How Do Linear Probes Emerge? A Circuit-Tracing Framework with Concept-Targeted Attribution

• URL: http://arxiv.org/abs/2608.27510v1
• Date: Thu, 27 Aug 2026 08:12:49 GMT
• Title: How Do Linear Probes Emerge? A Circuit-Tracing Framework with Concept-Targeted Attribution
• Abstract summary: Concept-Targeted Attribution (CTA) trains attribution graphs with respect to a linear probe direction.<n>We show that CTA yields probe-specific circuits that explain why an internal concept representation arises in a prompt.
• Score: 49.01642339390897
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Transcoder attribution graphs are usually trained to explain why a model assigns high probability to a particular next token. We introduce Concept-Targeted Attribution (CTA), which instead trains attribution graphs with respect to a linear probe direction. CTA therefore yields probe-specific circuits that explain why an internal concept representation arises in a prompt, independently of whether it is expressed in the generated token. Using Cross-Layer Transcoders, we show that these probe-targeted graphs contain predictive structure: graph-level features predict probe accuracy across four widely studied concept categories ($ρ= 0.91$, $R^2 = 0.84$), while local features identify the sparse components driving per-prompt classification. This connects probe performance to interpretable circuit structure, allowing us to ask not only whether a probe works, but which internal computations make it work. Causal ablations further show that probe-targeted and logit-targeted graphs capture functionally distinct mechanisms. Removing probe-relevant features reduces internal concept scores while largely preserving generated tokens, whereas removing logit-relevant features changes the generated token in 92% to 100% of cases with near-zero effect on probe scores. CTA provides a framework for moving from behavioral probe accuracy to mechanistic explanations of probe performance, enabling more detailed audits of internal concept representations, including safety-critical ones. Our code is available at https://github.com/vedantpalit/concept-targeted-attribution

Related papers

• Motion-Based Tokenization for Cross-Dataset Egocentric Gaze Modeling [13.408937731789814]
  We formulate event-aligned, fixed-horizon angular displacement as an interpretable, event-conditioned motion vocabulary.<n>We compare it with event-only, spatial, absolute-angle, learned vector-quantized, and continuous representations.<n>We show that angular-motion tokens have lower target-domain regret than frozen-codebook VQ tokens in one transfer direction.
  arXiv  Detail & Related papers  (2026-08-24T08:00:23Z)
• Interactions Between Crosscoder Features: A Compact Proofs Perspective [2.8682099615858156]
  We show how a compact proof of model performance can be constructed using a crosscoder.<n>We show that an error term arising in this proof can naturally be interpreted as a measure of interaction between crosscoder features.<n>We then show that clustering according to our interaction measure provides semantically meaningful feature clusters.
  arXiv  Detail & Related papers  (2026-06-08T00:15:44Z)
• ADAG: Automatically Describing Attribution Graphs [50.460651620833055]
  We introduce textbfADAG, an end-to-end pipeline for describing these attribution graphs which is fully automated.<n>We then introduce a novel clustering algorithm for grouping features, and an LLM explainer--simulator setup which generates and scores natural-language explanations of the functional role of these feature groups.<n>We run our system on known human-analysed circuit-tracing tasks and recover interpretable circuits, and further show ADAG can find steerable clusters which are responsible for a harmful advice jailbreak in Llama 3.1 8B Instruct.
  arXiv  Detail & Related papers  (2026-04-08T21:34:37Z)
• CodeCircuit: Toward Inferring LLM-Generated Code Correctness via Attribution Graphs [13.488544043942495]
  We aim to investigate whether the model's neural dynamics encode internally decodable signals that are predictive of logical validity during code generation.<n>By decomposing complex residual flows, we aim to identify the structural signatures that distinguish sound reasoning from logical failure.<n>Analysis across Python, C++, and Java confirms that intrinsic correctness signals are robust across diverse syntaxes.
  arXiv  Detail & Related papers  (2026-02-06T03:49:15Z)
• Automated Circuit Interpretation via Probe Prompting [0.0]
  We present probe prompting, an automated pipeline that transforms attribution graphs into compact, interpretable subgraphs.<n>Across five prompts, probe-prompted subgraphs preserve high explanatory coverage while compressing complexity.<n>Compared to geometric clustering baselines, concept-aligned groups exhibit higher behavioral coherence.
  arXiv  Detail & Related papers  (2025-11-10T11:53:36Z)
• Probing the Probes: Methods and Metrics for Concept Alignment [0.0]
  It is widely assumed that a high probe accuracy indicates a CAV faithfully representing its target concept.<n>We show that the probe's classification accuracy alone is an unreliable measure of concept alignment.<n>We introduce a novel concept localization method based on spatial linear attribution.
  arXiv  Detail & Related papers  (2025-11-06T12:34:05Z)
• Position-aware Automatic Circuit Discovery [59.64762573617173]
  We identify a gap in existing circuit discovery methods, treating model components as equally relevant across input positions.<n>We propose two improvements to incorporate positionality into circuits, even on tasks containing variable-length examples.<n>Our approach enables fully automated discovery of position-sensitive circuits, yielding better trade-offs between circuit size and faithfulness compared to prior work.
  arXiv  Detail & Related papers  (2025-02-07T00:18:20Z)
• Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models [55.19497659895122]
  We introduce methods for discovering and applying sparse feature circuits.<n>These are causally implicatedworks of human-interpretable features for explaining language model behaviors.
  arXiv  Detail & Related papers  (2024-03-28T17:56:07Z)
• Efficient Computation of Counterfactual Bounds [44.4263314637532]
  We compute exact counterfactual bounds via algorithms for credal nets on a subclass of structural causal models. We evaluate their accuracy by providing credible intervals on the quality of the approximation.
  arXiv  Detail & Related papers  (2023-07-17T07:59:47Z)
• Measuring the Interpretability of Unsupervised Representations via Quantized Reverse Probing [97.70862116338554]
  We investigate the problem of measuring interpretability of self-supervised representations. We formulate the latter as estimating the mutual information between the representation and a space of manually labelled concepts. We use our method to evaluate a large number of self-supervised representations, ranking them by interpretability.
  arXiv  Detail & Related papers  (2022-09-07T16:18:50Z)
• Should Graph Convolution Trust Neighbors? A Simple Causal Inference Method [114.48708191371524]
  Graph Convolutional Network (GCN) is an emerging technique for information retrieval (IR) applications. This work focuses on the local structure discrepancy of testing nodes, which has received little scrutiny. We analyze the working mechanism of GCN with causal graph, estimating the causal effect of a node's local structure for the prediction.
  arXiv  Detail & Related papers  (2020-10-22T15:21:47Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.