arXiv Machine Learning
Aug 12
Off-Axis, On Purpose: Where a Transformer Computes Concepts and Why it Does So
arXiv:2608. 10251v1 Announce Type: cross Abstract: A transformer's answer lives on one axis: the direction its unembedding reads.
What the Primary Lund Plane Omits: A Model-Stitching Diagnosis of a Jet-Tagger Representation Gap
The study investigates why Lund‑plane and Particle Transformer (ParT) jet taggers can perform differently by using model stitching to isolate representation from computation. By linking frozen early layers of one network to frozen late layers of another, the authors show that a known deficit in the primary Lund representation is purely representational, not computational, and that this gap is invisible in AUC but appears in background rejection at low signal efficiency. Rebuilding the representation on the full Cambridge/Aachen declustering tree largely closes the gap, revealing that the primary branch carries most of the discriminative power while secondary splittings mainly make the representation learnable.
Related stories
arXiv Machine Learning
Sep 29
The Residual Stream's Effective Depth
The paper introduces effective depth (Deff), a scalar diagnostic that treats a transformer’s layer‑wise residual stream as a discrete‑time process and measures how representation similarity decays with layer distance. Across sixteen decoder‑only language models, Deff reveals that most models exhibit a lower similarity decay than the closed‑form reference, indicating correlated residual updates rather than unused depth. The study also shows that this effect is robust to various controls and persists early in training, suggesting Deff is a global accumulated‑state diagnostic rather than a capability score.
arXiv Computation and Language
Sep 10
Through the Looking Glass: Directly Reading and Writing Transformers
The paper investigates how many transformer components influence a token prediction by measuring the absolute contribution of each unit and channel to the logit. It finds that thousands of components contribute to a single prediction, yet a small subset—often just dozens—carries the majority of the predictive mass. Across models ranging from 124 M to 7 B parameters, the proportion of the model involved in a prediction remains around one to three percent, independent of size, and the study demonstrates that specific components can be directly read and written to modify model behavior without additional training.
arXiv Machine Learning
5d ago
Combinatorial Residual Network: a decay-chain-aware classifier without explicit jet assignment
arXiv:2610. 40749v1 Announce Type: cross Abstract: Final states with several jets from the decays of heavy particles pose a combinatorial problem, since the jets must be assigned, at least implicitly, to their parent particles.
arXiv AI
Aug 10
Post-Grokking Collapse at the Representation-Readout Interface in Muon-Trained Transformers
arXiv:2608. 07436v1 Announce Type: new Abstract: Under the standard split, Muon gets hidden matrices and AdamW embeddings/output head.
arXiv AI
Jul 7
Input Pathways Shape Few-Shot, Not Zero-Shot, Binding in Tiny Transformers: A Fully-Enumerable Study
arXiv:2607. 04926v1 Announce Type: cross Abstract: How does the way information reaches a transformer -- as symbolic tokens, a clean per-factor "oracle" code, or an entangled perceptual vector -- shape whether it binds that information compositionally?
arXiv Machine Learning
Oct 22
Reader--Writer Coupling in Transformer MLPs
The paper investigates the dual roles of hidden channels in transformer MLPs, treating each channel as a reader that decides when to respond and a writer that determines what to add back to the residual stream. By examining the joint Jacobian of the MLP’s update, the authors identify a signed pairing coordinate that reveals a small core of strongly coupled reader–writer pairs, often with opposing signs. This opposed core appears early in training and persists across models, influencing how channels contribute to later layers and affecting the model’s output distribution.
arXiv Machine Learning
Sep 14
Where Decoder Cosine Similarity Fails for SAE Feature Flow Discovery
arXiv:2609. 12591v1 Announce Type: new Abstract: Foundation models are increasingly adapted through fine-tuning, model editing, and alignment procedures while retaining previously acquired capabilities.
arXiv Machine Learning
Jul 14
Gauge dependence and structured-output corruption in sign-branched repetition penalties: measurements across models, inference stacks, and alternative repetition controls
arXiv:2607. 09791v1 Announce Type: new Abstract: The multiplicative repetition penalty shipped across the LLM inference ecosystem (HuggingFace, vLLM, llama.
arXiv AI
Sep 7
How Does mHC Use Its Residual Streams? Selective Routing and Near-Identity Mixing
The paper investigates how the four‑stream manifold‑constrained hyper‑connection (mHC) residual pathway in DeepSeek‑V4‑Flash is actually used. It finds that read/write routing is concentrated, typically involving only two streams per block, and that the dominant stream shifts across layers while representations stay directionally distinct. Residual mixing is modest, mainly in early layers, and late mixing contributes little to performance, whereas early mixing is crucial for perplexity and task scores.
arXiv Machine Learning
Sep 17
Reaching Every Position Without Searching: Rotating Sparse Wiring on the Hypercube as a Substitute for Attention
arXiv:2609. 18145v1 Announce Type: new Abstract: Attention pays, at every layer and for every input, the cost of searching for whom to connect.
arXiv Machine Learning
Sep 18
Stiefel Attention: When the Geometry of Transformer Projection Matrices Dominates Optimizer Choice---and When It Does Not
The paper introduces Stiefel Attention, which constrains the query and key projection matrices of transformers to the Stiefel manifold and optimizes them with a Riemannian Adam variant. It demonstrates that this approach yields steepest‑descent updates, is well‑conditioned, and preserves learned attention geometry during weight decay. Empirical results show significant accuracy gains on modular arithmetic grokking and CIFAR‑10 patches, with the improvement attributed to a step‑scale‑free update rule rather than equivariance or projector changes.