Sign in
← AI PulseAug 31, 2026

Deep · research · Single-source brief

Concept-Targeted Attribution Explores Linear Probe Emergence

A new framework, Concept-Targeted Attribution (CTA), trains attribution graphs with respect to linear probe directions to explain the emergence of internal concept representations.

By Illumora Editorial

Source · Aug 31, 2026, 4:00 AM · On Illumora · Aug 31, 2026, 4:03 AM

Media from the primary source — shown here so you can stay on Illumora.

Rewritten from one allowlisted primary — not independent enterprise reporting. Lanes →

Brief drafted by Illumora’s editorial model from the linked primary source. Ops desk reviews flagged pieces. How we write →

Read the source →arXiv cs.CL — How Do Linear Probes Emerge? A Circuit-Tracing Framework with Concept-Targeted Attribution
Save

A recent paper on arXiv cs.CL introduces Concept-Targeted Attribution (CTA), a framework designed to explain how internal concept representations arise within a prompt. Unlike traditional transcoder attribution graphs that focus on next-token probability, CTA trains attribution graphs specifically for linear probe directions. This approach aims to reveal probe-specific circuits that illuminate the internal computations behind a concept's emergence, independent of its expression in the generated token.

Key Points

  • Concept-Targeted Attribution (CTA) trains attribution graphs relative to a linear probe direction.
  • This method explains why an internal concept representation arises in a prompt.
  • CTA operates independently of whether the concept is expressed in the generated token.
  • Cross-Layer Transcoders demonstrate that probe-targeted graphs contain predictive structure.
  • Local features identify sparse components driving per-prompt classification.
  • Causal ablations indicate that probe-targeted and logit-targeted graphs capture distinct functional mechanisms.
  • Removing probe-relevant features reduces internal concept scores while largely preserving generated tokens.

Context

According to the authors, transcoder attribution graphs typically explain why a model assigns high probability to a particular next token. CTA shifts this focus by training attribution graphs with respect to a linear probe direction. This allows for the identification of probe-specific circuits that explain the emergence of internal concept representations. The framework uses Cross-Layer Transcoders to show that these probe-targeted graphs possess predictive structure, with graph-level features forecasting probe accuracy across four widely studied concept categories.

Why It Matters

This research provides a method for understanding the internal computations that enable linear probes to function. For researchers, it offers a way to connect probe performance to interpretable circuit structures, moving beyond simply knowing if a probe works to understanding how it works. This can inform the development of more transparent and reliable AI systems.

What To Do

  • Review the paper's methodology for training attribution graphs with respect to linear probe directions.
  • Note the findings from causal ablations regarding the distinct mechanisms captured by probe-targeted versus logit-targeted graphs.
  • Consider how the Concept-Targeted Attribution framework could be applied to analyze internal concept representations in other model architectures.