The paper investigates how pretraining and alignment stages of transformer training affect weight updates in different matrix groups. By examining the residual stream’s shared coordinate system, it finds that alignment updates to query and key matrices align with the principal input subspace, whereas updates to output matrices remain largely random relative to the prediction subspace. This pattern persists across objectives and training stages, indicating that the geometry of updates is determined by the training phase rather than the specific objective.
By Valeria Ruscio, Keiran Thompson
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.
By Mark Oskin
arXiv:2608. 10251v1 Announce Type: cross Abstract: A transformer's answer lives on one axis: the direction its unembedding reads.
By Mark Oskin
The paper introduces the Communication Map, a method that charts every potential communication channel in a transformer model using only its weights. It generalizes previous coupling metrics into a single coefficient covering all 18 connection classes, revealing that 70‑89% of head pairs are non‑randomly oriented and identifying strong or avoiding couplings. The authors demonstrate the map’s utility by recovering known induction circuits and uncovering a two‑dimensional stream subspace whose removal eliminates induction capabilities across several models.
By Richard Zhe Wang
arXiv:2606. 31963v1 Announce Type: new Abstract: Modern LLM workflows move coordinate-indexed objects across checkpoints: steering vectors, sparse autoencoders, top-$k$ neuron sets, attribution lists, and merge alignments.
By John Sweeney
arXiv:2608.20530v1 Announce Type: new
Abstract: Speculative decoding accelerates language-model inference by drafting future tokens that the target model verifies in parallel. A diffusion-style block...
By Matan Rusanovsky, Yoav Miron, Roy Uziel, Omer Belhasin, Ran Zilberstein, Maor Ashkenazi, Michael Elad
arXiv:2609.24635v1 Announce Type: new
Abstract: When a language model reads an operation such as "Swap the contents of Box F and Box B", its forward pass writes keys and values for those tokens into...
By Lingfeng Wu, Behzad Shomali
The paper investigates whether a frozen transformer can use another model’s KV cache as a substitute for its own, finding that this approach fails because the reader only accepts caches within a narrow tolerable set. Experimental results show that none of eight substitution methods improve performance over a baseline, while a different strategy—adding the sender’s cache as extra context and only adapting queries—does outperform the baseline. The authors conclude that pursuing faithful cache substitution is misguided and that augmenting the reader’s own cache is the more promising direction.
By Adam Deng, Shicheng Liu, Yada Zhu
The paper proposes that two architectural assumptions—(1) attention and MLPs share a key‑value form <phi(S)>U, and (2) components read from an additive residual stream—are sufficient to answer three interpretability questions: component interaction, information routing, and token attribution. By treating these selections as a computational graph, the authors develop Unpack, a backward attribution method that validates interaction scores, recovered routes, and token attribution against established tests across models ranging from 160M to 6.9B parameters. The study also shows that contribution and causal effect can differ, with a recognizable signature in how components change when a task is removed.
By Po-Kai Chen, Aske Plaat, Niki van Stein
arXiv:2607. 24797v1 Announce Type: cross Abstract: In the literate human brain, reading and writing are two doubly-dissociable systems: a ventral decoding route (impaired in pure alexia) and a fronto-parietal encoding route (impaired in pure agraphia), sharing a partial orthographic core.
By Diego Salda\~na Ulloa
Delta N-gram introduces a forward‑only write path into a language model’s native hashed N‑gram table, preserving the pretrained table while adding evidence through a sparse patch Δ. The method uses a 4.1 M‑parameter writer that emits gated targets per event and applies a block‑averaged delta rule to keep each row a convex combination of its old value and the new targets. Experiments on a 0.6 B prototype and Qwen3.8‑Flash‑Next show that the delta writer maintains more document‑specific memory and reduces drift compared to online gradient writes, achieving stable performance over thousands of writes.
whyItMatters":"The approach allows continual updates to a language model’s memory without altering its pretrained parameters, enabling efficient, stable, and document‑specific knowledge retention during inference."
By Hao Wu
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.
By Pengxiang Zhao, Xing Li, Xianzhi Yu, Wei Guo, Zhenhua Dong