Interactions Between Crosscoder Features: A Compact Proofs Perspective

• URL: http://arxiv.org/abs/2606.09940v1
• Date: Mon, 08 Jun 2026 00:15:44 GMT
• Title: Interactions Between Crosscoder Features: A Compact Proofs Perspective
• Abstract summary: 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.
• Score: 2.8682099615858156
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Dictionary learning methods like Sparse Autoencoders (SAEs) and crosscoders attempt to explain a model by decomposing its activations into independent features. Interactions between features hence induce errors in the reconstruction. We formalize this intuition via compact proofs and make five contributions. First, we show how, \textit{in principle}, a compact proof of model performance can be constructed using a crosscoder. Second, we show that an error term arising in this proof can naturally be interpreted as a measure of interaction between crosscoder features and provide an explicit expression for the interaction term in the Multi-Layer Perceptron (MLP) layers. We then provide three applications of this new interaction measure. In our third contribution we show that the interaction term itself can be used as a differentiable loss penalty. Applying this penalty, we can achieve ``computationally sparse'' crosscoders that retain $60\%$ of MLP performance when only keeping a single feature at each datapoint and neuron, compared to $10\%$ in standard crosscoders. We then show that clustering according to our interaction measure provides semantically meaningful feature clusters, and finally that sleeper agents have significant interactions. Code is available at https://github.com/chainik1125/crosscoders-feature-interactions/tree/arxiv.

Related papers

• How Do Linear Probes Emerge? A Circuit-Tracing Framework with Concept-Targeted Attribution [49.01642339390897]
  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.
  arXiv  Detail & Related papers  (2026-08-27T08:12:49Z)
• Decomposing Query-Key Feature Interactions Using Contrastive Covariances [75.38737409771085]
  We study the query-key space -- the bilinear joint embedding space between queries and keys.<n>It is when features in keys and queries align in these low-rank subspaces that high attention scores are produced.
  arXiv  Detail & Related papers  (2026-02-04T16:50:02Z)
• Identifying Intervenable and Interpretable Features via Orthogonality Regularization [48.938969291033665]
  We disentangle the decoder matrix into almost orthogonal features.<n>This reduces interference and superposition between the features, while keeping performance on the target dataset essentially unchanged.<n>Our code is available under $texttthttps://github.com/mrtzmllr/sae-icm$.
  arXiv  Detail & Related papers  (2026-02-04T16:29:14Z)
• 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)
• Explaining Caption-Image Interactions in CLIP Models with Second-Order Attributions [28.53636082915161]
  Dual encoder architectures like Clip models map two types of inputs into a shared embedding space and predict similarities between them.<n>Common first-order feature-attribution methods explain importances of individual features and can, thus, only provide limited insights into dual encoders.<n>We first derive a second-order method enabling the attribution of predictions by any differentiable dual encoder onto feature-interactions between its inputs.
  arXiv  Detail & Related papers  (2024-08-26T09:55:34Z)
• Triple-Encoders: Representations That Fire Together, Wire Together [51.15206713482718]
  Contrastive Learning is a representation learning method that encodes relative distances between utterances into the embedding space via a bi-encoder. This study introduces triple-encoders, which efficiently compute distributed utterance mixtures from these independently encoded utterances. We find that triple-encoders lead to a substantial improvement over bi-encoders, and even to better zero-shot generalization than single-vector representation models.
  arXiv  Detail & Related papers  (2024-02-19T18:06:02Z)
• A Generically Contrastive Spatiotemporal Representation Enhancement for 3D Skeleton Action Recognition [10.403751563214113]
  We propose a Contrastive Spatiotemporal Representation Enhancement (CSRE) framework to obtain more discriminative representations from the sequences.<n>Specifically, we decompose the representation into spatial- and temporal-specific features to explore fine-grained motion patterns.<n>To explicitly exploit the latent data distributions, we employ the attentive features to contrastive learning, which models the cross-sequence semantic relations.
  arXiv  Detail & Related papers  (2023-12-23T02:54:41Z)
• Human-Object Interaction Detection via Disentangled Transformer [63.46358684341105]
  We present Disentangled Transformer, where both encoder and decoder are disentangled to facilitate learning of two sub-tasks. Our method outperforms prior work on two public HOI benchmarks by a sizeable margin.
  arXiv  Detail & Related papers  (2022-04-20T08:15:04Z)
• GEN-VLKT: Simplify Association and Enhance Interaction Understanding for HOI Detection [17.92210977820113]
  We propose Guided-Embedding Network(GEN) to attain a two-branch pipeline without post-matching. For the association, previous two-branch methods suffer from complex and costly post-matching. For the interaction understanding, previous methods suffer from long-tailed distribution and zero-shot discovery.
  arXiv  Detail & Related papers  (2022-03-26T01:04:13Z)
• Disentangled Representation Learning for Text-Video Retrieval [51.861423831566626]
  Cross-modality interaction is a critical component in Text-Video Retrieval (TVR) We study the interaction paradigm in depth, where we find that its computation can be split into two terms. We propose a disentangled framework to capture a sequential and hierarchical representation.
  arXiv  Detail & Related papers  (2022-03-14T13:55:33Z)
• BasConv: Aggregating Heterogeneous Interactions for Basket Recommendation with Graph Convolutional Neural Network [64.73281115977576]
  Within-basket recommendation reduces the exploration time of users, where the user's intention of the basket matters. We propose a new framework named textbfBasConv, which is based on the graph convolutional neural network. Our BasConv model has three types of aggregators specifically designed for three types of nodes.
  arXiv  Detail & Related papers  (2020-01-14T16:27:32Z)

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.