Skip to main content
Back

Anthropic's Jacobian Lens Finally Gets a Mechanistic Explanation

A new theoretical paper proves Anthropic's Jacobian lens isolates slow, context-stable features that grow across layers, reframing what its 'workspace' really measures.

Paper, Full Access
  • New theory shows Anthropic's Jacobian lens isolates leading Lyapunov modes, not a mystical workspace.
  • J-lens prefers features with three traits: amplify across layers, slow across tokens, stable across contexts.
  • Validated on a Bayes filter and toy transformer trained on Hidden Markov Model tokens.
  • Introduces omega-J-lenses, a frequency-tuned family that targets features at chosen temporal scales.
  • Explains why Claude's J-lens surfaced 'concentrate on X' features: X is coherent across all positions.
  • Reframes Anthropic's global workspace claims as expected statistics of slow amplifying modes.

A new theory paper explains why Anthropic’s Jacobian lens isolates a small set of directions in language-model activations. The method favors features that strengthen across layers, change slowly across token positions, and behave consistently across prompts. The authors identify these directions with leading Lyapunov modes, the perturbation patterns most strongly preserved or amplified by the network’s layer-to-layer dynamics.

Anthropic’s original study described this low-dimensional subspace as a global workspace, borrowing a cognitive-science term for information shared across multiple processes. The new theory paper gives the observations a mechanistic explanation. A J-lens hit establishes dynamical coherence; claims about working memory or consciousness require additional evidence.

A derivative map, averaged

The residual stream is the running vector representation that each transformer block reads and updates at every token position. A direction in that space is a linear pattern of activation, such as a weighted combination of many neurons.

At source layer v and target layer t, the Jacobian lens computes the mean input-output Jacobian of the intervening layers across a corpus. A Jacobian is a matrix of local derivatives: each entry measures how a small change at the source affects the target. The lens then applies the model’s unembedding matrix, which converts the transported representation into token scores.

  • Logit lens: Applies the final unembedding directly to an intermediate activation.
  • Tuned lens: Learns a mapping from intermediate activations to output logits.
  • Jacobian lens: Uses the model’s averaged local derivatives without fitting a separate decoder.

Anthropic reported that the resulting “J-space” accounts for roughly 10% or less of activation variance and appears primarily in the network’s middle layers. The theoretical problem was to explain why averaging Jacobians selected those directions from the much larger residual stream.