arXiv:2608. 03629v1 Announce Type: new Abstract: A companion paper studies when activation patching and weight-space ablation agree, inside an idealized model where a conditional computation is carried additively through a residual stream.
arXiv:2608. 03620v1 Announce Type: cross Abstract: Activation patching and weight-space ablation both claim a component is causally responsible for a behavior, yet they act on different objects: one forward pass versus the parameters behind every forward pass.
arXiv:2605. 23393v2 Announce Type: replace-cross Abstract: Mechanistic interpretability of transformers requires identifying not just which components matter but how they compose into the computational route that produced a prediction.
arXiv:2602. 18849v2 Announce Type: replace-cross Abstract: We develop a sensitivity analysis for transformer attention in a geometry aligned with tokenwise computation.
arXiv:2606. 21876v2 Announce Type: replace-cross Abstract: The Categorical Jacobian of Zhang et al.
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.
FishBack introduces a pullback Fisher geometry approach for activation steering in transformers, challenging the common Euclidean assumption of intermediate activation spaces. By deriving a closed‑form steering direction based on the Fisher information metric of the softmax layer, the method achieves target concept changes with minimal off‑target distortion, especially in early and middle layers. Experiments on GPT‑2 Small, Llama‑3‑8B, and Qwen3‑8B demonstrate significant reductions in off‑target KL divergence compared to existing steering baselines.
The paper investigates how new concepts can be integrated into unified multimodal models (UMMs) by separating generation and understanding objectives through a novel visual entity bound to a single task direction. Experiments show that the effectiveness of cross‑task usability depends on where the concept is injected into the shared computation, with a mid‑stack alignment objective achieving high concept acquisition with minimal loss to overall performance. The study highlights that unified weights alone are insufficient; the two directions must share a semantic format at the entry point for efficient concept integration.
The paper argues that diffusion-based action policies can use a frozen, observation‑free backbone as a reusable trajectory prior, with task adaptation handled entirely by the conditioning pathway. By pretraining a general action head on forward‑kinematics data and then freezing it, the authors show that a single backbone can match or outperform normally trained models on MimicGen and LIBERO. Their experiments reveal that a small 5 M‑parameter MLP backbone can rival large U‑Net and transformer backbones, indicating that action backbones are often over‑parameterized and that image‑style architectures may not be the best fit for low‑dimensional action generation.
arXiv:2609.01231v1 Announce Type: new Abstract: We prove that a multi-head scaled dot product attention can be viewed as a parameter identification strategy. The ratio of unidentified parameters to t...
arXiv:2607. 26988v1 Announce Type: cross Abstract: What types of decision problems can a causally masked, finite-precision transformer solve for inputs of arbitrary length?
The paper introduces OAttention, a token‑level attention mechanism that assigns each token a presence coefficient based on its hidden representation. This coefficient both gates the token’s output and weights its contribution to other tokens, making zero‑vector tokens behave as true zeros and enabling exact null‑receiver, null‑source, and empty‑support properties. The authors extend this idea to local O‑components and an O‑Transformer, and demonstrate small performance changes when retrofitting a pretrained TabPFN model.
The paper investigates whether a global workspace—a set of verbalisable, causally potent representations—emerges in transformer models that use recurrence instead of a stack of distinct layers. Using a Jacobian lens extended with a virtual‑unrolling adapter, the authors analyze two recurrent transformer architectures, Ouro‑2.6B and Huginn‑0125, and compare them to a standard Qwen3.6‑27B baseline. They find that a workspace does form in the iterated parts of both models, but recurrence alters how it can be accessed: Ouro reconstructs workspace content in every loop and requires writes and ablations across all loops, whereas Huginn forwards content across all recurrences but limits reads, writes, and ablations to a sliding window of about two recurrences.