposition)
54. [Mathematical Models of Computation in Superposition (arXiv)](https://arxiv.org/html/2408.05451v1)
55. [Complexity of Neural Computation in Superposition (arXiv 2026 update)](https://arxiv.org/html/2409.15318v3)
56. [Mathematical Models of Computation in Superposition Abstract](https://arxiv.org/abs/2408.05451)
57. [Toward a Mathematical Framework for Computation in Superposition](https://www.alignmentforum.org/posts/2roZtSr5TGmLjXMnT/toward-a-mathematical-framework-for-computation-in)
58. [Review of Mathematical Models of Computation](https://www.themoonlight.io/en/review/mathematical-models-of-computation-in-superposition)
60. [Mathematical Models of Computation in Superposition Text](https://arxiv.org/html/2408.05451v1)
62. [Mathematical Models of Computation OpenReview](https://openreview.net/pdf?id=OcVJP8kClR)
67. [Privileged vs Non-Privileged Bases Comparison](https://openreview.net/pdf/ea3c9a4135caad87031d3e445a80d0452f83da5d.pdf)
68. [Mechanistic Interpretability Glossary](https://www.neelnanda.io/mechanistic-interpretability/glossary)
69. [Inside the Black Box: A Practical Guide to MI](https://medium.com/@adnanmasood/inside-the-black-box-a-practical-field-guide-to-mechanistic-interpretability-b8757600e2de)
71. [Harrison Pim Blog on Privileged Bases](https://harrisonpim.com/blog/privileged-vs-non-privileged-bases-in-machine-learning)
72. [Bag-of-Words Superposition (BOWS)](https://arxiv.org/html/2603.09972v1)
81. [Feature Geometry and Correlations](https://openreview.net/forum?id=7akSRQS5Xh)
82. [Toy Models of Superposition Poster Detail](https://www.mlmi.eng.cam.ac.uk/files/2022_-_2023_advanced_machine_learning_posters/toy_models_of_superposition_reduced.pdf)
83. [Bag-of-Words Superposition Detail](https://arxiv.org/html/2603.09972v1)
84. [Superposition Hypothesis and Sparse Inference](https://medium.com/@liu.ma5kbuccaneer/superposition-hypothesis-and-sparse-inference-dcf7cbf880ad)
85. [From Data Statistics to Feature Geometry (ICLR 2026)](https://iclr.cc/virtual/2026/poster/10011287)
86. [The Sequence Opinion on Superposition](https://thesequence.substack.com/p/the-sequence-opinion-667-the-superposition)
88. [From Neurons to Neutrons (Meta AI)](https://ai.meta.com/research/publications/from-neurons-to-neutrons-a-case-study-in-mechanistic-interpretability/)
90. [Mechanistic Interpretability Taxonomy](https://arxiv.org/html/2511.19265v1)
93. [SAFE Architecture Empirical Results](https://arxiv.org/html/2503.03032v1)
94. [Steering LVLMs via SAE Latent Directions (SSL)](https://arxiv.org/html/2505.16146v2)
96. [Endogenous Steering Resistance (ESR)](https://www.ae.studio/research/esr)
97. [Feature Selection Strategy in Steering](https://openreview.net/pdf?id=Yz1gZJFRvT)
98. [Hanni et al. (2024) Quantitative Results](https://arxiv.org/abs/2408.05451)
99. [Anthropic Claude 3 SAE Safety Benchmarks](https://www.anthropic.com/news/mapping-mind-language-model)
100. [Toy Models Bounds and Constraints](https://transformer-circuits.pub/2022/toy_model/index.html)

**Sources:**
1. [substack.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEh71DBuAcqcgA-2kWgMHox95z_6y2sKJB8tfq8a5_p4LrFzi3LitHXMtVZGDSpgG7rrAwcCySvakO4vW1v-ly14DIQldD2TztKk8jvaUrS9qvUtA2IcfMjMMc0JsI4mj0O9PIwttEPjTWQ-UJY7OWael6-U4A=)
2. [meta.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGhndcFSCttnJqww_eAG5Jrl_gxzk5rAPGnX72uZp7Z9sWu_wEfPnZvnCy4VFnvIhPsRElbEtkD5tm7OxXNz9Qvs16SH4V0y725XbxXufzAe2ISVa55xJ1Gp1g_zov790rtsGQGe2YPJeAtbOp0Uis6O05p_fyAQllrTrpB-fs0NI2CsVapubf4J7mfeesmgY7mycgwSjo9XeuOIg2WBBZaDE3lKZE=)
3. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF1G5PqX5gadpLVW83HWJzfcCUovHfX-sJsP74ImQFeNAHzCyks88msnnZk5BD7Wm9T5FoxOZ5vR8YLKg4kK2ckVGZKvkQgVKFFaqUwvZT3GQp7OkiLdWMy)
4. [darioamodei.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGprWgINFsPh44utAZkN9Vf66ZE_8l58Nrl9ow4MAzZHD5sWGbOL6i5NLU1FWn2mZ6g2KzBekguur4hUGwTeRaLy4ZmRq05f5aikOKifO9th5uGiJuFmAf54_ul18f4vOtn7QFXthDD04ruid5D-TrqwACoa_k=)
5. [substack.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFkJnYSPOQ82IDp68XWi1pBRfLOtmNcZ1q8LtPdd9xW-81MF5PpzME5JPk1rbF_jUf9qWgp3pimSG9bOHsjbtoh188plQwDxZKTT7CRmaHyrDs4q8b1s5FJEv23ElAzsrAqUdln2HUMzJQqCjBbjGe1aYLQcEUejQAd-T6xLS1gqMEB)
6. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF8VsQfq015wxXugYdUN4SV5b-pItFw-Cr7zg--gJdrALAwa8YRTfwOu2wKAn9Q8YllmO_EskTUhJ_2MVswc5vnOIvJpjSxATKuzuSCJnOEEQvyBhbX8SSAliXlnlnYpT6Mm2HIslNUIsbfJi5RcGRK2qElWCziGTBMmclWfGR4x_BGOoEHKsYqr_2KQCFzfr83GT4tJB_eO7ZO5qnKbzshwg==)
7. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFDpoMxn1U62I7EJrvjm4ZAae-dTj_DbIoDMtccMgxY6Xp_OYYd_Y0wVMdZARcCRud0Z7i9cSlr1mc47EVuXXTsrrdSWnInLXLi36PWKuWzN0ym1mHo0o_9)
8. [alignmentforum.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFtBfB-Otp4JvJXV7cAvHzHn1Cou9q38faRhJ8zmV473dz5VgTjVCdrMDX_2SGKBr1i4fN5PIM-euZmQRQsU3Kl2DdNO1xxYMgz4gUprkZ7Iu013hsDiFp87KFwq-P8YuG9vaLmUyqYpckDrarzLX2pSB1BNuCyHe4l0rcgYW_MDwdSpeL8PI9dMNIqMAADKNb6wKJ_vVjGogn7WZI=)
9. [github.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHIYyANEBKoCJ_CKSGV96diFeuwid5cYmPg5EoePNVAup_KjaiSAVBGIo9T-1Pa7cJc-1kGnCyBuLS1J0X8WuXHb7DIh-FErh1ZH1RcJV1A0N8z73LTeg4Pl7UbbPw-mm9bM2_m)
10. [aisafety.info](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFVlUFnzqerEQ0ItRGgESA46uELlv7siiQxdUdzSfLrIwye5xefE_o2EL75w5_hIfLOg277taz84U4EGMZilMSeziSd_aWaQEKFpKl95_qQ0XZr1GWvHwhhOXEuXf11bf3Zlu_N1bKNnMsrW5R6wnOGHXJtLQBxkA==)
11. [transformer-circuits.pub](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFi99suwZpwXrM3qPiWAIyDjwc60dMc1YvT7TugZXbyxxM-wJguhuC8CJb9EjalSGppswyvELBf4XLdn10TSlUiUrJXVuRBo77H1lBD0VQp3CpJDUwXbESffzS88lK1XYxD8EvxysoFZKm8j2IRvW4=)
12. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEU_8Z2S5dfE4HAihen-m5DrJi6T-ccWlj01gZ1NPlgf0UORoePpQLcuQ2-uUJhtzn6Yi_pv6sN-9PgkqVe59i_r8c8gZjDfqALgsqHHfqX3TVu5s2qUmPB)
13. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHKnmh4yj_p8zRHYBa3YkDlSZe56yM9lTWb5tlthdF7nTOWW7j6aOIxVEo5TDv1rXLqW96_HnJ-DFXvsG6dU4CTiZzItgyHnOudvfWBSbEZ8-z0d7KClwVq6YhCIK0y12b-fNxP5ugN0R9pwXr8ZSOKyAO_KnHYT4qVQfEPl0hynCfD-3UTdXhVZnVXiMq47Aa_kRgBOQ==)
14. [towardsdatascience.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHHYuPFPGN6rGH0LqJ-aPZMMmNsn_kfG00L4y-9eUqEpHvvXeQ3qW1JtRXV03WiAdRT_t56yUfMzwtkhpfInCV5HJ4vW5L0WKIybwdC4MsYKC02Ej-ITRuspOi4o2vgpqG0rvWKdxDRljXKkAr6nOp0REMkpCNGgJySBlBrIsRzYD-JWOEF9_WJEiKOQMjPkvqpYFCtOUJqLtAfNAM8vtWkYk0=)
15. [alignmentforum.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFt0FmLwwC3L-7mJkjXqCxY8CInzP85rUSTTrr7_Qwl7t_z3RQfkevZcvar9ot0yfGdynZPKJ7imtmco_B9kGfwHu87dbkrXK36DBCqOuGMV5q02WeE-7UQmKyWRFZ-SDrf1lPS5rNul8s5ejlpvnjkh37Tr2QhTu4gHc3X-W_kzRRaSDBTgMMeZnDHcvu0a2UBrKq0q0fJh1JRoGvD)
16. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGEOXqObDxYR_whtXMWE96cCOMi44RtTe9OX7np2OSXOQmt4Ecozfra7OYQxsq4jlyY_JZhkNDaOqqlxMwuGBnYLLLu7odYnFUJpixLwt1Z6VL9bfdRTxcK)
17. [alignmentforum.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG8g6YiRdm_cVrpTpZP-jUX6Yq8Wlq825gc_rooc0kLuHfr7fno_reNpdlo1tp8RBcmShR3wV1NorLGKDM35WaIPebY78uYHhbHzoZgKVH142hpILqsVK0o5LNGbVoK_WOVAU4N07T-nDDI4bc5Sk5kvEDN3jhhAevrM12Gd1Knr2kv8AE2n6Q=)
18. [mit.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHk7seyKa1DPWUyxFArDL42GdyIpYa8DHOgEXNzzuYTOnzufCXn_n7ev_8w218E_1l5ytrbAyvnkZ8VWKY-wcu5WZ2C633N64-eh64DBZ6oQVrIHZVv5mBmH-o3bbW1y0Q=)
19. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEzlqvDZ1WgnGyI0AAPFyT5KWFfFrv_Q9z76WTEZ7CSPb_l18or9kbum-nDWxGF3T41g5V5BKfQkmkL6bada_Xg7j1l5lADy_L-QqIgIN7PnlWBtDPYnWj8)
20. [towardsai.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGAr8LfwI6j0qwUDXLygolsT3sEq6CmzzD0enWyK-b8ykT8GfmFnD0szlADGDPNf-h6PHLvnJDkPs65NMa7Wv4DJFC9MEPwOBwgURHaut1Fif11GjslYlqrjnKSNkZohXCBf1HOi0TcRfmFUqeYbv2vHygMGeY7zhc0AT0IrCzDCNFUOsvnZv78YtWGCA==)
21. [towardsdatascience.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHOXrCYUgTRJ6CdKJ4dwJRZII6nV7z24pa8UbRkf_I8_VCMY9sX2VNsORLvD-b6K8av7S0xRGnyEOZ3t0zwR_zmFUSDUyPy544AlvqfZhmHyYJY1lLKxPdBVbUZJmOjJH_tLe292Sg6GA4ykrtAbSkGUPanneb1nr9j0aimkC9VaMLf28cAv4Imntr68p5RFqOj3meRSQHTHjc8F28lj66ERg==)
22. [openai.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEnRYTL_s-T2eWyM8VCDmTTRYjBGDtFSckPbPwsfEj2zJwANO-mwnLqeLuNEOREAdckSVfr_0-e0p0G_K9BWiEWMLHc2-K6pvHEYVcPtZbKOWkY4RAqsoFY2Tuo3RIoS7AouJQQoHR72-l9tTa-)
23. [neelnanda.io](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEdy3xHMGqblWkw3-8mC6wJFQ4vwLwSCqV8E2yF2qpNrKA1T5PBJo9Oa5rwkjbLVrGWq0LZkZ4N4HJN-6wgKm0vFRs2-MsJz1vwQE2beEjVURQ6BbGmkvgK-60b8HxEhNJw64a7nnDJjxMysWLMTcD6DLfS)
24. [youtube.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFysyyiYn4ODn4ULY1S3JzmpqTsxU42tO-6sN0RTlk9oA6tUBaf2K5PS7t5NVDGmtk4e0ls-4YSHzhNmqvMGX1Cg51l92uiewQ2HcIrwaeLXdr8nFRUPndDL653XZjz8jo=)
25. [harrisonpim.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFF_X3h3NPxE5hDSG1pTLRw9ABnstrGQosicQudouUlW0AD6iy6Jo_X10QDDUgvQBtS1Yv7rUW6TXLmsQWii8f5Qm7DVtPbDQ7LSPdTJG_pe6XMK9XWU4O94EAJhA3x9cesBmDs14jwH49219XnAjTbMCrcHMDm1SzJtE9a8EuEc68yj1InjziP)
26. [openreview.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGsjREB5ejGKxH7RoCeB6YcrABYfIPnZZ5-nDs3HNKwsngfHkoztHyRP-6p7OdzsuvQ3NfzXbKnvqHnR1IweGM5CMLV0F_WP0PpLb0huQf7Wv4_h6X5pq2cQfXNCeiPFcYpUd-WppSh6-Ld6G93GmWh3oARA3W2zCWlfTeV)
27. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF_b7K_97zCF9BL9X1rpeXkClB4zUC7g4Fghhef_SbrMXsiLkpu6u2YTdyXf4y-9qfw3XxAGeO1nlztM6EHBwxsFLXl_nX-O9JMyDof83vsXbu42EPTnuz88BWNcctNdDqgCjiyRaakWSij8zgq5Osurjbd0GNe-MMHNEWzgaCqdCztEOKF1eqN2qvtTDRT4RwChMuD7v7swM-iBLAj70LkGMocPLXPD3d9sBWLd3Y=)
28. [cam.ac.uk](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGwBzXetCyLRJnmtyQYOI5Api0pH5cAwwfqg1-V7AQ35VkKErPAwNxGMcpFk139X0eG07HL-dUFVCcpjxWVH6wN5wIYNdvLacx_RYhri0JpBIBc88vNk4M6XD0jbQEHrdrDECz54dlMmdo60JluswlMYlPnEH8D9NYqz5iGFGUae45V8hZzN27YBDL3_LFCEY4-QWql3ON8ijsgmmkkm31_QQJ9D6GSJvfOCkHKcfDP)
29. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG89auB__-KML7vkLilRZTKADD1V0231GVMeR0QXz10_nXm2Ms1A9PaasDythjYOGDVSsbzUpNVL3FsBO7RuVyxjMf1KRBoKBtn44rYQ8K-r032Aa41E-It)
30. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF7KHwk77r9JcnDHgLRuy4wBa4jJ8vMjVmMVJPK0zTo9_YZTFrTK43mXYta5a9fPzfWuu9x6y6bB2g_YtUpCV5wCgQlKvLk_QCCaExsZ62hWu-1AJxLSFws)
31. [openreview.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEd_vMWYuGw_9ku1cSL_saj51bNfGwd86bImCRtT2jphA7e7NbrF1A9nDiWGwzWXgAIjgd0kLkhYa2E2q-dZL4H2JPApwn3bFJFHfG4EcG4_PN0Js2Cc8aVH1Ije7ZPeQ==)
32. [iclr.cc](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGD2stdSN9oErR5FUnF53iaIYkj9rGHlzYi0zZf6bF244u0ymDvZA0D-9v5r71R-itOVckb3fG63ChWrf9ExPDmuJKG_BA3X3p-hwepXr_bDsKmSJHREZVwDoVtNX_tz9FO)
33. [lesswrong.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE4QFlV-ilf4wLnPYqoMsRFNu5F4ErALnMvagxlU9LlhNYO999-270HUF2c7pXOyxC4jktgZMr_b-bzV8tvOVS9JNK6ydZhEkSVkUVYAcDCnuFLMxjhhcLtU8uNYDOQhVI4wsdcw94KqJDW9wYxK_0Pbzr0MUA8fYXBQ0KytwyUBCv0_9BXFJ1LQ40gfZNQhKll1VKDJngv)
34. [lesswrong.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFUUTXVCMF_oo3LT5I7-9O_k-NhODsuGa6zn0d0muq1tpBvh8eMN_6HyFd6PejMVthKhgMWj49aBbVhoA1HyouZSgh1W8_YDoC0Pn0p0-7lqJ_u0TTyoOyhNM2Xv2MD1jb8YeRFsc4V6IF00ixbHbyzZ_bW4jhQRgMALlKiatrSKQk=)
35. [openreview.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFav9QxLkMUwrrm7pawaByA-M57Em4Nds0aIPFfBTr5XNInhvIO7KRsUTSssBSp9N5vvH0qbetoYWV-z3G05pz_dtmRv50pRLORmTjEz649EAP2esX0dgxRk0diQLwQlg==)
36. [themoonlight.io](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFgw_UTnOtuNh3AwbaMrLl7nMTjE0mxhmau-RmwcHU-DnFfbxTyPFJAzmJ1t_U7S9acBiMuHUKQsMajaTnQ_w_MVhCTfy0MhVrnFlKGNDMjmRkylYtfcZccS4HY5YwZCHn2DrSCUc22ipENHgsijarP1yE8dNVadVjMRv09bnjBzZtrrLoyu9ZUCXmSJfgB)
37. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEKaTdLc6f9SndH6K6s6n9uKvQ92mfPhfSRirityCEv04qk2RYC-Z5ErWsrf_l69zE1CaFVNJFplfzMqkM-7ctx7NK56ubvWdLIdQ8ye1tfRCalMYK6)
38. [openreview.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFp9N8Wdx3CzoezgZhFUt_wZQDAyzXc4WKqIFB2WCpmNT1bwytp3Th2YtsXkKLBJyIg1ff1AJmNK_yYU89fsvIgdBP-GQG5EOuzYz2mSfv-DSSvwVnSbxTCvvc-lnA=)
39. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHhu1jkRgBOUJfBOAYdS9CgcqXEUlVhxRoBbSgdB5GMjP6n-Z8nz3oTSMm4F4gQigBlLw-gBFHYKZTNFLGGbV2cW61gs_NPxf7YfsVM8Of8U4OmG5DORUpq)
40. [alignmentforum.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHH_z6Ap3IxWb7o99ZcFvw_HlWREWRwVYmsyJlu7h-NFi2PF2nvTvdI_ezbjlacJe2daZVMijtk4-JkAfE7OOPytuRIjgWOdHiNLOpkif4KYqFkUpgZTA0QnSSmiGV-pC9X6D2VJ1s7LHp3P4rkkIx6fWPN15cpr6tULWF_3evxgCBRByGI8E4vQ-qsYmIEvRE7bNjeD5g8eCnZ3VxZWg==)
41. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFvZQzKXx8i5y487lPAlnsowiTx2U2byWvRNbeB8fIRmJ5_WPX0pcdYNrQNZybQ45kAv9uSOE99exCnKopiJEI0R_ea8FsttFS_BPl3bwCb3zTWqLUW6JpJ)
42. [Link](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHXnr4fdVpQgYXcxmCjLH53JJocvqdZ952AjTFQBLUWYRsoTHl4oA58ovj-xbooatShmfOeSzbOs-teLF1ZJ0fWNC7SWmueYaAUpOgVVsg4S6VssQTAidc0ImKUC7rA62cLs5jpwPD4MlwFtQvpNGI=)
43. [huggingface.co](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF3khe9rrIKnt9p860N-spaBu7UYg-59GmIcueGSGiSAAOFnV97Undz24-iVpnF2JRaXj0-Hy8vIfTfPfvWGdXz-2cITeQoJ4W-ZgOe4HbufBugo2DW6A5g2z1jwMSFFxvpwwU83uQQsvcVoB0-p9W1)
44. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQELhfRiyxWS-cNelqYgsrJjhpepp4OAkrhi7sL5YlJd01ALyyQMOdCJkcbtZzkfJ-Yymzd2k7TXxiij7Oj3B40RC6fFLk00TKUKWFaIH8J6Q4Z5hWKn73mS)
45. [github.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEkI9ulPItdvPjaZnaewgGnhBKsnwaW7ifcwh4hRyEi4lEdX5mHj-_Gp8Ur0tY_G7ZfzLqabvnWSQBqQbi5W_H4WcCuKXXbGF_xTlTAnkDK45IdsECXOtdu)
46. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFDTVfIgHPbyrLa4QYgRVvXgFxq2pdKY_GMlK1yb_SO6hkSuJh_fpbOKg3luL_8QhMDxNja-ia0cUBApnrLUgvHxHamEzuALVhcKS9EUcWJY22AZPkut4N-)
47. [openreview.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGaiq9eZ_nlQBUt6cJqjsuJ5anvnNX0arrk8yn-atM7xODdkzNWn2oRBEUMYD4mJ216oxQMSR-GRq8A8uVT_OfwN8JLhTXmWXKEYlRzoPfldbtjsReVvXBV8_33-cTJYw==)
48. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGdcge4N16BmPrHtjkpMrBpUYPWW0DNdVfBHOVnC6k5zsYsrpxXpSfz5YmIwfLajPSsVpN9GKcne6YrhKPJ5Mch4j_A3EutuSGA14cmyuAlekHnXnW3znJx)
49. [themoonlight.io](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGqSqtsMAkVY776izox9Xl3G27Wde7WV2B31mSMBHGWnVgxpIdqfsVDMLuq9xfadTBnLd5ufxYnpTN5uLXQXSuKbm3pqNkEVOp0EVvaL47lbvFQ2ddsXC_S6uwM67WLpvjRiY5c1-F3Ils0DNx1mCXY0xxd6G4lRMjaBpnB547QHyGrU6vSlXgyTN9wCmFtpXOD9EWIm7qMHIB_covoBWCTyaq3tHUTc-N6ldjgl9h5yVwfCOTbV6Fr83kYSGvh4kDfHqmi-bc=)
50. [aclanthology.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEedzp--hPJCHHkfH4x3OFTAfXVvx7TiAkZXm-qKci78QScbK9AdUwy038ur8m_Wn_8TG3nN-fHjul9CFEp8wEkBfZxGyFlCmao8-JBIJ-lfUgnLmCNThM-dSItSV3zIDhR6hzOIXwmeds=)
51. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE9jSRPpOu0iHk4g4PqeeswNod8uaNTB1atC4gSaf7ElGO0jXIdM06qFL5EqbHhV0S5u7EwxoxW-GqQYNfcwCj6aHEJIfy528BGt_JDC5k3tGQ9vuW4)
52. [github.io](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGV9B99GlFZumFpOPSPmt_7e4Z0lmDqGhIDSHNRhaAC4lGxwU1jda90-B9tCh9YH2WbvSTJKb7EeVLPx6Nr72NzblLtWvZAYZ4agVa81XnXLiJHjve14XAUHICdJsDMcuqoaD3Vxcl00rc2UZS2o02YSd4JQHeuENUJMv0Pq-S5i7ffQUqwk8eCD4F3uheLEZo2BqBlKaL4tyelZ6DeWXP5mV5rSeGdTL_rZp8gvSJWMMAWGlChyTsgHNjU88nakM3idKqLrn7rkQ==)
53. [openreview.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFPyTbz3Gw6B9R5jaEWrcEaThS2cYHnPRpwsc3wH0xiWhnANLasZUXD59E0G21ET9ZQJXsBEoeSrUs382hyqY7RykMr9bpY3luyA4yjVYEBGorTvGkjerxGVM-3Quo=)
54. [ae.studio](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFPJcasZRG88HO1xGEQB9N3VSwBpFkTXDQIS9zIqQqmVrITKKb76X1XAAiZoaE-bdn2klVwKV6liAepqIRmW9PkXB6kltZ26xu7WNHXtxxlqlhpeidlQtU=)
55. [anthropic.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGCztRZHu70lEPolBoFaN0mJrl_zYMn-Owke7PT7yeKWtnqTcEGsYNSoAY_JEObNuVCDwsPj8i9qy8u4ZLcY-kSlEUjoVlpnSg-rA58m5osvVP9KFHGMnfmy3Wxuen06VCqPx6cuQQYwSNOR5f6)