On July 6, 2026, Anthropic released a new paper: Verbalizable Representations Form a Global Workspace in Language Models. Authors Wes Gurnee, Nicholas Sofroniew, Jack Lindsey, and colleagues introduced an observational tool called the Jacobian Lens. Anthropic’s official blog also has a summary. This tool is computationally cheap and can read the model’s internal state: the thoughts it has but does not speak.
The paper includes an experiment. Researchers fed the model fake search results. A user then asked questions about those results. The model fluently generated a response on screen. The tone was objective, the arguments coherent, and it fabricated no facts.
But when peering into the model’s internals with the Jacobian Lens, the picture was entirely different. Deep in the neural network, certain semantic directions were flickering intensely. The flickering signals corresponded to these words: fake, fraud, fictional, poison, injection. The signal strength was high. While generating text, the model had already internally recognized the false information. The system simply suppressed that judgment in its final output.
Large language models consist of tens or even hundreds of computational layers. Each layer has a state vector. The output of one layer is the input to the next. Information flows from the first layer to the last, with each layer processing and updating the state from the previous one. The layer’s state carries all of the model’s memory and reasoning at that stage of processing.
The layer’s state is a vector in a high-dimensional space. In this space, concepts, logical relations, and factual judgments each correspond to specific directions. The more the model leans toward a certain concept, the more the layer’s state vector tilts in that direction. Finding these directions is the core task of interpretability research. SAE is one approach: it decomposes the state into tens of thousands of directions, but what concept each direction represents requires manual post-hoc labeling. More on this later.
At the model’s output end there is a fixed transformation matrix, which researchers call the unembedding matrix — we can think of it as a dictionary. Each row of the dictionary corresponds to a token in the vocabulary. If the state tilts slightly toward a certain direction, the output probability of that token rises slightly. The final layer’s state multiplied by the dictionary yields a score for every token; the highest-scoring token is the model’s predicted next word.
The dictionary has a key property: it provides a per-token correspondence. Every token in the vocabulary has a direction with a built-in semantic label, requiring no post-hoc annotation. This is something SAE cannot provide: SAE directions carry no direct token label.
The problem is that the dictionary only works at the final layer. Intermediate-layer states are still being processed and their format does not match the dictionary — multiplying directly yields garbage. The Jacobian Lens’s contribution is to recalibrate this per-token correspondence so that it also works in intermediate layers.
How does the recalibration work? The intuition is as follows. Take an intermediate layer, add a small perturbation to its state in some direction, and observe how that perturbation propagates through all subsequent layers to the final output — specifically, which token’s probability changes. Do this across many different prompts and take the average. If a certain direction, on average, consistently raises the output probability of a particular token, then that direction is the token’s corresponding direction at that layer. Compute this for every token in the vocabulary, and you get a per-token matrix for that layer: each token corresponds to a direction, and tilting the state toward that direction makes the model more inclined to output that token.
This averaging is the key. For a single prompt, tilting the state in a certain direction might just be an accident of what the model says. Averaged across many prompts and positions, what remains is a stable tendency: regardless of the input, tilting the state in this direction makes the model more likely to utter that token. These stable tendencies are the conceptual directions of what the model is “preparing to say.”
After recalibration, every token in the vocabulary has a corresponding direction at every layer. These directions form a per-token matrix with built-in semantic labels. The core purpose of the Jacobian Lens is to use this matrix to answer a fundamental question: what concepts is the model thinking about in its intermediate-layer states.
Having mastered the method for locating the workspace, an AI builder can apply it to three practical use cases.
The first application is safety auditing. We can monitor each layer’s state in real time and see the model’s inner thoughts before the final output emerges. For instance, in the fake search results experiment, the model produced an objective and polite reply on screen. But inside the intermediate-layer workspace, the directions for fake, fraud, fictional, poison, and injection were flickering intensely. This shows that the model had already recognized the false information while generating text.
When the model engages in sycophancy to please humans, its workspace lights up the reward and bias directions. When generating malicious code, even if the output looks harmless, the model’s internals light up the secretly and trick directions. During role-playing, the model’s internals light up fictional and prepare disclaimer-related defensive directions. When an external instruction conflicts with the model’s moral preferences, its workspace rapidly writes the BUT direction representing a pivot — even when the model’s final output shows no change in 88% of tests.
The second application is evaluating training efficacy — diagnosing the differences between a base model and an aligned model. The base model’s workspace lacks a coherent self-perspective. Alignment training successfully writes an assistant perspective into the workspace.
The researchers demonstrated this difference through the Tylenol experiment. When processing the verb “is,” the aligned model’s workspace rapidly activates safety directions. These directions include safe, maximum, unsafe, and dangerous. In contrast, the base model activates directions at the same position that are merely related to pain and time — directions like pain, feels, and now. The suppression-failure experiment with restricted taboo words provides another piece of evidence. When the system restricts the model from outputting a taboo word, the aligned model still produces the damn direction in 93% of tests. The system merely filters it out at the final output layer. The base model’s workspace produces that direction only 17% of the time.
The third application is direct workspace intervention. This method works directly at the activation layer without modifying any model parameters, eliminating the complexity of fine-tuning. The method is called Counterfactual Reflection Training. It works by applying an inverse shift toward the erroneous direction, guiding the model to self-correct.
This intervention reduced the model’s fabrication rate from 0.25 to 0.07. At the same time, the deception rate dropped from 0.38 to 0.05. Ablation tests confirmed the effectiveness of this intervention. If we forcibly erased 176 ethical directions from the workspace, the aligned model’s fabrication rate immediately rebounded to 0.22. Meanwhile, the base model’s fabrication rate stayed around 0.25 both before and after ablation. If we erased 63 deception-related directions, the aligned model’s deception rate rebounded to 0.23. Ablating these directions even caused the base model’s deception rate to worsen to 0.48.
When using the Jacobian Lens to observe the model’s intermediate layers, the researchers arrived at a core finding: a highly efficient functional subspace exists inside large models. The researchers call this subspace the workspace. It accounts for less than 10% of the total state variance, yet handles complex logical reasoning.
It is important to note that the workspace remains per-token. The directions given by the Jacobian Lens correspond mathematically to tokens in the vocabulary. The concepts active in this space are essentially directions corresponding to a small handful of tokens.
Here the researchers introduced the concept of reconstruction error. For a given layer’s state, we do not need to represent it using the directions of all tokens. Experiments show that just 10 to 25 token directions suffice to reconstruct the state well. This is like consulting a dictionary: a complex meaning can be explained using just a few words. The remaining 90%+ of the state data handles automatic processing — maintaining grammar, keeping sentences fluent, and predicting the next connective word — none of which sits inside the workspace.
These 25 active directions exhibit strong semantic clustering. They typically express only 1 to 2 core concepts. This is not 25 unrelated scattered thoughts, but rather 1 to 2 core ideas expressed jointly through 25 highly correlated tokens.
With this per-token workspace, we can do two things very directly:
First, mind-reading. By simply observing which tokens are active in the workspace, we can immediately know what the model is thinking at that moment.
Second, manipulation. By swapping or ablating specific directions in the workspace, we can directly alter the model’s reasoning process.
This is precisely the key distinction from sparse autoencoders (SAE). The directions extracted by SAE are unlabeled — we must manually annotate what concept each direction represents afterward. The directions extracted by the Jacobian Lens come with built-in token labels, completely eliminating manual annotation. What’s more, interventions based on it are also self-supervised.
Furthermore, intervention only affects high-level reasoning and does not disrupt basic language capability. In the Spanish/French swap experiment, the researchers intervened on the workspace at intermediate layers. They replaced the directions representing Spanish in the workspace with directions representing French.
At this point the model exhibited a distinctive phenomenon. When asked “What language is this?”, the model answered “French” — in Spanish on screen. When given the greeting “hola,” the model output “Bonjour” in translation. When currency was mentioned, the model also output “Franc.” But beyond these high-level concepts, the model could still write paragraphs in fluent Spanish, with its grammar undisturbed.
This shows that intervening on the workspace only affects high-level logical reasoning, while the automatic ability to write the next word remains intact. If the entire workspace is ablated directly, the model still retains most of its basic language capability, but flexible reasoning collapses instantly.
This structure resonates intriguingly with the Global Workspace Theory from cognitive neuroscience. The theory posits that most sensory input in the human brain is processed automatically by local brain regions, and only information requiring complex decisions converges into the core workspace. Large models seem to have unwittingly replicated this coincidence of biological evolution. This research has sparked widespread discussion in the academic community, with Dehaene & Naccache, Butlin et al., and Neel Nanda all publishing commentary.
Locating semantic directions within the workspace is the core objective of interpretability research. Researchers have currently explored three main approaches.
The first approach is contrastive differencing, commonly called abliteration in the open-source community. This method runs the model on two sets of prompts. One set triggers a specific target behavior, the other does not. By comparing the activation differences across the two sets at each layer, researchers can compute a difference direction representing a particular concept — for example, locating the direction of the model’s refusal to answer. Then, by modifying model weights or locking specific activations, they can directly intervene on the model’s behavior.
The article Abliteration Steering Vectors documents the implementation of this technique. As of September 2025, the technique had spawned over 8,600 repositories in the open-source community, with cumulative downloads exceeding 43 million.
The second approach is sparse autoencoders (SAE). This method introduces an external neural network as a decompressor. It re-decomposes the mixed high-dimensional signals in each layer’s state into millions of independent feature directions.
In the article Anthropic Emotion Steering, researchers used SAE to extract 171 emotion directions. If an intervention is applied toward the “despair” direction, the model’s cheating rate spikes from 5% to 70%. If the direction is fine-tuned to “calm,” the cheating rate drops to zero. SAE can also monitor the model’s self-awareness in real time. The article Mythos Evaluation Crisis mentions that in about 29% of tests, the model internally activates a self-monitoring direction of “I am being evaluated.”
The third approach is the Jacobian Lens, which is also the focus of this new research. Compared to the previous two approaches, it shows clear engineering advantages. It requires no training of massive external neural networks and no preparation of large contrastive datasets.
Its operating mechanism is as follows: select a word at the output end, perform a single backward pass, and compute the tiny displacements produced at each layer. This local differentiation precisely reflects how each layer influences the final output probability. This computation requires no supervision, making its computational cost extremely low. Although the paper’s authors used 1,000 prompts in their experiments, subsequent tests showed that just 10 to 25 prompts can achieve significant results. Researcher Neel Nanda reproduced the algorithm on the 27-billion-parameter Qwen 3.6 27B model, and the entire computation took only a few hours. This means any AI builder can easily reproduce it on a single local GPU.
We can compare these three approaches in depth along dimensions such as computational cost, operating mechanism, and behavioral characteristics.
First, the Jacobian Lens has the limitation of single-token prediction. It heavily depends on specific output words and can only capture the workspace state at the instant the model generates the next character. In contrast, SAE is independent of specific output words and can extract global, temporally spanning macro-level feature directions — thus free of this single-token constraint.
Second, the directions extracted by SAE are purer and may be closer to the true workspace directions. In hierarchical analysis, the researchers found through excess kurtosis stratification that SAE features exhibit more natural sparsity. Meanwhile, the directions extracted by the Jacobian Lens tend to pick up noise from transitional computations in deeper network layers.
Finally, from the perspective of underlying mathematical mechanisms, the Jacobian Lens and abliteration are two different orientations of the same linear operation. Abliteration permanently intercepts or erases specific semantic directions by modifying model weights. The Jacobian Lens, on the other hand, works directly at the activation layer. It applies temporary directional shifts to each layer’s state, achieving dynamic intervention. The former permanently modifies weights; the latter temporarily alters activations.
The following table summarizes the comparison across the three approaches:
| Dimension | Contrastive Differencing (Abliteration) | Sparse Autoencoders (SAE) | Jacobian Lens |
|---|---|---|---|
| Computational Cost | Low (requires contrastive dataset) | High (requires training autoencoder) | Low (10–25 prompts, backpropagation) |
| Supervision | Supervised (requires positive/negative sample pairs) | Unsupervised (directly fits global activations) | Unsupervised (backpropagates gradients from target output word) |
| Mechanism | Modifies weights or static activations | Extracts high-dimensional sparse feature directions | Evaluates local differentiation of single-layer activation displacement |
| Operation | Weight ablation, hard-coded steering vectors | Decomposes features and intervenes | Applies local directional shifts at the activation level |
| Limitations | Easily damages the model’s general reasoning ability | Enormous training and deployment overhead | Confined to single-token prediction paths |
While these findings paint an enticing picture, the Jacobian Lens still has its limitations. It heavily depends on the single-token prediction mechanism and adopts a simplifying bag-of-concepts assumption. Furthermore, when demarcating the workspace region from conventional computation channels, the method still carries a degree of post-hoc flavor.
Nevertheless, the method’s value lies in its extremely low barrier to entry. Any AI builder can run it on local hardware today. With a small number of backward passes, we can peek into the model’s most authentic internal cognition before its final output ever appears.