Every Component is a Lookup: Token Attribution and Composition from a Single Decomposition

• URL: http://arxiv.org/abs/2605.23393v1
• Date: Fri, 22 May 2026 09:03:01 GMT
• Title: Every Component is a Lookup: Token Attribution and Composition from a Single Decomposition
• Abstract summary: We exploit a shared key-value template $(S)U$ to develop Unpack, a backward recursion that decomposes credit through both sublayers.<n>On GPT-2 small, the method recovers all three composition connections described by Wang et al. (2023), including the mode-specific routing of each connection.<n>Across the Pythia family from 160M to 6.9B parameters, this suppression pattern is consistently recovered at every scale.
• Score: 1.3290717106567922
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Mechanistic interpretability of transformers requires identifying not just which components matter but how they compose into the computational route that produced a prediction. Both attention and MLP follow a shared key-value template $φ(S)U$. We exploit this structure to develop Unpack, a backward recursion that decomposes credit through both sublayers, producing interaction strengths between any two components, named end-to-end paths with K/Q/V composition labels, and per-token attribution from a single forward pass, without intervention, gradients, or auxiliary training. We evaluate on the indirect object identification task. On GPT-2 small, the method recovers all three composition connections described by Wang et al. (2023), including the mode-specific routing of each connection (K, Q, or V). To test token-level attribution beyond trivial copying, we compare two occurrences of the same name in the same decomposition: the first mention retains strong credit while the duplicate-detection position is suppressed, a pattern absent in matched control prompts. Across the Pythia family from 160M to 6.9B parameters, this suppression pattern is consistently recovered at every scale, demonstrating that the method tracks mechanistic structure without ground-truth circuit labels. Code is available at https://github.com/Fun-Cry/unpacklm.

Related papers

• Intrinsic Structure: Spectral Identifiability for Mechanistic Interpretability [51.56484100374058]
  We prove the first identifiability theorem for a mechanistic-interpretability primitive.<n>The spectrum is an identifiable, model-intrinsic fingerprint with a stated error bar, not a legible decomposition.
  arXiv  Detail & Related papers  (2026-08-10T19:42:01Z)
• SEER: A Self-Grounded Evidence Interface for Controlled Spatial Relation Classification [71.9783246097687]
  SEER hides candidate relations during pair localization, constructs a query-specific view with explicit subject/object roles, and retains the full image and sparse box geometry as complementary evidence.<n>For relation-choice protocols with exact inverse support, an optional refinement swaps the entity roles and changes the forward decision only when exactly one visual state obeys the corresponding inverse relation.<n>The unchanged protocol yields +4.35 to +11.79 on all 2,434 filtered EmbSpatial pair-relation questions across three models.
  arXiv  Detail & Related papers  (2026-08-04T13:16:15Z)
• SPARE: Structural Parameter-Free Affinity Regularization for Flow Matching [30.43836162456455]
  Regularizing internal representations has emerged as an effective accelerator, yet existing methods split into two families with complementary costs.<n>We propose Structural.<n>free Affinity Regularization (SPARE), a regularizer that matches the pairwise affinities of intermediate tokens to those of the clean latents.<n>On ImageNet $256 times 256$ with SiT backbones under matched 400K-iteration budgets, SPARE adds no encoder, head, or parameters and only 0.08 GB of training memory.
  arXiv  Detail & Related papers  (2026-08-03T09:50:51Z)
• Causal dictionary learning reveals and validates transcription-factor binding features in genomic language models [5.237172334460829]
  We introduce a framework that combines sparse dictionary learning with causal intervention to extract, validate, and causally test interpretable features in genomic foundation models.<n>We show that the naive validation of such features against position weight matrices is severely confounded by GC composition and repetitive elements.<n>We establish that specific features are emphcausally used to represent cell-type-specific TF binding, not merely motif presence.
  arXiv  Detail & Related papers  (2026-07-21T22:54:46Z)
• SliceGraph: Mapping Process Isomers in Multi-Run Chain-of-Thought Reasoning [60.650808962786364]
  We propose SliceGraph, a problem-model-cell graph built by mutual-kNN over sparse activation-key Jaccard similarity between CoT slices.<n>We show that correct CoTs sharing the same normalized answer split into multiple process families.<n>We also show that process families navigate the same atlas with distinct transition kernels under matched null controls.
  arXiv  Detail & Related papers  (2026-05-14T09:37:26Z)
• Geometric Factual Recall in Transformers [57.48371649045765]
  A common view casts internal weight matrices as associative memories over pairs of embeddings, requiring memorization counts that scale linearly with the number of facts.<n>We develop a theoretical and empirical account of an alternative, emphgeometric form of memorization in which learned embeddings encode structure directly.<n>In a controlled setting where a single-layer transformer must memorize random vectors from subjects to a shared attribute set, we prove that a logarithmic embedding dimension suffices.<n>We extend these results to the multi-hop setting, providing constructions with and without chain-of-thought that exhibit a provable
  arXiv  Detail & Related papers  (2026-05-12T17:22:22Z)
• Pairwise matrices for sparse autoencoders: single-feature inspection mislabels causal axes [2.741152471987327]
  Standard sparse-autoencoder protocol labels each feature from its top-activating contexts and validates by single-feature steering.<n>We propose the pairwise matrix protocol, co-varying steering coefficient with joint condition, and report three findings the standard one-corner protocol misses on Qwen3-1.7B-Instruct, replicated on Gemma-2-2B-it.<n>All three findings reproduce on Gemma with model-specific damage signatures; the matched-geometry control is CI-separated by 10x.
  arXiv  Detail & Related papers  (2026-05-04T21:11:21Z)
• The Structural Attention Tax: How Retrieval Format Hijacks In-Context Learning Independent of Content [11.226669792850645]
  We formalise a phenomenon we term the structural attention tax: knowledge graph (KG) triples.<n>KG triples capture 2-3x more attention per token than semantically equivalent natural-language text.<n>We develop a formal framework decomposing attention scores into semantic and structural components.
  arXiv  Detail & Related papers  (2026-04-21T08:47:00Z)
• To Copy or Not to Copy: Copying Is Easier to Induce Than Recall [5.057026826740146]
  Language models must arbitrate between parametric knowledge stored in their weights and contextual information in the prompt.<n>This work presents a mechanistic study of that choice by extracting an empharbitration vector from model activations on a curated dataset.
  arXiv  Detail & Related papers  (2026-01-17T14:46:29Z)
• SETR: A Two-Stage Semantic-Enhanced Framework for Zero-Shot Composed Image Retrieval [4.230223288110963]
  Zero-shot Composed Image Retrieval (ZS-CIR) aims to retrieve a target image given a reference image and a relative text, without relying on costly triplet annotations.<n>Existing CLIP-based methods face two core challenges: (1) union-based feature fusion indiscriminately aggregates all visual cues, carrying over irrelevant background details that dilute the intended modification, and (2) global cosine similarity from CLIP embeddings lacks the ability to resolve fine-grained semantic relations.
  arXiv  Detail & Related papers  (2025-09-30T09:41:52Z)
• Token Merging for Training-Free Semantic Binding in Text-to-Image Synthesis [98.21700880115938]
  Text-to-image (T2I) models often fail to accurately bind semantically related objects or attributes in the input prompts. We introduce a novel method called Token Merging (ToMe), which enhances semantic binding by aggregating relevant tokens into a single composite token.
  arXiv  Detail & Related papers  (2024-11-11T17:05:15Z)
• Semi-DETR: Semi-Supervised Object Detection with Detection Transformers [105.45018934087076]
  We analyze the DETR-based framework on semi-supervised object detection (SSOD) We present Semi-DETR, the first transformer-based end-to-end semi-supervised object detector. Our method outperforms all state-of-the-art methods by clear margins.
  arXiv  Detail & Related papers  (2023-07-16T16:32:14Z)
• Triplet Contrastive Learning for Unsupervised Vehicle Re-identification [55.445358749042384]
  Part feature learning is a critical technology for fine semantic understanding in vehicle re-identification. We propose a novel Triplet Contrastive Learning framework (TCL) which leverages cluster features to bridge the part features and global features.
  arXiv  Detail & Related papers  (2023-01-23T15:52:12Z)
• 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)
• Inducing Transformer's Compositional Generalization Ability via Auxiliary Sequence Prediction Tasks [86.10875837475783]
  Systematic compositionality is an essential mechanism in human language, allowing the recombination of known parts to create novel expressions. Existing neural models have been shown to lack this basic ability in learning symbolic structures. We propose two auxiliary sequence prediction tasks that track the progress of function and argument semantics.
  arXiv  Detail & Related papers  (2021-09-30T16:41:19Z)

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.