arXiv:2608. 06839v1 Announce Type: new Abstract: Artificial Neural networks (ANNs) are often treated as black-box models, making explainability a central challenge in deep learning.
arXiv:2608.29530v1 Announce Type: cross Abstract: Modern systems in artificial intelligence (AI) somehow excel in domains for which they seem poorly suited. Intelligence has traditionally been modele...
arXiv:2603. 06592v2 Announce Type: replace-cross Abstract: Contemporary studies in mechanistic interpretability have uncovered many puzzling phenomena in the neural information processing of Transformer-based language models, such as induction heads, function vectors, and the Hydra effect.
The study investigates whether large language models (LLMs) develop internal representations similar to cognitive maps used by animals for flexible behavior. Using relational reasoning tasks and causal mosaic experiments, the authors found that LLMs beyond ten billion parameters can infer unstated relationships, generalize to synthetic entity names, and form abstract relational structures that combine with entity identities. These relational representations are stored as pairwise links and composed retrospectively, mediated by a sparse, distributed key‑value system that can be perturbed to implant false memories and alter behavior predictably.
arXiv:2608.29034v1 Announce Type: cross Abstract: A wide range of methods have been proposed for interpreting language models, delivering important insights into their inner workings. However, differ...
arXiv:2608.28980v1 Announce Type: cross Abstract: Can the specialized architectures that machine learning has traditionally built for structured data be replaced by language-based models? This questi...
The paper introduces SAE‑IDF, a data‑driven approach that uses sparse autoencoders to encode symbols into interpretable sparse codes, re‑weights dimensions by rarity, and matches units across heterogeneous symbolic systems. On 12 tasks spanning six symbolic system types, SAE‑IDF outperforms dense LLM baselines (Gemma‑2‑9B and Llama‑3.1‑8B) by up to 45 points, consistently across layers and configurations. Each sparse dimension carries a human‑readable meaning, allowing matches to decompose into named concepts, thereby enabling a scalable, universal symbolic backbone.
Fine‑tuning reshapes internal representations of large language models, affecting attention patterns and layer‑wise activations. The study shows that components identified by EAP as important for task performance cluster in specific layers, yet these layers do not align with those undergoing the largest representational changes. Additionally, overlapping EAP components across different tasks do not guarantee cross‑task transfer and can even degrade performance when tasks differ in nature.
The paper investigates how large language models encode and use relational information among tokens across transformer layers. By analyzing activations from prompts that require inferring relationships among three cyclic tokens (months, hours, weekdays, musical notes), the authors find a consistent layerwise progression: intermediate layers capture pairwise relationships, while later layers encode the full three‑token relationship to predict the next token. They also identify geometrically structured token relationships that do not influence prediction, and show that constraining models to use only causally relevant joint representations improves next‑token accuracy.
arXiv:2510. 25013v2 Announce Type: replace-cross Abstract: Mechanistic interpretability aims to reverse-engineer large language models (LLMs) into human-understandable computational circuits.
arXiv:2606. 17945v1 Announce Type: new Abstract: Large language models provide a tractable system for asking how intelligence itself emerges, rather than only how LLMs can be engineered.
The paper investigates how transformers acquire deep semantic dependencies, proposing a mechanistic framework that frames learning as a competition between surface statistics and deep semantics. It identifies a "Gradient Starvation" effect that suppresses error signals for sparse semantic dependencies early in training, delaying structural reasoning until a sudden phase transition. The study also explains the success of Chain-of-Thought strategies and introduces a topology‑aligned contrastive objective that improves variable binding performance by more than twice the gain of standard fine‑tuning.