The paper introduces Contrastive Projection, a method that reads a transformer’s internal states by differencing the hidden states of two closely matched prompts and projecting the difference through the unembedding layer. This approach cancels shared components and highlights the distinctions between prompts, effectively revealing steering vectors and domain-to-domain mappings such as metaphor. The technique is training‑free, operates at every position, sub‑layer, and head, and has been validated across multiple architectures and initialization seeds.
arXiv:2608.30720v1 Announce Type: new Abstract: Representational similarity is foundational to analyses of deep networks, yet distances between point-valued representations are not intrinsically tied...
arXiv:2607. 03502v1 Announce Type: cross Abstract: Frontier LLMs can perform multi-step reasoning over content-free filler tokens like dots or counting sequences, producing correct answers with no visible chain-of-thought (CoT).
The paper introduces Sparse Readout Prism (SRP), a method that decomposes a language model’s readout matrix into sparse features, allowing logit‑lens scores to be expressed as sums of feature contributions. SRP reveals that lens readings depend on the corpus used to fit the readout, a phenomenon called corpus conditionality, and that the dominant readout feature remains stable across different corpora. By replacing the original readout with SRP’s sparse approximation, the authors recover 8.9–17.3 percentage points more of the tested logit differences than six geometric‑relation baselines, and ablating features shifts logit differences proportionally to their SRP contributions.
arXiv:2510. 25013v2 Announce Type: replace-cross Abstract: Mechanistic interpretability aims to reverse-engineer large language models (LLMs) into human-understandable computational circuits.
What do a language model's hidden states say about the organization of a single text? From one forward pass, without training, we score every token position on two properties.
arXiv:2606. 12629v3 Announce Type: replace-cross Abstract: The standard basis of transformer hidden states is a training-free, architecture-general feature basis for detecting concepts and, in language models, steering them; with no learned dictionary.
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.
The paper presents a mean‑field analysis of attention in language models, defining an average attention kernel that propagates representations layer by layer. When conditioned on a whole corpus, the kernel predicts the average evolution of representation geometry; when conditioned on a single context, it predicts the expected geometry for that context. The difference between actual attention and the mean‑field prediction—called the mean‑field deviation—captures context‑specific computation, revealing how models diverge from average behavior during training and in few‑shot tasks.
arXiv:2606. 08105v1 Announce Type: new Abstract: When attention concentrates on a single token, a sink, what is the model actually computing?
The paper compares latent representations in Selective State Space Models (SSMs) like Mamba and Transformers such as Pythia using Sparse Autoencoders. Across a 10‑million token corpus, 99.98% of Mamba features align closely with Pythia’s, supporting the Universality Hypothesis that core semantic representations are similar across architectures. A tiny 0.02% of features diverge, with Mamba’s recurrent bottleneck causing it to compress syntactic anomalies into polysemantic neurons, whereas Pythia’s attention can isolate distinct formatting edge‑cases.
arXiv:2608. 10251v1 Announce Type: cross Abstract: A transformer's answer lives on one axis: the direction its unembedding reads.
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.