Anthropic's J Space: Measurable Capacity and Interventions

Anthropic describes a "J space" to isolate, measure, and manipulate "verbalizable" content within language model representations. Beyond the cognitive analogy proposed by its authors, the device relies on precise geometric construction and computational choices. Here’s what the definitions, key figures, and conditions under which the interventions remain valid reveal.

Measured Capacity and Fragility Zones of Interventions

Anthropic reports that the excess variance explained by the projection onto J space stabilizes beyond k = 25. k ≈ 25 is interpreted as a natural elbow, beyond which projecting onto J space no longer provides excess information compared to randomly drawn vectors. Numerically, this additional variance represents about 10%. For comparison, in the case where the dimension d is 4096 and k equals 25, the expected proportion of explained variance for randomly chosen vectors is about 25 out of 4096, or less than 1% assuming independence, indicating a manifold significantly more restricted than the entirety of the space. However, calculating the exact non-negative k-sparse projection remains an NP-hard problem. It is noted that this exact projection is primarily necessary for the offline evaluation of the FVE, and that an approximation can count the variance twice if the retained J-lens vectors are correlated. In terms of interventions, a case of fragility arises when the source and target vectors are nearly collinear: the pseudo-inverse becomes ill-posed, the Gram matrix approaches singularity, the coefficients become unstable, and may even turn negative, expelling the result outside the cone of J space.

What J Space Covers in the Model

J space is defined as the set of k-sparse linear combinations of J-lens vectors with non-negative coefficients. For any subset of indices, a cone is associated, formed by the non-negative combinations of the corresponding J-lens vectors. Geometrically, J space is the union of these cones, rather than a linear subspace. The J-lens vectors constitute an overcomplete framework in the representation space, with one such vector per token in the vocabulary and per layer.

Origin of J-lens Vectors and Impact Measurement

The approach begins with the semantic representations associated with a token located at position t in layer ℓ of the residual flow of a Transformer. The average causal effect of a modification to this representation on the final layer L corresponds to an average calculated over different prompts, across all token positions as well as all future context positions. After conversion to logit scores via the unembedding matrix, this effect is presented, line by line, as a J-lens vector. The effect on the propensity to produce a token i is obtained through the dot product with its J-lens vector. Under the assumption of an approximately uniform distribution of J-lens vectors on the sphere, the dot products between distinct vectors are on average close to zero, except for semantically close pairs; for a dimension of 4096, the characteristic order of magnitude of such a dot product is about 0.015.

Verbalizable Part, Residual, and Strict Orthogonality

The verbalizable component of a representation is its orthogonal projection onto J space, viewed as an approximate k-sparse expansion on the J-lens vectors. Sparsity retains approximately orthogonal and semantically distant vectors, such that tokens that receive high scores belong to clusters close to the k chosen vectors. The non-verbalizable component is the residual of this projection. Results from convex optimization guarantee strict orthogonality between the projection and the residual, which nullifies their dot product.

Effective Dimension: Defining and Interpreting the FVE

The capacity of J space is quantified via the fraction of explained variance, defined as the expectation, over prompts and positions, of the ratio between the energy of the projection and the total energy. By orthogonality, it is also expressed as the expectation of the complement of the energy ratio of the residual. To isolate the signal beyond the effect of k arbitrary vectors, the excess variance subtracts the FVE obtained with k random vectors. Empirically, this measure stabilizes beyond k = 25 and reaches about 10% excess, while 25 random vectors would explain on average a fraction of variance on the order of 25 out of 4096, less than 1%, under the assumption of independence. These results are interpreted as indicating that a significant portion of verbalizable content occupies a very low-dimensional manifold, with 25 being much less than 4096.

Intervening Without NP-Hardness: 2D Regression and Conditions

For computable causal manipulations, Anthropic substitutes the exact conical projection with a linear regression scheme in a plane spanned by two J-lens vectors, source and target. A two-column matrix is formed, a representation is projected via the Moore-Penrose pseudo-inverse, and then the two components of the coefficient vector are swapped to substitute the concepts. This reduction to 2D provides a closed-form expression and avoids NP-hard optimization. The analysis links these regression coefficients to the exact coefficients of the sparse decomposition: when both vectors are active, their exact coefficients are positive and small perturbations remain within the cone; the condition for nullifying certain terms arises from the minimization property of the projection, and another nullification is approached via the quasi-orthogonality guaranteed by a concentration lemma from Lévy. It is concluded that the regression swap is a good approximation of an exact permutation of the coefficients. However, this approach becomes ill-posed if the two vectors are nearly collinear: the pseudo-inverse deteriorates, coefficients may turn negative, and the result may exit the cone of J space.