Artificial intelligence news.
Straight from the original announcements.

Anthropic's J++ Lens Sharpens Readouts of What AI Models Represent Internally

October 29, 2026By the AI News Bites Editorial TeamSource: Anthropic Interpretability (Transformer Circuits)

Anthropic's interpretability team has described the J++ lens, an improved version of its Jacobian lens for reading out the concepts a model represents at intermediate layers. According to the team, it surfaces intermediate latent variables used for the model's reasoning more reliably than the earlier J-lens or the R-lens.

Key points
  • Anthropic's interpretability team introduced the J++ lens, an improved version of the Jacobian lens (J-lens).
  • The J++ lens clusters activations at each layer, fits a separate Jacobian to each cluster and combines them with learned weights.
  • It also adds the R-lens improvement (layerwise relevance propagation) and post-hoc filtering of spurious readout tokens.
  • The result is still a single linear map per layer, so it can replace the J-lens directly.
  • Code, lenses and an interactive Neuronpedia demo have been released.

What the J++ lens changes

The work appears in the October 2026 edition of Circuits Updates, a series on the Transformer Circuits site where the Anthropic interpretability team shares developing ideas. The team asks readers to treat the results like a colleague sharing preliminary experiments at a lab meeting, not like a mature paper.

The J++ lens comes out of an Anthropic Fellows Program project led by Kola Ayonrinde, with Jack Lindsey as co-author. It builds on the Jacobian lens (J-lens), which Anthropic introduced in a recent paper as a technique for reading out the concepts a model represents at its intermediate layers.

Fixing noisy gradients

The J-lens is computed by averaging Jacobians over a large number of source token positions, with every position weighted equally. According to the team, some of these positions contribute very noisy gradients that corrupt the average.

The main change in the J++ lens is to cluster the activations at each layer and fit a separate Jacobian to each cluster. These cluster-specific Jacobians are then combined using learned weights. In practice, the team says, those weights mostly serve to discard the clusters that produce poor readouts.

The J++ lens also includes two further elements:

  • The improvement introduced by the R-lens, which uses layerwise relevance propagation when performing backward passes to compute the Jacobians.
  • Post-hoc filtering of spurious readout tokens.

Drop-in replacement and release

Despite the added steps, the result is still a single linear map at each layer. This means the J++ lens can be used as a drop-in replacement for the J-lens, according to the team.

Across Anthropic's evaluations, the J++ lens more reliably surfaces intermediate latent variables used for the model's reasoning than the J-lens or the R-lens. The full post contains more details and experiments.

Anthropic has also released its code and lenses, along with an interactive demo hosted on Neuronpedia. For researchers working on interpretability, this means they can try the method directly and swap it in where they already use the J-lens. The source describes the work as preliminary, and the team says some of the strands in these updates are ones it expects to publish more on in the coming months.

AnthropicJ++ lensJ-lensR-lensNeuronpedialanguage modelsresearch findings
Source
Anthropic Interpretability (Transformer Circuits): Circuits Updates — October 2026

Related news

Databricks launches AI Index tracking enterprise model costs and usage
ResearchNovember 16, 2026, 18:00

Databricks launches AI Index tracking enterprise model costs and usage

Databricks has launched the Databricks AI Index, a public resource tracking enterprise AI model adoption, coding-agent performance and spending. The company says the…

JetBrains asks 12 AI models for 1,200 programming hot takes
ResearchNovember 15, 2026, 18:02

JetBrains asks 12 AI models for 1,200 programming hot takes

Twelve AI models overwhelmingly attacked Python, Rust and TypeScript in a JetBrains experiment that collected 1,200 opinions about programming languages. The models also…

AI safety report calls for disclosures on chain-of-thought monitoring
Safety and ethicsNovember 10, 2026, 13:31

AI safety report calls for disclosures on chain-of-thought monitoring

A new report from the Institute for AI Policy and Strategy calls on frontier AI developers to disclose standardized indicators showing whether their models’…