Bag of Dims: Training-Free Mechanistic Interpretability via Dimension-Level Sign Patterns

• URL: http://arxiv.org/abs/2606.12629v2
• Date: Wed, 17 Jun 2026 04:35:07 GMT
• Title: Bag of Dims: Training-Free Mechanistic Interpretability via Dimension-Level Sign Patterns
• Abstract summary: We show the standard basis of transformer hidden states already provides a training-free, architecture-general feature basis.<n>We validate this Bag of Dims framework across seven models spanning language.
• Score: 0.0
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: We show the standard basis of transformer hidden states already provides a training-free, architecture-general feature basis. Individual dimensions encode semantic content via their signs (+/-1) and confidence via their magnitudes, acting as independent binary registers; a feature is a subset of dimensions with a consistent sign pattern, read by counting sign agreements with no learned rotation. We validate this Bag of Dims framework across seven models spanning language (Qwen 3.5-4B, Gemma 3-4B, Mistral 7B, Qwen3-32B), vision (DINOv2, ViT-Base), and audio (AST). Signs alone carry predictive content: unit-magnitude sign patterns preserve 60-93% top-5 next-token accuracy through the LM head, and decoder-free Hamming scoring reaches 80-90% top-4096. From a single-token cache (one forward pass per token, no context, no labels), we detect 175 categories at AUC 0.97-0.99 by sign agreement; a trained probe adds only +0.018 AUC and converges to axis-aligned weights. These features are causally operative: they survive the K/V attention projections, trace to the FFN neuron coalitions that write them (random-weight controls never reproduce this), and flipping a feature's signs during the live forward pass suppresses its concept across four language models, magnitude-matched and concept-specific. Dimensions stay independent throughout (pairwise mutual information below 0.006 bits). The structure is not specific to language: the same per-dimension signs appear in self-supervised vision (DINOv2, 9/12 ImageNet superclasses), supervised vision (ViT-Base, 11/12), and audio (AST, 50/50 ESC-50 categories), so it reflects transformer training in general, not the language-modeling objective. The standard basis already suffices for feature reading at one forward pass, no optimization, no GPU-days. The open problem shifts from finding the right rotation to cataloging what each dimension encodes.

Related papers

• NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction [76.38136259305638]
  We introduce NCP-ArchPreview, a latent-space language model that pushes autoregressive pretraining beyond standard next-token prediction (NTP)<n> NCP-ArchPreview builds a latent space by constructing a product-quantized concept vocabulary directly from its hidden states, and subsequently learns to predict future concepts via a dedicated Concept Module.<n>We scale this architecture to 8.9B parameters and train it on 5.73T tokens from the Dolma-3 dataset, marking the largest demonstration of a latent-space language model to date.
  arXiv  Detail & Related papers  (2026-09-09T18:12:43Z)
• TractoGraphVLM: A Unified Vision-Language Framework for White Matter Tractography [0.30586855806896046]
  TractoGraphVLM is a unified framework for four tasks, bundle classification, text-to-tract, anatomical, and visual question answering.<n>GPS transformer produces bundle embeddings aligned with a frozen BiomedBERT text encoder via contrastive learning.<n>BioGPT decoder with visual captions generates tokens and answers.
  arXiv  Detail & Related papers  (2026-08-16T16:14:48Z)
• UniProbe: A Learnable Token-Level Hallucination Detector for Large VLMs using Multi-Structural Internal Representations [61.85916280854257]
  Large Vision-Language Models (LVLMs) achieve impressive visual reasoning and dialogue capabilities, yet frequently hallucinate content unsupported by the visual input.<n>We introduce textbfUniProbe, a lightweight, unified, learnable detector that models a frozen LVLM's heterogeneous computational trace from a single forward pass.
  arXiv  Detail & Related papers  (2026-08-11T12:01:59Z)
• Kathleen Writes: Autoregressive Generation and Data Scaling Without Attention [0.0]
  We show that a byte-level, attention-free architecture can match strong baselines on classification at 450-700K parameters, without pretraining.<n>We introduce FORM DISTANCE, a non-parametric, gaming-resistant instrument for "reads like text"<n>We also report four architectural additions that did not help, and a computed lexicon reaching 94% of a learned table's top-1 accuracy at one fifth of the parameters.
  arXiv  Detail & Related papers  (2026-08-05T10:44:26Z)
• Closing the Modality Gap in Zero-Shot HAR: Contrastive Training and Separability-Optimized Prototypes on IMU Data [0.0]
  We evaluate seven configurations combining three inference methods with two training pipelines on the PAMAP2 dataset.<n>We find that the modality gap is a training-time phenomenon governed by the objective.<n>We also demonstrate that overall accuracy is a misleading primary metric when test-set class distributions are imbalanced.
  arXiv  Detail & Related papers  (2026-06-09T12:39:41Z)
• Automated Proving of Shannon-Type Entropy Inequalities via Fine-Tuned Language Models and Guided Tree Search [50.16356451328644]
  Proving Shannon-type entropy inequalities is a fundamental task in information theory.<n>We investigate whether small-scale large language models, fine-tuned on atomic proof steps, can automate this process.<n>We show that GPT-5.5 solves 1.7% samples under zero-shot prompting while Psitip solves 33.3% samples.
  arXiv  Detail & Related papers  (2026-06-04T05:43:12Z)
• Right Makes Might: Aligning Verified Hidden States Empowers RL Reasoning [55.264863369127774]
  Current methods reduce each correct rollout to a single reward bit, ignoring the geometric structure shared among their hidden states.<n>We propose Hidden-Align, an auxiliary loss function that aligns the last-layer hidden states of correct rollouts at the anchor token during RL training, with zero overhead in both training and inference.<n>On eight mathematical reasoning benchmarks, Hidden-Align improves average pass@1 over the DAPO baseline by 3.8, 6.2, and 5.4 percentage points on Qwen3-1.7B, 4B, and 14B respectively, with consistent pass@k gains across all three scales.
  arXiv  Detail & Related papers  (2026-06-02T06:51:15Z)
• Kronecker Embeddings: Byte-Level Structured Token Representations for Parameter-Efficient Language Models [0.0]
  Large language models route every input through a learned embedding table of shape.<n>We introduce Kronecker Embeddings, a deterministic byte-level character-position factorization.<n>It eliminates 91--94% of input-side trainable parameters at frontier scale.
  arXiv  Detail & Related papers  (2026-05-28T06:53:18Z)
• Unextractable Protocol Models: Collaborative Training and Inference without Weight Materialization [58.14514930760722]
  We consider a decentralized setup in which the participants collaboratively train and serve a large neural network.<n>In this setup, we explore the possibility of unmaterializable weights, where a full weight set is never available to any one participant.<n>We introduce Unextractable Protocol Models (UPMs): a training and inference framework that leverages the sharded model setup.
  arXiv  Detail & Related papers  (2026-05-22T10:24:57Z)
• Amplifying, Not Learning: Fine-Tuned AI Text Detectors Amplify a Pretrained Direction [51.56484100374058]
  Text detectors amplify a pretrained typicality axis; they do not construct an AI-vs-human boundary.<n>On raw encoders before any task supervision, projecting onto centroid(AI)-centroid(HC3) achieves NYT-vs-HC3 AUROC 0.806/0.944/0.834 across three architectures.<n>On RoBERTa-base, raw projection exceeds fine-tuning; on RoBERTa-base, full fine-tuning reduces discrimination below raw on both fluent-formal populations tested.
  arXiv  Detail & Related papers  (2026-05-20T19:08:38Z)
• UniRefiner: Teaching Pre-trained ViTs to Self-Dispose Dross via Contrastive Register [30.077265994339303]
  We argue that any token failing to encode location-aligned semantics should be treated as a spurious artifact.<n>We propose UniRefiner, a universal refinement framework that teaches pre-trained Vision Transformers to self-dispose of these artifacts.<n>Our method requires only a few epochs of fine-tuning on 5k images to refine diverse ViTs, including massive models like EVA-CLIP-8B and InternViT-6B.
  arXiv  Detail & Related papers  (2026-05-19T10:00:14Z)
• THEIA: Learning Complete Kleene Three-Valued Logic in a Pure-Neural Modular Architecture [0.0]
  THEIA is a 2.75M modular neural architecture that learns the complete Kleene three-valued logic (K3) truth table from task data without external symbolic inference or hand-encoded K3 gate primitives.<n> Transformer baselines also reach >99% on all 39 rules, and flats match THEIA on Phase-1 accuracy within 0.04pp.
  arXiv  Detail & Related papers  (2026-04-13T10:44:15Z)
• Learning the Signature of Memorization in Autoregressive Language Models [3.6048665052465663]
  We introduce the first transferable learned attack, enabled by the observation that fine-tuning any model on any corpus yields unlimited labeled data.<n>This removes the shadow model bottleneck and brings membership inference into the deep learning era.
  arXiv  Detail & Related papers  (2026-04-03T17:17:51Z)
• Next-Embedding Prediction Makes Strong Vision Learners [68.55755328850634]
  We train models to generate embeddings to perform predictive tasks directly.<n>Next-Embedding Predictive Autoregression (NEPA) achieves strong results across tasks.<n>We believe generative pretraining from embeddings provides a simple, scalable, and potentially modality-agnostic alternative to visual self-supervised learning.
  arXiv  Detail & Related papers  (2025-12-18T18:59:58Z)
• Perception Encoder: The best visual embeddings are not at the output of the network [70.86738083862099]
  We introduce Perception (PE), a vision encoder for image and video understanding trained via simple vision-language learning.<n>We find that contrastive vision-language training alone can produce strong, general embeddings for all of these downstream tasks.<n>Together, our PE family of models achieves best-in-class results on a wide variety of tasks.
  arXiv  Detail & Related papers  (2025-04-17T17:59:57Z)
• Telling Left from Right: Identifying Geometry-Aware Semantic Correspondence [80.6840060272386]
  This paper identifies the importance of being geometry-aware for semantic correspondence. We show that incorporating this information can markedly enhance semantic correspondence performance. Our method achieves a PCK@0.10 score of 65.4 (zero-shot) and 85.6 (supervised) on the challenging SPair-71k dataset.
  arXiv  Detail & Related papers  (2023-11-28T18:45:13Z)
• SignBERT+: Hand-model-aware Self-supervised Pre-training for Sign Language Understanding [132.78015553111234]
  Hand gesture serves as a crucial role during the expression of sign language. Current deep learning based methods for sign language understanding (SLU) are prone to over-fitting due to insufficient sign data resource. We propose the first self-supervised pre-trainable SignBERT+ framework with model-aware hand prior incorporated.
  arXiv  Detail & Related papers  (2023-05-08T17:16:38Z)
• Unifying Language Learning Paradigms [96.35981503087567]
  We present a unified framework for pre-training models that are universally effective across datasets and setups. We show how different pre-training objectives can be cast as one another and how interpolating between different objectives can be effective. Our model also achieve strong results at in-context learning, outperforming 175B GPT-3 on zero-shot SuperGLUE and tripling the performance of T5-XXL on one-shot summarization.
  arXiv  Detail & Related papers  (2022-05-10T19:32:20Z)
• HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units [81.53783563025084]
  We propose an offline clustering step to provide aligned target labels for a BERT-like prediction loss. A key ingredient of our approach is applying the prediction loss over the masked regions only. HuBERT shows up to 19% and 13% relative WER reduction on the more challenging dev-other and test-other evaluation subsets.
  arXiv  Detail & Related papers  (2021-06-14T14:14:28Z)

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.